Distributed information processing method, system, storage medium and computer program product

Abstract

The application relates to the technical field of distributed information processing, and discloses a distributed information processing method, a distributed information processing system, a storage medium and a computer program product. The method comprises the following steps: organizing a plurality of processing units into at least one processing group, the maximum communication step number between any two units in the group being a preset constant; in response to the acquisition of input data, distributing the input data to all the processing units; determining an activated unit and recording the propagation distance information of the activated unit based on the matching relationship between the input data and the local reference data of each unit; determining the decision weight of each activated unit based on the propagation distance information, the smaller the propagation distance, the higher the weight; and finally outputting a processing result based on the decision weight. The application realizes constant-step processing delay irrelevant to the total number of units by limiting the communication step number and combining distance-weighted decision, effectively eliminates the centralized control bottleneck, and significantly improves the robustness and expansibility of the system.

Description

Technical Field

[0001] This invention relates to the field of distributed information processing technology, specifically to a distributed information processing method, system, storage medium, and computer program product. Background Technology

[0002] With the rapid development of artificial intelligence, the Internet of Things, and big data technologies, distributed information processing systems have evolved from early centralized master-slave architectures to massively parallel processing architectures. Traditional von Neumann architectures, limited by the "memory wall" and "power wall," struggle to meet the ever-increasing real-time demands of the edge. To overcome this bottleneck, the industry's technological development path has gradually shifted towards in-memory computing, near-memory computing, and dedicated domain architectures (DSA). Against this backdrop, optimizing communication efficiency through physical topology reconstruction has become a key direction in the evolution of distributed systems. From the initial bus and ring topologies to later mesh and network-on-chip (NoC) topologies, technological advancements have consistently focused on reducing data transmission latency between nodes to meet the stringent requirements of complex scenarios ranging from industrial real-time monitoring to large-scale data retrieval.

[0003] Current distributed information processing systems, such as mainstream neural network inference accelerators, large-scale data retrieval systems, and industrial monitoring systems, typically employ the following patterns: In terms of hardware topology, they often use mesh or toroidal topologies, or rely on routers for full interconnection; in terms of processing logic, they generally follow a closed-loop process of "master controller scheduling—node computation—data aggregation—master controller decision-making." For example, existing technologies often use hierarchical index structures such as FM-index for gene sequence alignment, or employ a centralized master-station architecture for fault location in smart grids. During processing, these systems typically have a master controller uniformly sending control signals. Each node, upon receiving the signal, performs local computation and transmits the results back to the central node for reduction and fusion, either hierarchically or via specific routing protocols.

[0004] However, there is a significant gap between the current state of technology and the growing expectations for real-time performance and low power consumption, mainly in the following three aspects: First, communication latency worsens with scale. Whether it is ring hop-by-hop propagation or grid routing, the number of communication steps between nodes usually increases linearly by O(N) or logarithmically by O(logN) with the total number of nodes N. This makes it impossible to meet the real-time response requirements at the nanosecond or even microsecond level when processing large-scale data (such as whole genome alignment).

[0005] Secondly, there is a bottleneck in centralized control. The existing reduction operation relies heavily on the main controller for scheduling. The main controller not only bears a huge computing load, becoming a performance bottleneck of the system, but also has the risk of single point of failure, and generates high power consumption during data aggregation.

[0006] Finally, power consumption and scalability are difficult to achieve simultaneously. Traditional solutions often require activating all nodes to participate in each round of computation or rely on complex routing logic to maintain full-to-full interconnection, resulting in a large amount of ineffective power consumption. As the system scales up, latency and power consumption deteriorate simultaneously, making it difficult to deploy on resource-constrained edge devices and unable to achieve true constant-time information processing.

[0007] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0008] The main objective of this application is to provide a distributed information processing method, system, storage medium, and computer program product, which aims to achieve constant-step processing latency independent of the total number of processing units and reduce the system's dependence on centralized control.

[0009] To achieve the above objectives, this application proposes a distributed information processing method, which includes: Multiple processing units are organized into at least one processing group, wherein the maximum number of communication steps between any two processing units within the same processing group is a preset constant; In response to the acquisition of input data, the input data is distributed to all processing units; Based on the matching relationship between the input data and the reference data stored locally in each processing unit, the active processing unit is determined from all processing units, and the propagation distance information of each active processing unit is recorded. Based on the propagation distance information, the decision weight of each activation processing unit is determined, wherein the smaller the propagation distance represented by the propagation distance information, the higher the decision weight; Based on the decision weights of each activation processing unit, the processing result of the input data is output.

[0010] In one embodiment, organizing the plurality of processing units into at least one processing group includes: Each processing group is configured as a star topology containing a central unit and at least one member unit; The member units are only connected to the central unit, and there is no direct communication between the member units.

[0011] In one embodiment, the preset constant is 2.

[0012] In one embodiment, the activation processing unit includes an initial activation processing unit and an indirect activation processing unit; The step of determining the active processing unit from all processing units based on the matching relationship between the input data and the reference data locally stored in each processing unit includes: The processing unit that directly satisfies the preset conditions based on the matching relationship is determined as the initial activation processing unit; A processing unit that is activated in response to receiving an activation indication signal is identified as an indirect activation processing unit, wherein the activation indication signal is generated by the activation processing unit and sent to other processing units within the same processing group.

[0013] In one embodiment, the propagation distance information is the number of activation steps; The recording of the propagation distance information of each of the activation processing units includes: The activation step count of the initial activation processing unit is recorded as an initial value; The activation step count of the indirect activation processing unit is recorded as the step value carried by the activation indication signal plus one.

[0014] In one embodiment, the propagation of the activation indication signal satisfies a preset termination condition, which includes: reaching a preset maximum number of steps, or no new processing unit being activated.

[0015] In one embodiment, the decision weight is calculated using a monotonically decreasing function, where the independent variable of the monotonically decreasing function is the propagation distance represented by the propagation distance information.

[0016] In one embodiment, distributing the input data to all processing units includes: The input data is simultaneously sent to all processing units within the same clock cycle via a global broadcast bus.

[0017] In one embodiment, the global broadcast bus is an H-tree, mesh broadcast, wireless broadcast, or optical broadcast structure.

[0018] In one embodiment, the reference data includes at least one of prototype vectors, hash values, threshold intervals, or statistical features.

[0019] In one embodiment, the processing unit includes a non-volatile memory for permanently storing the reference data.

[0020] Furthermore, to achieve the above objectives, this application also proposes a distributed information processing system, which includes: Multiple processing units are organized into at least one processing group, wherein the maximum number of communication steps between any two processing units within the same processing group is a preset constant; A global broadcast bus is configured to distribute the input data to all processing units in response to the acquisition of input data; Each processing unit includes: Reference data storage, configured to store reference data locally; A matching degree calculation circuit is configured to calculate the matching degree between the input data and the reference data; A threshold comparator is configured to determine whether the matching degree meets a preset condition. Step counter, configured to record the number of active steps; The neighbor list register is configured to store information about neighboring units within the processing group to which this processing unit belongs. A state machine configured to control activation, step counting, and the sending and forwarding of activation indication signals; The voting accumulator is configured to aggregate the category information and activation steps of all activated processing units, calculate the decision weight based on the activation steps, and output the processing results.

[0021] In one embodiment, the preset constant is 2.

[0022] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the distributed information processing method described above.

[0023] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the distributed information processing method described above.

[0024] One or more technical solutions proposed in this application have at least the following technical effects: This invention organizes processing units into a topology with limited intra-group communication steps, constraining the communication steps between any two intra-group units to a preset constant. This decouples information processing latency from the total number of processing units, achieving constant-step processing latency. Simultaneously, by recording the propagation distance information of activated processing units and performing weighted decision-making based on propagation distance, processing units closer to the activation source have higher decision weights, improving decision accuracy and response speed. Furthermore, each processing unit autonomously activates and propagates its activation signal based on local matching relationships, eliminating the need for hop-by-hop scheduling by the main controller, thus removing centralized control bottlenecks and improving system robustness and scalability. A global broadcast bus ensures synchronous distribution of input data, further reducing latency. This system and method are particularly suitable for edge computing, real-time monitoring, and large-scale data retrieval scenarios with strict requirements for latency, power consumption, and scalability. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating an embodiment of the distributed information processing method of this application. Figure 2 This is a schematic diagram of the processing group configuration provided in Embodiment 2 of the distributed information processing method of this application; Figure 3 This is a schematic diagram of the communication path between member units provided in Embodiment 2 of this application; Figure 4 This is a schematic diagram of the internal architecture of the processing unit provided in Embodiment 3 of the distributed information processing method of this application; Figure 5 This is a schematic diagram of the distributed information processing timing operation provided in Embodiment 5 of the distributed information processing method of this application; Figure 6 This is a schematic diagram of the processing unit implementation based on resistive random access memory (RRAM) in-memory computation provided in Embodiment 7 of the distributed information processing method of this application; Figure 7 This is a schematic diagram of the system architecture of the hardware operating environment involved in the distributed information processing method in the embodiments of this application.

[0028] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0030] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0031] Example 1: like Figure 1 As shown, this embodiment provides a distributed information processing method. This method achieves constant-level processing latency decoupled from system scale by constructing specific topological constraints and introducing a weighted decision-making mechanism based on propagation distance. Specifically, it includes the following steps: Step S100: Organize multiple processing units into at least one processing group, wherein the maximum number of communication steps between any two processing units within the same processing group is a preset constant.

[0032] In this embodiment, the system includes multiple processing units. A processing unit is the smallest functional unit in the system that performs matching operations and participates in activation propagation. Each processing unit locally stores a copy of the reference data and has independent matching operation capabilities. In different application scenarios, the processing unit can take on different physical forms; for example, it may be a neural network processing unit in a neural network inference scenario, a smart electronic device in a power grid fault location scenario, or a hash calculation unit in a gene sequence alignment scenario. This embodiment does not specifically limit its physical form.

[0033] These processing units are organized into at least one processing group. A processing group is a logical collection of multiple processing units organized according to certain rules. Processing units within the same processing group can communicate with each other, and the number of communication steps is constrained by a preset constant. The specific topology of the processing group can be configured according to actual needs, and this embodiment does not impose specific limitations on it. The criteria for dividing the processing groups can be set according to actual application needs, such as dividing according to the category label of the reference data, or according to the clustering results of the feature space, according to geographical regions, or according to functional modules, etc.

[0034] The core of this step lies in locking the communication latency within a fixed range through topological constraints. In traditional distributed systems, the number of communication steps between nodes often increases linearly or logarithmically with the total number of nodes N, resulting in poor system scalability. In this embodiment, by limiting the "maximum number of communication steps to a preset constant," information interaction between any two processing units requires at most a constant number of hops. This preset constant is typically a small integer, such as 2, 3, or 4, with the specific value depending on the internal connection structure of the processing group. This design frees the communication latency within the system from the total number of processing units, thus decoupling latency from scale. Whether the system scales to hundreds or millions of processing units, the maximum communication latency within the group is always limited to the time range corresponding to this preset constant, providing a structural foundation for the real-time performance of large-scale parallel processing.

[0035] In one exemplary implementation, the above constraint can be achieved by configuring each processing group as a star topology, in which case the maximum number of communication steps between any two processing units within the same processing group is 2. However, it should be understood that the star topology is merely one specific implementation of the constraint that the maximum number of communication steps is a preset constant. In other embodiments, as long as the organizational structure used can guarantee that the maximum number of communication steps between any two processing units within the same processing group is a certain preset constant, it can be applied to the method of this embodiment.

[0036] Step S200: In response to the acquisition of input data, the input data is distributed to all processing units.

[0037] When the system receives input data, it distributes the input data to all processing units. The input data can be a feature vector to be processed, a hash sequence to be retrieved, or real-time sensor data to be monitored, etc. This embodiment does not limit the specific type of input data. For example, in a neural network inference scenario, the input data can be a feature vector obtained after preprocessing the image to be classified; in a gene sequence alignment scenario, the input data can be the hash value of the short read sequence to be queried; in a power grid fault location scenario, the input data can be the electrical quantity measurements collected in real time by each monitoring point.

[0038] The distribution process can be implemented using hardware structures such as a global broadcast bus. The core of this step is "distributing input data to all processing units" to ensure that each processing unit can process the same input data in parallel. Various methods can be used for distribution, such as physical implementations like H-tree structures, grid broadcast structures, wireless broadcast structures, or optical broadcast structures, or other methods like serial broadcasting and multi-level relays. This step does not limit the specific distribution method or timing synchronization precision, as long as each processing unit receives the input data, thereby triggering subsequent parallel processing flows.

[0039] Step S300: Based on the matching relationship between the input data and the reference data stored locally in each processing unit, determine the active processing unit from all processing units and record the propagation distance information of each active processing unit.

[0040] After receiving the input data, each processing unit performs the following processing independently.

[0041] (a) Determine the activation processing unit Each processing unit determines whether it should be identified as an active processing unit based on whether there is a matching relationship between the received input data and its own locally stored reference data.

[0042] Reference data is the basic data stored locally by each processing unit for matching and comparing with the input data. In different application scenarios, reference data can take different forms. For example, in neural network inference scenarios, reference data can be prototype vectors of various categories; in gene sequence alignment scenarios, reference data can be hash vectors of reference genome fragments; and in power grid fault location scenarios, reference data can be characteristic intervals of electrical quantities under normal operating conditions.

[0043] Matching relationships are the basis for determining the degree of correlation between input data and reference data. Matching relationships can be measured in various ways, such as cosine similarity, dot product, Euclidean distance, Hamming distance, interval membership calculation, or hash equality judgment. Preset conditions are the specific criteria for determining whether a matching relationship is valid, such as the matching result being greater than a threshold, less than a threshold, or falling within a threshold interval. This embodiment does not limit the specific calculation method of the matching relationship or the specific form of the preset conditions.

[0044] When the matching relationship between the input data and the reference data stored locally in a certain processing unit meets the preset conditions, the processing unit is determined to be the active processing unit.

[0045] (ii) Recording propagation distance information For each processing unit that is identified as an active processing unit, the system records its propagation distance information.

[0046] The propagation distance information is used to characterize how far the activation state has spread within the processing group, that is, how far an activated processing unit is from the occurrence of the initial matching event. There are various ways to record the propagation distance information, and this embodiment does not specifically limit this method.

[0047] In one implementation, the propagation distance information can be the activation step count, i.e., the number of hops the signal takes to travel from the initial activation unit to the current unit; a smaller step count indicates a closer distance to the initial matching event. In another implementation, it can be the time difference, i.e., the time interval between the transmission and reception of the activation signal; a shorter time indicates a closer distance. Alternatively, it can be the signal attenuation rate, i.e., the degree of energy attenuation during signal propagation; a smaller attenuation indicates a closer distance.

[0048] By recording propagation distance information, the system not only knows which processing units are activated, but also perceives the proximity between each activated processing unit and the initial matching event—the closer the processing unit, the more direct the correlation between its reference data and input data, and the higher the credibility of its voting opinion. This provides crucial data support for distance-based refined decision-making in subsequent steps.

[0049] Regarding the specific rules for recording propagation distance information, different types of activation processing units can have different recording methods: For processing units that are determined to be activated based on their own matching calculation results directly meeting preset conditions, their activation is triggered by the direct matching of input data and local reference data, without depending on the activation state of other processing units, and their propagation distance information can be recorded as the initial value; for processing units that are determined to be activated by other activated processing units, their propagation distance information can be updated and recorded based on the received propagation distance information. Thus, the propagation distance information increases progressively as the activation state propagates within the processing group, forming a distance metric for the activation propagation path. The subsequent embodiment 3 will elaborate on the determination methods of the initial activation processing unit and the indirect activation processing unit, as well as the specific rules for recording propagation distance information, in more detail.

[0050] Step S400: Based on the propagation distance information, determine the decision weight of each activated processing unit, wherein the smaller the propagation distance represented by the propagation distance information, the higher the decision weight.

[0051] After the propagation distance information of all activated processing units is recorded, the system determines the decision weight for each activated processing unit based on its propagation distance information. The decision weight is the voting weight assigned to each activated processing unit according to its propagation distance information.

[0052] This step establishes the core decision-making logic of this invention: the smaller the propagation distance represented by the propagation distance information, the higher the corresponding decision weight. This means that the activation processing unit that is closer to the initial matching event will gain greater influence in the final decision.

[0053] The underlying principle of this design is that in distributed information processing, processing units closer to the activation source tend to have a higher correlation between their stored reference data and the input data, or they are located on the core path of information propagation, thus their voting opinions should have higher importance. Conversely, the greater the propagation distance, the more likely the unit was activated through multiple relays, and its direct correlation with the original input may decrease step by step. Through this distance-based weight allocation, the system can effectively suppress the interference of edge noise nodes and enhance the influence of the core activation region on the final result, thereby significantly improving the accuracy and robustness of decision-making.

[0054] In specific implementation, the determination of the decision weights can be calculated through a preset monotonically decreasing function, that is, the weights are expressed as a function of the propagation distance w = f(d), where the function f satisfies the monotonically decreasing property: if d1 < d2, then f(d1) > f(d2). There are various choices for the specific form of the monotonically decreasing function, such as the reciprocal form, exponential decay form, etc. Specific examples of the monotonically decreasing function will be described in more detail in subsequent Embodiment 4.

[0055] Step S500, based on the decision weights of each of the activation processing units, output the processing result of the input data.

[0056] After determining the decision weights of each activation processing unit, the system performs weighted aggregation based on the category information and decision weights corresponding to each activation processing unit, and outputs the final processing result.

[0057] Specifically, the system collects the category labels of all activation processing units (i.e., the category identifiers associated with the reference data stored in this processing unit) and their corresponding decision weights; then summarizes the weights by category, calculates the weighted total score for each category; and finally outputs the category with the highest weighted total score as the processing result.

[0058] The weighted aggregation mechanism in this step has two characteristics: First, weighted aggregation makes the activation processing units with small propagation distances contribute more to the final decision, reflecting the principle of "the closer the distance, the more reliable"; second, the weighted aggregation mechanism does not require all processing units to participate in the decision-making. Only the category information of the activated processing units will be taken into consideration, and the unactivated processing units do not participate in the voting, which ensures that the decision result is only based on that part of the reference data actually related to the input data.

[0059] Through the above steps, this embodiment constructs a complete technical chain from topological constraints to distance recording and then to weighted decision-making. It organizes the processing units into a topological structure with limited intra-group communication steps, records the propagation distance during the activation propagation process, and finally makes a weighted decision based on the propagation distance, decoupling the processing delay from the total number of processing units and achieving a constant-step processing delay. At the same time, the distributed autonomous activation mechanism eliminates the dependence on the main controller and improves the robustness and scalability of the system.

[0060] The distributed information processing method provided in this embodiment optimizes the decision-making quality through the distance perception mechanism while ensuring a constant-level low delay, achieving efficient and accurate distributed information processing.

[0061] Specifically, the process of a single inference can be broken down into: input data distribution, parallel matching operations among processing units, activation state propagation and propagation distance recording (the delay of which depends on the maximum number of communication steps within the group, i.e., a preset constant), and weighted decision output. The delays of each of the above steps are independent of the total number of processing units in the system, and the total number of processing steps is constant.

[0062] To verify the above effects, tests were conducted in an exemplary implementation with a scale of 1024 processing units, resulting in a single-processing latency in the nanosecond range. Further scaling simulations showed that as the number of processing units increased from hundreds to hundreds of thousands, the average propagation steps remained stable near a small constant value, and the total processing latency remained constant, not increasing with scale. This fully demonstrates that the method of the present invention achieves a truly constant-step processing latency independent of the total number of processing units.

[0063] Example 2: Based on Embodiment 1 above, this embodiment further defines the organizational form of the processing group. Specifically, organizing multiple processing units into at least one processing group includes: Each processing group is configured as a star topology containing a central unit and at least one member unit; wherein the member units communicate only with the central unit and there is no direct communication between the member units.

[0064] The processing unit is the smallest functional unit in the system that performs matching operations and participates in activation propagation. Each processing unit locally stores a copy of the reference data and has independent matching operation capabilities. In different application scenarios, the processing unit can take on different physical forms, and this embodiment does not impose any specific limitations on this.

[0065] The processing units within each processing group are divided into two categories according to their roles: central units and member units. The central unit, acting as the communication hub within the processing group, undertakes the core function of signal relay. Taking a processing group containing one central unit and k member units as an example, the communication connections within this group are as follows: the central unit establishes bidirectional communication connections with each member unit, forming k independent communication links; however, there is no direct physical communication link between any two member units.

[0066] like Figure 2 As shown, processing group 1 is configured to contain 1 central unit and 5 member units; processing group 2 is configured to contain 1 central unit and 6 member units; processing group 3 is configured to contain 4 member units... The criteria for dividing processing groups can be set according to actual application needs. For example, they can be divided according to the category labels of the reference data, or according to the clustering results of the feature space, geographical regions, or functional modules.

[0067] It should be noted that a communication connection refers to a physical or logical channel that enables direct information exchange between processing units. In this embodiment, the specific implementation of the communication connection can be a wired electrical signal connection (such as PCB traces, through-silicon vias (TSVs), metal straight-through cables, etc.), a wireless signal connection (such as on-chip wireless interconnect), an optical interconnect, or a virtual connection established through logical grouping. This embodiment does not limit the specific physical implementation of the communication connection between different processing units, as long as it enables information exchange between the processing units.

[0068] It should be noted that in traditional fully connected or mesh topologies, the communication paths between nodes are complex and unpredictable. Multiple communication paths of varying lengths may exist between any two nodes, easily leading to communication congestion and uncontrollable latency. However, in the star topology described in this embodiment, there are no direct physical links between member units. This means that if any two member units need to exchange information, they must do so through the central unit.

[0069] This design, where there are no direct communication connections between member units, may seem to increase the number of hops per communication (compared to the single-hop direct approach in a fully connected topology), but it is actually the key structural foundation for achieving constant-step latency. The reason is that while a fully connected topology may only require one hop per communication, its connection complexity increases by O(N²) with the number of nodes, becoming impractical for ultra-large-scale deployments. In contrast, the connection complexity of a star topology is only O(N), and through the deterministic relay of the central unit, the number of communication steps is strictly locked to a fixed constant value.

[0070] Based on the above connection relationship, this embodiment can strictly guarantee that the maximum number of communication steps between any two processing units within the same processing group is a preset constant. As a specific implementation, the preset constant is 2.

[0071] To better understand the effect of this technology, the following analysis will consider the signal flow direction. For example... Figure 3 As shown, assume there are member unit A and member unit B within the processing group. When member unit A needs to send information to member unit B, since there is no direct connection between the member units, the signal transmission path is as follows: Step 1: Member unit A sends a signal to the central unit of its group. At this point, the communication step count is 1.

[0072] Step 2: After receiving the signal, the central unit forwards it to member unit B. At this point, the communication step count is 2.

[0073] The communication process has now ended, with a total of 2 communication steps.

[0074] Similarly, communication between member units and the central unit (whether from member to center or from center to member) requires only one step. The number of steps for communication between the central unit and itself can be considered as 0.

[0075] Therefore, regardless of the number of member units within the processing group, the maximum number of communication steps between any two processing units within the group is always strictly limited to within 2. This characteristic fundamentally solves the problem in existing technologies where communication latency increases linearly or logarithmically with the number of nodes.

[0076] Several node interconnection schemes exist in the prior art, but none have achieved the aforementioned effect. For example, patent application CN118631605A from Kunyun Information Technology Co., Ltd. discloses a topology circuit with a bus ring connection, where communication between nodes is a ring-hop-by-hop transmission, and the number of communication steps has a linear O(N) relationship with the number of nodes N. Patent application CN114365082A from Google LLC discloses a scheme for providing parallel computing and communication between nodes organized in a grid or toroidal topology, where the number of communication steps between nodes depends on the topology dimension and increases with scale. The RCE-NN accelerator proposed by Korea Aerospace University adopts a hierarchical star (H-star) topology, but achieves all-to-all communication between neurons through a specially designed NoC router, without limiting member units to only connecting to the central unit and without direct connections between members.

[0077] This embodiment explicitly limits communication between member units and the central unit, with no direct communication between member units. This ensures a highly deterministic communication path, thus completely decoupling the number of communication steps from the processing group size. This is one of the key structural features that distinguishes this invention from the aforementioned prior art solutions.

[0078] It should be understood that although this embodiment uses a star topology as an example for detailed description, those skilled in the art can choose other topologies that can achieve a maximum communication step count of a preset constant according to actual application scenarios. For example, in a tree topology, communication between any two leaf nodes needs to be forwarded through their lowest common ancestor node, and its maximum communication step count depends on the depth of the tree; in a hierarchical star topology, multiple star groups are interconnected through upper-level central units, and the inter-group communication step count can also be limited to a constant range. As long as the organizational structure adopted can ensure that the maximum communication step count between any two processing units within the same processing group is a preset constant, and that this constant does not increase with the increase in the total number of processing units, it falls within the protection scope of this invention.

[0079] This embodiment configures the processing group in a star topology, limiting member units to communicate only with the central unit and eliminating direct communication between member units. This provides a concrete structural implementation at the physical connection level for ensuring the maximum number of communication steps in the method described in claim 1 is a preset constant. This structure completely decouples intra-group communication latency from the number of processing units within the group. Regardless of whether the system scales to hundreds or millions of processing units, the maximum number of communication steps between any two units within the group is always strictly limited to a constant number of steps (e.g., 2 steps), ensuring the system maintains nanosecond-level response speeds even under ultra-large-scale deployments. Simultaneously, the design of the central unit as a communication hub simplifies signal routing logic and reduces the complexity of communication scheduling.

[0080] Example 3: Based on the above embodiments, this embodiment elaborates on the classification and determination mechanism of the activation processing unit and the specific recording method of the propagation distance information. Specifically, the activation processing unit in this embodiment includes an initial activation processing unit and an indirect activation processing unit. The step of determining the activation processing unit from all processing units based on the matching relationship between the input data and the reference data locally stored in each processing unit includes: The processing unit that directly satisfies the preset conditions based on the matching relationship is determined as the initial activation processing unit; the processing unit that is activated in response to receiving the activation indication signal is determined as the indirect activation processing unit, wherein the activation indication signal is generated by the activation processing unit and sent to other processing units in the processing group.

[0081] In this embodiment, the activation processing unit includes an initial activation processing unit and an indirect activation processing unit. The distinction between these two types of activation processing units is based on the triggering source of their activation: whether it is triggered directly by a matching relationship or by a propagation signal from another activated processing unit.

[0082] The initial activation processing unit refers to the processing unit that is determined based on the matching relationship between the input data and the local reference data, which directly meets the preset conditions.

[0083] Specifically, such as Figure 4 As shown, after receiving the input data, each processing unit independently calculates the matching degree between the input data and the locally stored reference data. The matching relationship can be measured in various ways, such as cosine similarity, dot product, Euclidean distance, Hamming distance, interval membership calculation, or hash equality judgment, or a combination of multiple methods can be used. This embodiment does not specifically limit the method.

[0084] Taking dot product as an example, the processing unit calculates the dot product between the input vector and the locally stored reference vector, which serves as the matching score. If multiple methods are combined, the matching scores can be added together using an addition tree. The final matching score is then given to the comparator for judgment. When the matching score meets a preset condition, such as exceeding a preset threshold in the threshold register, the processing unit is determined as the initial activation processing unit, i.e., the state machine is updated to active.

[0085] This activation determination process is completed entirely in parallel locally by each processing unit, without the intervention of the central control unit, thus greatly reducing the system's response latency. Since the activation of the initial activation processing unit is triggered by the direct matching of input data and local reference data, and does not depend on the activation state of other processing units, it represents the location of the reference data most closely related to the input data and is the source of information propagation.

[0086] An indirect activation processing unit refers to a processing unit that is activated in response to receiving an activation indication signal.

[0087] The activation indication signal is generated by an activated processing unit and sent to other processing units within its processing group. This signal serves as the carrier for the propagation of the activation state within the processing group, and can specifically be an electrical pulse, optical pulse, message packet, or wireless beacon. Once an activated processing unit (whether initially activated or indirectly activated) completes its activation, it generates an activation indication signal based on its stored neighbor unit information and sends it to neighboring processing units within its processing group. A neighboring processing unit refers to another processing unit within a processing group that has a direct communication connection with the current processing unit. In the star topology of this embodiment, the definition of neighbor relationships depends on the role of the current processing unit: for a member unit, it has only one neighboring processing unit, namely the central unit of the group, because member units only communicate with the central unit, and there are no direct communication connections between member units; for a central unit, its neighboring processing units include all member units within the group, because the central unit has established a direct communication connection with every member unit in the group. Neighbor unit information refers to data stored locally by the processing unit to record the identifiers of its neighboring processing units. This information is written into the configuration register of each processing unit during the system initialization phase (see the neighbor list in the diagram). When a processing unit needs to send an activation indication signal to other processing units, it determines the target of the signal based on its stored neighbor unit information. The specific storage format of the neighbor unit information can be implemented in various ways, such as recording the unique identifiers of neighbor processing units in the form of a neighbor list, or marking the neighbor relationships of each processing unit in the form of a bitmap. This embodiment does not specifically limit this.

[0088] When an inactive processing unit receives an activation indication signal, it is identified as an indirectly activated processing unit. Unlike the initially activated processing unit, the activation of an indirectly activated processing unit is not directly triggered by the matching relationship between the input data and the local reference data, but is propagated from other activated processing units via topology connections. It represents the processing unit associated with the initial matching event in the topology space.

[0089] The two-tiered mechanism of initial activation and indirect activation described above constructs an information propagation path from core matching to association diffusion. The initial activation processing unit corresponds to the reference data with the highest matching degree to the input data, serving as the source of information propagation. The indirect activation processing units correspond to the reference data adjacent to or close to the initial activation unit in the topological space; these may represent different variants within the same category, or similar but slightly different matching results. Through this mechanism, the system can not only identify the best-matching processing unit but also awaken related neighboring units through topological connections, thereby providing richer contextual information for decision-making.

[0090] It should also be noted that the propagation distance information is used to characterize how far the activated state has spread within the processing group. In this embodiment, the number of activation steps is used as the specific metric for the propagation distance information.

[0091] In this embodiment, the propagation distance information is the number of active steps; in another embodiment, the number of active steps is stored in a pedometer.

[0092] The recording of the propagation distance information of each of the activation processing units includes: The activation step count of the initial activation processing unit is recorded as an initial value; The activation step count of the indirect activation processing unit is recorded as the step value carried by the activation indication signal plus one.

[0093] Specifically, for the initial activation processing unit, since it is activated directly through the matching relationship without any intermediate nodes, its activation step count is recorded as the initial value. This initial value is typically set to 0. A step count of 0 indicates that this processing unit is the location where the initial matching event occurred, and its matching relationship with the input data is the most direct and reliable.

[0094] For the indirect activation processing unit, its activation step count is recorded as the step value carried by the received activation indication signal plus one. Specifically, the activation indication signal carries the current activation step count of the sending processing unit. The receiving processing unit extracts the step value from this signal, adds one to it, and records it as its own activation step count.

[0095] For example, an initial activation processing unit with an activation step count of 0 sends an activation indication signal to its neighbors, carrying the step value 0. Upon receiving this signal, a neighboring processing unit, if not currently activated, is identified as an indirect activation processing unit and its activation step count is recorded as 0+1=1. If the processing unit with a step count of 1 continues to forward activation indication signals (carrying the step value 1) to its neighbors, the next-level processing unit receiving the signal records its activation step count as 1+1=2. This process continues in a similar manner.

[0096] Since the step value increases by 1 with each forward, the size of the step value is directly positively correlated with the number of topological hops between the processing unit and the initial activation source. Through this hop-by-hop accumulation method, the system can accurately construct a distance field centered on the initial activation source—the smaller the step value, the closer the processing unit is to the initial activation source, and the more direct the association between its locally stored reference data and the input data; the larger the step value, the farther the processing unit is from the initial activation source, and its association with the input data is obtained through multiple relays.

[0097] To ensure the convergence and stability of the system, the propagation of the activation indication signal must meet preset termination conditions. This embodiment specifies two main termination conditions: reaching a preset maximum number of steps, or no new processing unit being activated.

[0098] The first termination condition is reaching a preset maximum number of steps. The system presets a maximum number of steps. When the number of activation steps recorded by a processing unit reaches this preset maximum number, the processing unit will no longer forward activation indication signals to its neighbors. For example, when the system presets a maximum propagation step count of 3, a processing unit whose step counter reaches 3 will no longer generate or send new activation indication signals after completing its own activation. Thus, the propagation range is limited to a finite neighborhood centered on the initial activation source, preventing invalid global flooding and keeping computational and communication overhead within a reasonable range.

[0099] The second termination condition is when no new processing unit is activated. In a round of propagation, if all activated processing units send activation indication signals to their neighbors, but no new inactive processing units are triggered to activate, the propagation process naturally terminates. This situation typically occurs when all neighbors of an activated processing unit have been activated, or when all processing units within the processing group to which the activated processing unit belongs have been activated.

[0100] The combination of the two termination conditions ensures that key information can be disseminated to a sufficient range to cover all nearby reference data worth incorporating into the decision-making process, while effectively avoiding the waste of resources caused by infinite loop propagation, thus achieving an optimal balance between computing resources and information coverage.

[0101] In this embodiment, by dividing the activation processing unit into an initial activation processing unit and an indirect activation processing unit, two triggering paths for activation are defined: direct matching trigger and propagated signal trigger, and an information propagation path from core matching to associated diffusion is constructed. By using the activation step number as a specific measure of the propagation distance information, and stipulating that the initial activation processing unit records the initial value and the indirect activation processing unit records the increment rule of adding one to the received step number value, the topological distance between each activation processing unit and the initial matching event is accurately quantified. By setting two termination conditions of a preset maximum number of steps and no new activation, the propagation range is effectively controlled, ensuring the convergence of the system. The above mechanisms work together, providing an accurate data basis for the weighted decision-making based on the propagation distance in the subsequent step S400, while eliminating the dependence on the main controller, improving the robustness and scalability of the system.

[0102] Embodiment 4: Based on the above embodiment, this embodiment elaborates in detail the specific calculation method of the decision weight. Specifically, the decision weight is calculated through a monotonically decreasing function, and the independent variable of this monotonically decreasing function is the propagation distance characterized by the propagation distance information.

[0103] A monotonically decreasing function refers to a function whose function value strictly decreases as the independent variable increases. In this embodiment, with the propagation distance d as the independent variable and the decision weight w as the dependent variable, it satisfies that when d1 < d2, w(d1) > w(d2). This monotonically decreasing property ensures the mathematical realization of the core rule that the smaller the propagation distance, the higher the decision weight.

[0104] The design principle of the monotonically decreasing function lies in: in a distributed information processing network, the processing unit with a closer propagation distance often means a higher degree of association with the initial activation source in the topological structure or feature space. The initial activation unit is the node with the highest direct matching degree to the input data, representing the core features of the input data. The units closer to the initial activation source usually have higher similarity between the stored reference data and the input data or belong to the same category cluster; while as the propagation distance increases, the indirectly activated units may be awakened through multiple relays, and the direct association degree between their reference data and the input data gradually decreases, and may even contain noise or interference information.

[0105] Therefore, by assigning higher decision weights to the units with closer distances through the monotonically decreasing function, the contribution of the core activation region to the final decision can be effectively enhanced, while suppressing the interference of the edge noise nodes, thus significantly improving the accuracy and robustness of the decision.

[0106] As a specific implementation manner, the monotonically decreasing function can adopt the form of a reciprocal function. With the propagation distance as d, the calculation formula for the decision weight w is: w = 1 / (d + 1) Where d represents the propagation distance information. When the propagation distance d is 0, that is, for the initial activation processing unit, its weight w = 1 / (0+1) = 1; when the propagation distance d is 1, the weight w = 1 / (1+1) = 0.5; when the propagation distance d is 2, the weight w = 1 / (2+1) ≈ 0.33; and so on. This reciprocal function form has the characteristics of simple calculation and hardware-friendly implementation. The division operation can be approximated by shift operation or lookup table, which is suitable for embedded systems with strict constraints on hardware resources.

[0107] As another alternative implementation, the monotonically decreasing function can also be in the form of an exponential function: w = e^(-d) Where e is the natural constant (approximately 2.71828). When the propagation distance d is 0, the weight w = e 0 = 1; when the propagation distance d is 1, the weight w = e-1 ≈ 0.368; when the propagation distance d is 2, the weight w = e-2 ≈ 0.135. The exponential function has a steeper decay characteristic than the reciprocal function, and can more strongly suppress the weights of distant nodes. This form is suitable for application scenarios that are sensitive to noise and where the decision result is expected to be mainly dominated by the core activation unit.

[0108] Alternatively, a piecewise function can be used. For example, a distance threshold D_th can be set (e.g., D_th = 1). When the propagation distance d ≤ D_th, the weight w = 1; when the propagation distance d > D_th, the weight w = 0. This piecewise function form is suitable for scenarios with extremely high real-time requirements and where only core nodes need to participate in the decision-making process. The decision result is entirely determined by the activation processing unit closest to the initial activation source.

[0109] It should be understood that the forms of the reciprocal function, exponential function, and piecewise function described above are merely illustrative examples. Those skilled in the art can choose any function that satisfies the monotonically decreasing property, based on the sensitivity to noise and the requirements for decision-making accuracy in the actual application scenario. For example, a linearly decreasing function w = a - b × d (where a and b are constants and b > 0, and the range of values ​​for the propagation distance d should ensure that the weights are non-negative), or a Gaussian function w = exp(-d² / 2σ²) (where σ is a parameter controlling the attenuation rate), can all be used as specific implementations of the monotonically decreasing function. As long as the weighting function satisfies the monotonically decreasing property that the smaller the propagation distance, the higher the decision weight, it falls within the protection scope of this invention.

[0110] To more intuitively demonstrate the technical effects of this embodiment, the following comparison will be made between this scheme and the traditional equal-weighted voting scheme.

[0111] In the equal-weighted scheme, regardless of the propagation distance, all activation processing units have the same voting weight (usually 1), and the final decision depends only on the distribution of the number of category labels of each activation processing unit.

[0112] Suppose that in a processing group, there are the following activation conditions: there is one initial activation processing unit (propagation distance d = 0), whose reference data matches the input data with an evaluation of 0.95; there are five indirect activation processing units (propagation distance d = 1), whose reference data matches the input data with an average evaluation of 0.6.

[0113] If a traditional equal-weighted voting scheme is used, each activation unit has a weight of 1. Using the matching degree as a measure of each unit's contribution to the candidate category, the final decision score is: (1×0.95 + 5×0.6) / (1+ 5) = 3.95 / 6 ≈ 0.658. In this case, a large number of indirect activation units with low matching degrees dilute the high-matching-degree contribution of the core units, causing the weighted average score to be biased towards mediocre values, and masking the advantage of the core high-matching-degree units.

[0114] If the reciprocal weighting scheme of this embodiment (w = 1 / (d + 1)) is adopted, the weight of the initial activation processing unit (d = 0) is 1.0, and the weight of each of the five indirect activation processing units (d = 1) is 0.5. The final decision score is: (1.0 × 0.95 + 5 × 0.5 × 0.6) / (1.0 + 5 × 0.5) = (0.95 + 1.50) / 3.5 = 2.45 / 3.5 = 0.7.

[0115] As can be seen from the comparison, under the reciprocal weighting scheme in this embodiment, the decision score of 0.7 is higher than that of 0.658 under the equal weighting scheme, and is closer to the true matching degree value of 0.95 for the core high-matching unit. This indicates that by monotonically decreasing the weights, the high-matching degree features of the core processing unit are effectively preserved and amplified, while noise and interference information that may exist in distant units are effectively suppressed.

[0116] This embodiment employs a monotonically decreasing function to calculate decision weights, using propagation distance as the independent variable. This achieves a precise mathematical expression that a smaller propagation distance corresponds to a higher decision weight. Based on the correlation between propagation distance and matching reliability, this mechanism effectively enhances the dominant role of the core activation region in the final decision through differentiated weights, while suppressing noise interference that may be introduced by distant indirect activation units. This significantly improves the decision accuracy and robustness of distributed information processing. Furthermore, this embodiment provides various specific implementation examples of monotonically decreasing functions, allowing this solution to be flexibly configured according to actual needs in different application scenarios.

[0117] Example 5: Based on the above embodiments, this embodiment elaborates on the input data distribution mechanism and the specific types of reference data. Specifically, in this embodiment, the input data is distributed to all processing units by simultaneously sending the input data to all processing units within the same clock cycle via a global broadcast bus.

[0118] The global broadcast bus is a key hardware infrastructure for achieving constant-step latency. In traditional bus architectures, data is often transmitted via time-division multiplexing or hop-by-hop forwarding, resulting in differences in the time it takes for different processing units to receive data. This time difference accumulates as the system scales up, severely impacting the efficiency of parallel processing—some processing units may have already completed matching operations while others have not yet received the input data, and the overall system processing latency is limited by the slowest data arrival time. In this embodiment, the global broadcast bus employs a dedicated physical routing design to ensure that input data can be synchronously transmitted from a single source to all target processing units.

[0119] As a preferred implementation, the global broadcast bus can adopt an H-tree structure. The H-tree structure is a common equal-length routing topology used for clock signal distribution and broadcast data transmission. Its basic form is a recursively constructed H-shaped routing: starting from the root node, the signal travels along the trunk to the midpoint of the first-level H-shaped branch, and then extends in four directions to form the next-level H-shaped branches, and so on recursively until all target processing units are covered. The core feature of the H-tree structure is that the physical path length from the root node to any leaf node (i.e., the target processing unit) is completely equal. This equal-length physical path ensures, at the physical level, that the signal delay time from the root node to each processing unit is consistent. In practical implementation, the signal lines of the H-tree structure employ impedance matching design to reduce signal reflection and attenuation.

[0120] As an alternative implementation, mesh broadcasting can also be used in Network-on-Chip (NoC) architectures. Mesh broadcasting utilizes the interconnected structure of a two-dimensional mesh network to broadcast data packets simultaneously in both the horizontal and vertical directions. Although the physical distances between nodes may not be exactly equal, the time difference of arrival can be controlled within an acceptable range through proper pipeline design and repeater insertion.

[0121] Furthermore, in certain specific scenarios, wireless broadcasting or optical broadcasting can also be used. For example, in an on-chip wireless interconnect scheme, the transmitting antenna simultaneously broadcasts signals to all receiving antennas; in an optical interconnect scheme, the light source distributes the optical signal to each photodetector simultaneously through a beam splitter. As long as it can be ensured that all processing units receive input data within the same clock cycle, it can be used as a specific implementation of a global broadcast bus.

[0122] It should be noted that the H-tree structure, grid broadcast, wireless broadcast, and optical broadcast described above are merely exemplary implementations of the global broadcast bus, and this embodiment does not limit the specific physical structure of the global broadcast bus. As long as the broadcast structure used can ensure that input data is delivered to each processing unit simultaneously within the same clock cycle, the function of this step can be achieved.

[0123] In one embodiment, such as Figure 5 As shown, the entire processing can be divided into 5 stages. In stage T1, the input data is broadcast globally via the broadcast bus. In stage T2, each processing unit performs parallel matching of the input data, specifically calculating the matching degree between the input data and the local reference data. In stage T3, each processing unit activates autonomously based on the matching degree (at this time, the number of steps for the activated processing unit is 0). In stage T4, the activated processing unit, i.e. the initial activation unit, sends an activation pulse carrying the step count to the neighboring units, at which point the step count increases from 0 to 1. If the initial activation unit is not the central unit, it first propagates to the central unit, which then receives it and forwards it to the neighboring units, at which point the step count increases from 1 to 2. In stage T5, each processing unit votes weighted by the step count to output a decision.

[0124] This embodiment uses a global broadcast bus to send input data to all processing units simultaneously within the same clock cycle, eliminating the time difference in data arrival and ensuring that all processing units start matching operations synchronously. This improves the efficiency of parallel processing and the determinism of decision-making, and provides hardware support at the data distribution level for achieving constant-step processing latency that is independent of the total number of processing units.

[0125] Example 6: Based on the above embodiments, this embodiment provides supplementary explanations regarding the specific types of reference data stored locally by each processing unit.

[0126] In this embodiment, the reference data stored locally by each processing unit includes at least one of prototype vectors, hash values, threshold ranges, or statistical features.

[0127] Reference data is the basic data stored locally by each processing unit for matching and comparing with the input data. The form of the reference data determines the functional positioning of the processing unit and the specific measurement method of the matching relationship described in step S300 of claim 1. By setting different types of reference data, the distributed information processing method of the present invention can flexibly adapt to different application scenarios without modifying the core topology and propagation mechanism.

[0128] When the reference data is a prototype vector, the processing unit is mainly used for pattern recognition or classification tasks. In this case, the input data is typically the feature vector of the object to be classified, and the matching relationship can be measured by calculating the similarity between the input vector and the prototype vector, for example, using dot product or cosine similarity calculation. In image recognition scenarios, the reference data can be a typical image feature vector of a certain category (i.e., the prototype of that category). The processing unit calculates the dot product between the input image features and the prototype vector; the larger the dot product value, the higher the matching degree between the input image and that category.

[0129] When the reference data is a hash value, the processing unit is mainly used for large-scale data retrieval or sequence alignment tasks. In this case, the input data is the hash value of the object to be queried, and the matching relationship can be measured by calculating the Hamming distance between the input hash value and the reference hash value. In gene sequencing scenarios, the reference data can be k-mer hash fingerprints extracted from each fragment in the reference genome. The processing unit calculates the Hamming distance between the hash value of the input short read sequence and the reference hash value; the smaller the distance, the higher the matching degree, meaning the higher the similarity between the input sequence and that genomic fragment.

[0130] When the reference data is a threshold range, the processing unit is mainly used for anomaly detection or monitoring alarm tasks. In this case, the input data is real-time monitored values ​​(such as temperature, pressure, current, voltage, etc.), and the matching relationship can be measured by determining whether the input value falls within the reference threshold range. In industrial monitoring or power grid fault location scenarios, the reference data can be the normal fluctuation range of various parameters under normal operating conditions of the equipment. If the input value falls within this range, it is determined to be a match (normal state); if it deviates from this range, it is determined to be a mismatch (abnormal state), thereby triggering the activation of the corresponding processing unit.

[0131] When the reference data is a statistical feature, the processing unit is mainly used for data stream analysis or trend prediction tasks. In this case, the input data is the statistics of the current time window (such as mean, variance, histogram distribution, etc.), and the matching relationship can be measured by calculating the probability distribution similarity between the current statistic and the reference statistical feature. In the sliding window anomaly detection scenario, each processing unit maintains statistical features within a window of different sizes as reference data. When new data arrives, the degree of fit between the current window statistic and the reference statistical feature is calculated to determine whether an anomaly exists.

[0132] By setting up multiple reference data types as described above, this embodiment demonstrates the strong versatility of the distributed information processing system. The system can flexibly adapt to various heterogeneous application scenarios, from artificial intelligence inference to industrial control, without requiring modifications to the core topology and propagation mechanism—the organization of processing units, activation propagation logic, and weighted decision-making process remain unchanged. For different application scenarios, only the corresponding type of reference data and the matching relationship calculation logic need to be configured, which also reflects the technical universality of the distributed information processing method described in claim 1.

[0133] Example 7: This embodiment provides a distributed information processing system. This system implements the distributed information processing methods described in the above embodiments through a specific hardware architecture, mapping the logical steps at the method level to circuit modules and signal flows at the physical level, thereby ensuring constant-step delay and decentralized processing at the hardware level. The system mainly includes: Multiple processing units are organized into at least one processing group, wherein the maximum number of communication steps between any two processing units within the same processing group is a preset constant; A global broadcast bus is configured to distribute the input data to all processing units in response to the acquisition of input data; Each processing unit includes: Reference data storage, configured to store reference data locally; A matching degree calculation circuit is configured to calculate the matching degree between the input data and the reference data; A threshold comparator is configured to determine whether the matching degree meets a preset condition. Step counter, configured to record the number of active steps; The neighbor list register is configured to store information about neighboring units within the processing group to which this processing unit belongs. A state machine configured to control activation, step counting, and the sending and forwarding of activation indication signals; The voting accumulator is configured to aggregate the category information and activation steps of all activated processing units, calculate the decision weight based on the activation steps, and output the processing result. The smaller the activation steps, the higher the decision weight. Specifically, the multiple processing units are organized into at least one processing group, wherein the maximum number of communication steps between any two processing units within the same processing group is a preset constant. This organization is manifested in a specific hardware physical connection topology—the communication links of processing units belonging to the same processing group are laid out in a star topology. In a star topology, each processing group contains a central unit and at least one member unit. The central unit establishes direct physical communication links with each member unit, while there are no direct physical connections between member units. This constrained topology ensures that communication between any two member units within the same processing group must be forwarded through the central unit, and the maximum number of communication steps is strictly limited to 2. Whether at the PCB routing level or the chip layout design level, this topological constraint limits the physical communication path length between any two units, thereby avoiding the increase in communication latency caused by network expansion at the hardware level, and providing a structural basis for achieving constant-step processing latency.

[0134] The global broadcast bus is configured to distribute the input data to all processing units in response to the acquisition of the input data.

[0135] In terms of hardware implementation, the global broadcast bus can adopt an H-tree structure, a mesh broadcast network, or a dedicated broadcast channel. Its core design goal is to ensure the equal length and synchronization of signal transmission, so that input data can arrive at the input ports of all processing units simultaneously within one clock cycle.

[0136] Taking an H-tree structure as an example, starting from the root node, the signal is transmitted along the trunk to the midpoint of the first-level H-shaped branch, and then extends in four directions to form the next-level H-shaped branch, and so on recursively until all processing units are covered. The H-tree structure ensures that the physical path length from the root node to any processing unit is completely equal, guaranteeing the consistency of signal delay at the physical level. In the specific implementation, the signal lines adopt impedance matching design to reduce signal reflection and attenuation, and repeater drivers are inserted step by step to ensure signal integrity.

[0137] This hardware synchronization mechanism eliminates the time skew of data arrival, providing a unified starting moment for the parallel matching calculations of all subsequent processing units, which is the basis for achieving nanosecond-level response speed.

[0138] Furthermore, each processing unit, as the core computing node of the system, integrates multiple functional modules to collaboratively complete the matching, activation, and propagation tasks. These include a reference data memory, a matching degree calculation circuit, a threshold comparator, a step counter, a neighbor list register, and a state machine.

[0139] A reference data memory is configured to store reference data locally. This memory can be static random access memory (SRAM), dynamic random access memory (DRAM), or non-volatile memory (such as flash memory or resistive random access memory, RRAM). Depending on the application scenario, the reference data can be prototype vectors, hash values, threshold ranges, or statistical features, etc. Storing the reference data locally within the processing unit avoids the high latency and power consumption caused by frequent access to external main memory during processing, achieving the data locality advantage of in-memory computing, and enabling each processing unit to independently complete matching operations without intervention from the main controller. In one embodiment, as... Figure 6 As shown, Figure 6 This paper demonstrates a processing unit implementation scheme based on resistive random access memory (RRAM) in-memory computation, which is a specific hardware implementation of a single processing unit in the distributed information processing system of this invention. This scheme utilizes the analog computation characteristics of the RRAM cross-array to directly perform the dot product operation between the input vector and the reference prototype vector within the storage cell, thereby significantly reducing data transfer power consumption and accelerating matching speed. As shown in the figure, the RRAM cross-array consists of multiple word lines (WL1~WL3) and multiple bit lines (BL1~BL4), with an RRAM storage cell (R) placed at each cross-point. The conductance value of each RRAM cell is programmed to the magnitude of the corresponding reference vector component (i.e., the weight value of the prototype vector, i.e., the reference data). The input vector, i.e., the input data, is applied to the word lines in the form of voltages (V1~V3). According to Ohm's law and Kirchhoff's current law, the current output on each bit line (I1~I4) is equal to the sum of the products of the conductance of each column of RRAM cells and the corresponding input voltage, which directly represents the dot product result between the input vector and the prototype vector stored in that column. This simulation computation process achieves in-memory computation, avoiding the power consumption and latency overhead of frequent external memory accesses in traditional von Neumann architectures. The output current of each bit line is fed into a comparator (COMP) and compared with a reference current (I_ref). This reference current corresponds to a preset threshold for the matching degree. If the output current of a bit line is greater than or equal to the reference current, it means that the similarity between the input vector and the prototype vector of that column meets the condition, and the comparator outputs an activation signal. This activation signal, along with the locally stored category information (i.e., the category label corresponding to the prototype), is sent to the "Digital State Machine & Neighbor List" module. The digital state machine is responsible for maintaining the activation state and step counter of this processing unit (initially with the step count set to 0), and forwarding activation indication signals (carrying step count information) to the central unit or member units based on the neighbor unit information within the same star group recorded in the neighbor list. At the same time, this module also receives activation signals from other processing units, realizing hop-by-hop incrementing of the step count and autonomous propagation.

[0140] A matching degree calculation circuit is configured to calculate the matching degree between the input data and the reference data. The specific implementation of this circuit depends on the method of measuring the matching relationship. If the matching degree is a dot product or cosine similarity, the circuit can consist of a multiplier array and an adder tree. The multiplier array calculates the product of each component of the input vector and the reference vector in parallel, and the adder tree accumulates the products at multiple levels, finally outputting the matching degree score. If the matching degree is a Hamming distance, the circuit can consist of an XOR gate array and a counter. The XOR gate array compares the differences between each bit of the input hash value and the reference hash value in parallel, and the counter counts the number of different bits, outputting the Hamming distance value. In analog computing scenarios, this circuit can also utilize Kirchhoff's laws to implement current-mode analog multiplication and addition operations through a resistor or memristor cross array. The input voltage is applied to the word line, the reference conductance is stored at the cross point, and the bit line output current directly represents the multiplication and addition result, in exchange for higher energy efficiency.

[0141] A threshold comparator is configured to determine whether the matching degree meets a preset condition. The comparator receives a score signal output from the matching degree calculation circuit and compares it with a preset threshold voltage or value. When the matching degree score meets the preset condition (e.g., exceeds the threshold), the comparator outputs an activation trigger signal. This threshold can be a preset fixed value or an adjustable value configured through a register to adapt to the sensitivity requirements of different application scenarios.

[0142] A step counter is configured to record the number of active steps. This counter typically consists of a set of flip-flops or registers used to store the propagation distance information of the current processing unit. Its initial value can be reset to 0, and upon receiving an activation indication signal from a neighboring unit, it is incremented by one based on the step value carried in the signal and then written back. The bit width of this counter determines the maximum number of propagation steps that can be recorded; for example, a 2-bit counter can record 0 to 3 steps, which is sufficient for constant-step propagation. When the step counter reaches the preset maximum number of steps, the state machine will no longer forward the activation indication signal to neighboring units, thus terminating the propagation.

[0143] The neighbor list register is configured to store information about neighboring units within the processing group to which this processing unit belongs. This register stores the connection relationships of the current processing unit in the topology—during initialization, the system writes the corresponding neighbor unit identifiers into this register based on the processing group number and role within the group. In a star topology, the neighbor list of a member unit only contains the identifier of the central unit of its group, and the neighbor list of the central unit contains the identifiers of all member units in its group. This list guides the direction of activation signal transmission, ensuring that signals propagate only within the processing group along a preset topology path, avoiding global flooding.

[0144] The state machine is configured to control activation, step counting, and the transmission and forwarding of activation indication signals. It is the control core of the processing unit, internally defining multiple states such as idle, matched, active, and propagation states.

[0145] In the idle state, the state machine waits for input data to arrive from the global broadcast bus. Once the input data is written to the input register of the processing unit, the state machine transitions to the matching state.

[0146] In the matching state, the state machine initiates the matching degree calculation circuit, sending the input data and reference data from the reference data memory into the calculation circuit for matching operations. After the operation is completed, the threshold comparator judges the matching result. If the matching degree meets the preset condition, the state machine jumps to the active state, sets the step counter to its initial value (e.g., 0), and generates an activation indication signal based on the neighbor unit information stored in the neighbor list register, sending it to each neighbor processing unit. If the matching degree does not meet the preset condition, the state machine returns to the idle state, waiting for the next round of input data.

[0147] If a processing unit receives an activation indication signal from a neighboring processing unit while in an idle or matched state, the state machine determines whether the current processing unit is already in an active state. If not, the state machine transitions to the active state, extracts the step value from the received activation indication signal, updates the step counter to that step value by one, and forwards the activation indication signal to all neighboring processing units except the sender, based on the contents of the neighbor list register. If the unit is already active, the activation indication signal is ignored, and activation is not repeated.

[0148] In the active state, the processing unit outputs its active state signal, category label, and the current value of the step counter for the voting accumulator to read. If the step counter has reached the preset maximum number of steps, the state machine, after completing the active state output, will no longer forward the active indication signal and will return to the idle state to wait for the next round of processing.

[0149] Through precise state machine control, the processing units achieve autonomous activation and propagation logic without intervention from the main controller. Each processing unit is activated only when the matching degree condition is met or when an activation indication signal is received; inactive processing units remain in a static standby state. Combined with the design of storing reference data locally in the processing units, this architecture achieves energy efficiency characteristics where dynamic power consumption is only consumed when activated, and static low power consumption is maintained when inactive. It is particularly suitable for edge computing and embedded applications with strict power consumption requirements.

[0150] The voting accumulator is configured to aggregate the category information and activation steps of all activated processing units, calculate the decision weight based on the activation steps, and output the processing results.

[0151] This module can be a standalone hardware unit, connected to the activation status output of each processing unit via dedicated signal lines; alternatively, it can be a combinational logic circuit shared by the processing units. During the voting phase of each round of processing, the voting accumulator polls the activation status of each processing unit. For each active processing unit, the voting accumulator reads its category label and step counter values.

[0152] The voting accumulator calculates the decision weight of the processing unit based on the read step value, where a smaller activation step count results in a higher decision weight. Weight calculation can be implemented using a lookup table (LUT) or an arithmetic logic unit (ALU). For example, for a scheme using the reciprocal function w = 1 / (steps + 1), the weight value can be calculated using a divider or a shift approximation circuit; for a scheme using the exponential function w = exp(-steps), the pre-calculated weight value can be stored in a lookup table and read directly using the step value as the address.

[0153] The voting accumulator sums the weights by category, calculating a weighted total score for each category. After summing, the voting accumulator compares the total scores of each category and outputs the category with the highest score as the final decision result. This result is the system's output after processing the input data.

[0154] In one embodiment, the voting accumulator is set as a globally independent hardware computing unit, relying on a dedicated global broadcast bus to complete real-time data linkage and docking with all distributed processing units.

[0155] Through the hardware architecture described above, this embodiment materializes the logical steps in the method embodiment into specific circuit behaviors.

[0156] When input data arrives, the global broadcast bus synchronously distributes the input data to all processing units within the same clock cycle. The matching degree calculation circuits within each processing unit operate in parallel, completing the matching operation between the input data and local reference data within one clock cycle. A threshold comparator judges the matching result; processing units that meet the conditions enter the active state through state machine control, record the initial step count, and send an activation indication signal to neighboring units. The activation indication signal propagates hop-by-hop within the restricted topology, and the receiving unit's state machine updates its step counter based on the received signal and continues forwarding. The propagation process terminates when the preset maximum number of steps is reached or no new unit is activated. A voting accumulator collects the category information and step values ​​of all active processing units, calculates the score for each category according to weighted rules, and outputs the final result.

[0157] It should be noted that this full-domain performance simulation is based on the Verilog standardized RTL hardware model to build the test environment. The main frequency of the full-domain controller is fixed at 200MHz, and the total number of distributed processing units deployed across the entire domain is 1024. The benchmark is a traditional Mesh distributed topology computing system of the same scale.

[0158] The distributed information processing system provided in this embodiment organizes multiple processing units into processing groups with limited intra-group communication steps. Combined with the synchronous distribution mechanism of a global broadcast bus and the collaborative work of the integrated reference data memory, matching degree calculation circuit, threshold comparator, step counter, neighbor list register, and state machine within each processing unit, the distributed information processing method described in the above embodiments is implemented at the hardware level. The voting accumulator calculates the decision weight based on the number of activation steps (the smaller the number of activation steps, the higher the decision weight) and outputs the processing result, fully implementing the core weighted decision mechanism of this invention. Through the parallelism and decentralized architecture of the circuits, this system achieves high energy efficiency and high reliability while maintaining constant-step processing latency, making it suitable for various application scenarios with stringent real-time requirements, such as neural network inference acceleration, gene sequence alignment, and power grid fault location.

[0159] Example 8: This embodiment uses three representative application scenarios as examples to illustrate in detail the specific applications of the above-mentioned distributed information processing method and system and the significant technical effects they bring. It should be understood that this embodiment is merely illustrative and is not intended to limit the scope of protection of this invention.

[0160] Scenario 1: Neural Network Inference This scenario uses handwritten digit recognition (MNIST dataset) as an example to demonstrate the application of this invention in accelerating neural network inference.

[0161] The system comprises 1024 processing units. In this scenario, each processing unit is specifically a neural network processing unit (PE). Each PE locally stores reference data, specifically a prototype vector of a handwritten digit for a particular category. The 1024 PEs are organized into 10 processing groups, corresponding to the 10 categories of digits 0 to 9. Each processing group is configured in a star topology, consisting of a central PE and several member PEs. Member PEs communicate only with the central PE; there are no direct communication connections between member PEs. Therefore, the maximum number of communication steps between any two PEs within the same processing group is 2.

[0162] It should be noted that all star topology architectures across the entire domain are equipped with a central control unit, which simultaneously stores corresponding category-specific reference matching data and participates in the entire process of front-end data comparison and linkage activation.

[0163] When a handwritten digit image is input into the system, the image is first converted into a 784-dimensional feature vector as input data. The system simultaneously transmits the input data to all PEs within the same clock cycle via a global broadcast bus. The global broadcast bus is implemented using an H-tree structure, ensuring that the physical path length from the root node to each PE is equal, thus guaranteeing consistent signal delay. The system operates at a clock frequency of 200MHz, with a single cycle time of 5ns.

[0164] After receiving the input data, each PE independently calculates the matching degree between the input vector and its local prototype vector (specifically, the dot product operation in this scenario). The matching degree calculation circuit inside the PE consists of 784 multipliers and a single-level adder tree, which can complete the dot product operation within one clock cycle. The threshold comparator compares the dot product result with a preset threshold. PEs that meet the condition are determined as the initial activation processing unit, and their step counter is set to 0.

[0165] The initial activation PE generates an activation indication signal, carrying its PE identifier, step value 0, and category label, and sends it to its neighboring PEs, i.e., the center PE of the group. Upon receiving the activation indication signal, if the center PE is not yet activated, it is identified as an indirect activation processing unit, sets its step counter to 1, and generates a new activation indication signal (carrying a step value of 1) to forward to other member PEs in the group. Member PEs receiving this signal, if not yet activated, are identified as indirect activation processing units, and their step counters are set to 2.

[0166] After the propagation ends, the voting accumulator collects the category labels and step values ​​of all activated PEs, calculates the decision weight based on the step value (the smaller the number of activated steps, the higher the decision weight), sums up the scores of each category in a weighted manner, and outputs the category with the highest score as the recognition result.

[0167] At a scale of 1024 PEs, the test results for the MNIST handwritten digit recognition task in this scenario are as follows: recognition accuracy of 97.8%, and a single inference latency of 50ns (input broadcast 5ns + parallel matching 5ns + activation propagation up to 2 hops 10ns + voting output 5ns, totaling approximately 50ns after adding engineering overhead). Scale-up simulations show that as the number of PEs increases from 100 to 100,000, the average propagation steps remain stable between 2.04 and 2.20, and the total processing latency remains constant, demonstrating the core effect that the processing latency is independent of the total number of PEs.

[0168] Scenario 2: Gene Sequence Alignment This scenario uses human genome short read sequence alignment as an example to demonstrate the application of this invention in the field of bioinformatics.

[0169] In gene sequencing data analysis, it is necessary to align the short reads obtained from sequencing onto a reference genome to determine their origin. Traditional methods such as BWA-MEM and Bowtie rely on hierarchical index structures like the FM-index for step-by-step searching. The location delay increases logarithmically with the size of the reference genome, making it difficult to achieve constant-step searches.

[0170] This scenario divides the human reference genome (approximately 3 billion base pairs) into approximately 120 million overlapping seed fragments, extracting the k-mer hash value as its feature for each fragment. The system contains a corresponding number of processing units, each storing the hash vector of one fragment as reference data. Processing units are organized into multiple processing groups based on chromosome number and location interval—fragments located close to each other within the same chromosome form a processing group, and each processing group is configured with a star topology. The division of processing groups follows a preset division rule (here, chromosomal regions). The category label for all processing units is chromosome number + starting position.

[0171] When a short read sequence arrives, its k-mer hash value is extracted as input data and simultaneously sent to all processing units via a global broadcast bus. Each processing unit calculates the Hamming distance between the input hash value and its local reference hash value. If the Hamming distance is less than a preset threshold, the processing unit is identified as an initially activated processing unit, and its step count is recorded as 0. Activated units send an activation indication signal to the central unit of their processing group, which then forwards it to other members within the group, with a maximum propagation step count of 2.

[0172] The voting accumulator collects chromosomal location information and activation steps from all activation processing units, calculates the score for each candidate position by weighting the number of steps, and outputs the top K positions with the highest scores as the alignment results. This method transforms the traditional O(log N) tree-structured search of FM-index into a constant-step flattened matching: all processing units calculate the matching degree in parallel, the activation signal propagates within the star group in at most 2 steps, weighted voting is completed in constant time, and the localization delay is independent of the total number of fragments N in the reference genome.

[0173] In simulation tests, the latency of a single short read comparison decreased from approximately 2.3 milliseconds for BWA-MEM to approximately 0.8 microseconds for this method (under ideal hardware parallel conditions), representing a speedup of approximately 2875 times, while maintaining an accuracy of 99.1% (considering neighboring segment co-validation). The core reason for this acceleration is that the traditional FM-index method requires searching level by level along the tree index starting from the root node, with the number of search steps being related to the index depth (logarithmic level); while this invention flattens the search process into a constant-step operation by broadcasting and matching all segments in parallel, achieving complete decoupling between processing latency and database size.

[0174] Scenario 3: Distributed Fault Location in Smart Grids This scenario uses distributed fault location in a smart grid as an example to demonstrate the application of this invention in the field of industrial real-time control. It should be understood that this scenario is merely illustrative, and those skilled in the art can apply this invention to other similar distributed real-time control scenarios such as rail transit protection and oil pipeline monitoring.

[0175] In smart grids, quickly and accurately locating fault areas is crucial for isolating faults and ensuring the safe and stable operation of the power grid. Traditional fault location solutions often employ a centralized master station architecture, where monitoring data from each substation or feeder needs to be aggregated at the master station for unified analysis. Communication and processing delays increase linearly with the scale of the power grid, typically reaching the hundreds of milliseconds level, which is insufficient to meet the millisecond or even microsecond response speed requirements of relay protection.

[0176] This scenario organizes multiple Intelligent Electronic Devices (IEDs) into at least one processing group. Each processing group contains a central IED and at least one member IED, configured in a star topology. Member IEDs communicate only with the central IED; there are no direct communication connections between member IEDs. In actual deployment, the grouping method can be determined based on the power grid topology. For example, multiple IEDs within the same substation can be grouped into one processing group, or multiple IEDs on the same feeder can be grouped into one processing group. The preset grouping rules can be formulated according to the actual situation of the protected area. Therefore, the maximum number of communication steps between any two IEDs within the same processing group is 2. This means that regardless of the expansion of the power grid, information interaction between devices within the group requires at most two hops: member → center → member, fundamentally solving the problem of communication latency increasing with the number of nodes in traditional solutions.

[0177] Each IED internally stores reference data, which describes the electrical characteristics of the area where the IED is located under normal conditions, such as current amplitude range, voltage phase angle reference value, and harmonic content threshold. This reference data is specifically of the threshold range type.

[0178] When a fault such as a short circuit occurs in the power grid, abnormal electrical quantities such as voltage drops and current surges are acquired as input data. In response to the acquisition of this input data, the system simultaneously transmits it to all IEDs via a global broadcast bus. This global broadcast bus can be implemented using a high-speed communication network within the substation, such as HSR (High Availability Seamless Ring Network) or TSN (Time Sensitive Network), ensuring that fault characteristic data can be instantly synchronized to all relevant devices, avoiding the time loss associated with hierarchical reporting in traditional solutions.

[0179] After receiving input data, each IED independently calculates the deviation between the real-time measured value and local normal characteristics (such as current surges and impedance calculations) as the matching degree. A threshold comparator determines whether the deviation meets a preset condition (such as exceeding a preset threshold). If the condition is met, the IED is identified as the initial activation processing unit, and its activation step count is recorded as the initial value 0. This process is completed entirely in parallel locally by each IED without the need for intervention from the main station, greatly shortening the fault determination time.

[0180] The initial activation of the IED generates an activation indication signal. In a power grid scenario, this activation indication signal is specifically manifested as a GOOSE message (Generic Object Oriented Substation Event). The GOOSE message is a fast message transmission mechanism defined in the IEC 61850 standard, used to transmit protection and control information between IEDs. This GOOSE message carries propagation step information. The initially activated IED sends GOOSE messages to its neighboring IEDs within its processing group based on the neighbor cell information stored in its neighbor list register.

[0181] An IED that receives a GOOSE message but is not yet activated is activated, becoming an indirect activation processing unit. Its activation step count is recorded as the received step value plus one, and it continues to forward GOOSE messages to its neighboring IEDs. Due to the constraints of the star topology, the number of steps a GOOSE message propagates within the group is strictly limited: member IED_A detects a fault activation (step 0) and sends a GOOSE message to the central IED; the central IED receives and activates (step 1) and forwards it to other member IEDs; other member IEDs receive and activate (step 2). The propagation process terminates when a preset maximum number of steps (e.g., 3) is reached or no new IED is activated. This autonomous propagation mechanism based on a neighbor list allows fault information to spread rapidly within a local area without the need for unified scheduling by a main station, achieving decentralized collaborative awareness.

[0182] The voting accumulator aggregates the area identifiers (i.e., category information) and activation steps of all activated IEDs, and calculates the decision weight based on the activation steps, where the smaller the activation steps, the higher the decision weight. The decision weight is calculated using a monotonically decreasing function, such as the reciprocal function w = 1 / (d+1). The IED closest to the fault point (with 0 steps) has the highest weight, and the weight of IEDs slightly farther away (with 1 or 2 steps) decreases accordingly. Based on the decision weights of each activated processing unit, the system outputs the processing result of the input data—that is, the fault location result: for example, if the central IED and multiple member IEDs in a certain area are all activated, and the IED with the smaller step count has the highest accumulated weight score, then the area is determined to be a fault area, and the corresponding circuit breaker tripping command is directly triggered.

[0183] In simulations of the IEEE 39-bus standard test system, the accuracy of three-phase short-circuit fault location reached 100%, with a total delay of only 1.2 microseconds from fault occurrence to trip command issuance (under ideal hardware conditions), far lower than the 50-200 milliseconds of traditional solutions. The technical reason for this significant improvement lies in the fact that traditional centralized master station architectures require each node to upload measurement data to the master station level by level, which then analyzes the data and issues control commands level by level. Communication and processing delays increase linearly with the scale of the power grid. In contrast, this invention, through global broadcast synchronous triggering, local parallel determination by each IED, constrained propagation in a star topology, and weighted voting output, locks the delay of the entire fault location process within a constant step range, achieving complete decoupling of processing delay from the scale of the power grid.

[0184] This embodiment fully demonstrates the cross-domain versatility and significant technical effects of the distributed information processing method and system of this invention through three application scenarios: neural network inference, gene sequence alignment, and smart grid fault location. In the neural network scenario, the inference latency is 50 ns and is independent of the number of PEs; in the gene sequence alignment scenario, the speedup reaches 2875 times with an accuracy of 99.1%; in the smart grid fault location scenario, the location latency is 1.2 microseconds with 100% accuracy, far superior to traditional solutions. These three scenarios collectively verify the effectiveness of the core mechanism of topological constraints—distance recording—weighted decision-making, as well as the invention's contributions in terms of constant-step processing latency, decentralized robustness, and cross-domain versatility.

[0185] Example 9: This embodiment, based on the above embodiments, further demonstrates the versatility of the distributed information processing method of the present invention in various other application scenarios. It should be understood that the following scenarios are merely illustrative and are not intended to limit the scope of protection of the present invention.

[0186] Scenario 1: Hierarchical progressive filtering (constant-step index query based on parallel voting) In large-scale database retrieval, traditional solutions typically use tree-based index structures (such as B+ trees, LSM trees, FM-indexes, etc.) for step-by-step searches. Query latency increases logarithmically with the size of the database and the depth of the index, making it difficult to meet the real-time query requirements of ultra-large-scale datasets.

[0187] This scenario organizes processing units into a multi-level index structure. Each processing unit represents an index node, storing the hash range or numerical range of the corresponding index interval as reference data. All processing units are divided into multiple processing groups according to index level and coverage. When a query request arrives, the query features are used as input data and simultaneously sent to all index nodes via a global broadcast bus. Each processing unit independently determines whether the query features fall within its own coverage range; if the condition is met, it is determined as the initially activated processing unit. Activated processing units propagate activation signals along the topology within their processing groups and record propagation distance information. The voting accumulator performs weighted voting based on the category information (i.e., the data branch it points to) and propagation distance of each activated processing unit, according to the rule that smaller propagation distances have higher weights, quickly locating candidate branches and then performing precise matching directly within those branches.

[0188] Unlike traditional hierarchical index traversal queries with a latency of O(log N), this method involves simultaneous voting by index nodes at all levels, flattening the tree-like search into a constant-step operation. Simulation tests show that with 100 million records and a three-level index configuration, the retrieval latency is reduced from 67 microseconds to 1 microsecond, achieving a speedup of 67 times, and the query latency is independent of the database size.

[0189] Scenario 2: Incremental cache prefetching (similarity-aware broadcast cache) In applications such as web services and database queries, cache hit rate directly impacts system response speed. Traditional caching solutions often employ exact matching strategies, meaning a cached entry will only be hit if the query request is exactly the same as the cached entry, failing to leverage the similarity between queries to improve the hit rate.

[0190] In this scenario, each entry in the cache is converted into a feature vector as reference data and stored in each processing unit. When a new query arrives, the query features are used as input data and simultaneously sent to all processing units via a global broadcast bus. Each processing unit calculates the similarity between the query features and the local cached feature vector in parallel. If the similarity meets a preset condition, the processing unit is determined as the initially activated processing unit. Activated processing units propagate activation signals along the topology within the processing group and record propagation distance information. A voting accumulator performs weighted voting based on the category information (i.e., cache entry identifier) ​​and propagation distance of each activated processing unit to determine whether the cache has been hit. If a hit is found, the cache result is returned; simultaneously, the most similar set of data is prefetched into a higher-speed cache level based on the voting results.

[0191] Compared to traditional exact match caching, this method can hit cached content that is similar to but not exactly the same as the query, significantly improving the cache hit rate. Simulation results show that in a test with 100 cache prototypes and 1000 queries, the similarity cache hit rate increased from 0% to 2.8%, and the matching time was independent of the number of cache entries.

[0192] Scenario 3: Approximate Probability Filtering (Multiple Filters in Parallel Voting) In distributed systems, network routing, and database queries, probabilistic filters (such as Bloom filters, counting Bloom filters, and cuckoo filters) are widely used to quickly determine whether an element belongs to a set. Traditional solutions typically use a single filter, which inherently conflicts between false positive rate and query speed.

[0193] In this scenario, multiple probabilistic filters of different types are deployed on different processing units. Each processing unit stores the parameter configuration of a filter as reference data—for example, the hash function seed and bit array of a Bloom filter, and the counter array of a counting Bloom filter. When querying whether an element belongs to a set, the hash value of that element is used as input data and simultaneously sent to all processing units via a global broadcast bus. Each processing unit independently runs its deployed probabilistic filter to determine whether the input element might belong to the set. If it is determined to be "possibly," it is identified as the initially activated processing unit. Activated processing units propagate activation signals along the topology within the processing group and record propagation distance information. A voting accumulator performs weighted voting based on the judgment results of each activated processing unit and the propagation distance, and merges the results to arrive at the final conclusion.

[0194] Compared to traditional single-filter schemes, this method uses multiple heterogeneous filters to make parallel judgments and then weighted fusion, which can improve the judgment speed while maintaining the same false positive rate, or reduce the false positive rate without reducing the speed. Simulation results show that in a scenario where three Bloom filters with different configurations vote in parallel, the judgment accuracy is on par with the best single filter, but the time speedup under hardware parallel conditions reaches 10 times.

[0195] Scenario 4: Sliding Window Aggregation (Multi-Window Parallel Anomaly Detection) In scenarios such as data stream monitoring, industrial sensor monitoring, and financial risk control, it is often necessary to continuously calculate the statistical characteristics of different time windows to detect anomalies. Traditional solutions require sequential calculation of each window, and the detection latency increases linearly with the number of windows.

[0196] In this scenario, multiple processing units are configured as sliding windows of different sizes. Each processing unit maintains a statistical data structure of a specific window size as reference data—such as the mean, variance, and histogram distribution of the data within the window. When new data arrives, it is used as input data and simultaneously sent to all processing units via a global broadcast bus. Each processing unit independently updates the statistics within its window and calculates the anomaly score for the current window. If the anomaly score meets a preset condition, the processing unit is determined as the initially activated processing unit. Activated processing units propagate activation signals along the topology within the processing group and record propagation distance information. A voting accumulator aggregates the anomaly scores and propagation distances of all processing units, calculates a comprehensive anomaly score according to the rule that smaller propagation distances have higher weights, and quickly determines whether there are anomalies in the current data stream.

[0197] Unlike traditional methods that require window-by-window calculation, this method performs parallel calculations across all windows, making the detection latency independent of the number of windows. Simulations show that with a configuration of four windows (sizes 10, 20, 50, and 100), the traditional sequential calculation takes 20.98 milliseconds, while the parallel voting method requires only 1 microsecond (under ideal hardware conditions), achieving a speedup of 20976 times. Furthermore, the overall accuracy of multi-window voting is 51.5%, which is better than the 48.75% of a single fixed window.

[0198] Scenario 5: Parallel Active Learning (Multi-Source Parallel Labeling) In the data labeling phase of machine learning, active learning strategies reduce labeling costs by selecting the most valuable unlabeled samples and submitting them to labeling sources. Traditional approaches typically query multiple labeling sources sequentially, with labeling time directly proportional to the number of labeling sources and the number of samples.

[0199] In this scenario, multiple annotation sources (including different AI models, human experts, or crowdsourcing platforms) are assigned to different processing units. Each processing unit stores the annotation capabilities and historical accuracy of its corresponding annotation source as reference data. When a model's prediction confidence for a batch of samples is low and external annotation is needed, the feature vectors of these samples are used as input data and simultaneously sent to all processing units via a global broadcast bus. Each processing unit assigns a label and confidence score to each sample in parallel. If the annotation confidence score of an annotation source for a sample meets a preset condition, that processing unit is determined as the initially activated processing unit. Activated processing units propagate activation signals along the topology within the processing group and record propagation distance information. A voting accumulator aggregates the annotation results and propagation distances of all processing units, performs weighted voting according to the rule that smaller propagation distances have higher weights, and merges them to obtain the final label for each sample. The initial weight of each annotation source can be preset based on its historical annotation accuracy.

[0200] Unlike traditional active learning that serially queries annotation sources, this method decouples annotation time from the number of samples and annotation sources. Simulations show that, with a configuration of 10 annotation sources and 100 samples, the annotation completion time is reduced from 169 microseconds in serial mode to 1 microsecond under ideal hardware parallel conditions, and the accuracy after multi-source voting fusion is improved from 95% to 100%.

[0201] This embodiment further demonstrates the cross-domain versatility and significant technical acceleration effect of the distributed information processing method of this invention through five extended application scenarios: hierarchical progressive filtering, incremental cache prefetching, approximate probabilistic filtering, sliding window aggregation, and parallel active learning. In each of these scenarios, this method eliminates the latency factors related to the number of processing units or data scale in traditional schemes by adopting a flattened processing approach of "participating all processing units in parallel in matching operations and voting decisions," achieving constant-step processing latency and obtaining speedups of tens to tens of thousands of times. This fully verifies the technical universality of the core mechanism of topological constraints—distance records—weighted decision-making of this invention, as well as its practical value in various heterogeneous application scenarios.

[0202] This application provides a distributed information processing system, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the distributed information processing method in Embodiment 1 above.

[0203] The following is for reference. Figure 7 This document illustrates a structural diagram suitable for implementing a distributed information processing system according to embodiments of this application. The distributed information processing system in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 7 The distributed information processing system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0204] like Figure 7As shown, the distributed information processing system may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the distributed information processing system. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the distributed information processing system to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a distributed information processing system with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0205] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0206] The distributed information processing system provided in this application, using the distributed information processing method in the above embodiments, can achieve a constant-step processing delay independent of the total number of processing units and reduces the system's dependence on centralized control. Furthermore, other technical features of this distributed information processing system are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0207] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0208] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0209] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the distributed information processing method in the above embodiments.

[0210] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0211] The aforementioned computer-readable storage medium may be included in a distributed information processing system; or it may exist independently and not be assembled into a distributed information processing system.

[0212] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the distributed information processing system, enable the distributed information processing system to implement the distributed information processing method of the above embodiments.

[0213] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0214] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0215] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0216] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described distributed information processing method. This enables constant-step processing latency independent of the total number of units, effectively eliminating centralized control bottlenecks and significantly improving the robustness and scalability of the system. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the distributed information processing method provided in the above embodiments, and will not be repeated here.

[0217] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the distributed information processing method described above.

[0218] The computer program product provided in this application can achieve a constant-step processing delay independent of the total number of units, effectively eliminating the centralized control bottleneck and significantly improving the robustness and scalability of the system. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the distributed information processing method provided in the above embodiments, and will not be repeated here.

[0219] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A distributed information processing method, characterized in that, The method includes: Multiple processing units are organized into at least one processing group, wherein the maximum number of communication steps between any two processing units within the same processing group is a preset constant; In response to the acquisition of input data, the input data is distributed to all processing units; Based on the matching relationship between the input data and the reference data stored locally in each processing unit, the active processing unit is determined from all processing units, and the propagation distance information of each active processing unit is recorded. Based on the propagation distance information, the decision weight of each activation processing unit is determined, wherein the smaller the propagation distance represented by the propagation distance information, the higher the decision weight; Based on the decision weights of each activation processing unit, the processing result of the input data is output.

2. The distributed information processing method according to claim 1, characterized in that, The process of organizing multiple processing units into at least one processing group includes: Each processing group is configured as a star topology containing a central unit and at least one member unit; The member units are only connected to the central unit, and there is no direct communication between the member units.

3. The distributed information processing method according to claim 2, characterized in that, The preset constant is 2.

4. The distributed information processing method according to claim 1, characterized in that, The activation processing unit includes an initial activation processing unit and an indirect activation processing unit; The step of determining the active processing unit from all processing units based on the matching relationship between the input data and the reference data stored locally in each processing unit includes: The processing unit that directly satisfies the preset conditions based on the matching relationship is determined as the initial activation processing unit; A processing unit that is activated in response to receiving an activation indication signal is identified as an indirect activation processing unit, wherein the activation indication signal is generated by the activation processing unit and sent to other processing units within the same processing group.

5. The distributed information processing method according to claim 4, characterized in that, The propagation distance information is the number of activation steps; The recording of the propagation distance information of each of the activation processing units includes: The activation step count of the initial activation processing unit is recorded as an initial value; The activation step count of the indirect activation processing unit is recorded as the step value carried by the activation indication signal plus one.

6. The distributed information processing method according to claim 5, characterized in that, The propagation of the activation indication signal satisfies a preset termination condition, which includes: reaching a preset maximum number of steps, or no new processing unit being activated.

7. The distributed information processing method according to claim 1, characterized in that, The decision weights are calculated using a monotonically decreasing function, where the independent variable of the monotonically decreasing function is the propagation distance represented by the propagation distance information.

8. The distributed information processing method according to claim 1, characterized in that, The step of distributing the input data to all processing units includes: The input data is simultaneously sent to all processing units within the same clock cycle via a global broadcast bus.

9. The distributed information processing method according to claim 8, characterized in that, The global broadcast bus can be an H-tree, grid broadcast, wireless broadcast, or optical broadcast structure.

10. The distributed information processing method according to claim 1, characterized in that, The reference data includes at least one of the following: prototype vector, hash value, threshold range, or statistical feature.

11. The distributed information processing method according to claim 1, characterized in that, The processing unit includes a non-volatile memory for permanently storing the reference data.

12. A distributed information processing system, characterized in that, include: Multiple processing units are organized into at least one processing group, wherein the maximum number of communication steps between any two processing units within the same processing group is a preset constant; A global broadcast bus is configured to distribute the input data to all processing units in response to the acquisition of input data; Each processing unit includes: Reference data storage, configured to store reference data locally; A matching degree calculation circuit is configured to calculate the matching degree between the input data and the reference data; A threshold comparator is configured to determine whether the matching degree meets a preset condition. Step counter, configured to record the number of active steps; The neighbor list register is configured to store information about neighboring units within the processing group to which this processing unit belongs. A state machine configured to control activation, step counting, and the sending and forwarding of activation indication signals; The voting accumulator is configured to aggregate the category information and activation steps of all activated processing units, calculate the decision weight based on the activation steps, and output the processing result. The smaller the activation steps, the higher the decision weight.

13. The distributed information processing system according to claim 12, characterized in that, The preset constant is 2.

14. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the distributed information processing method as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the distributed information processing method as described in any one of claims 1 to 11.

Images