Distributed computing and communication system with brain-like consciousness and implementation method thereof

By introducing brain-like consciousness state management module and efficient communication protocol in distributed computing and communication systems, the problems of low resource utilization, insufficient computing efficiency and poor robustness of traditional systems are solved, and independent learning and adaptability are achieved, which improves the ability to achieve The overall performance and intelligence level of the system.

CN118972387AActive Publication Date: 2024-11-15SHANGHAI JIAOTONG UNIV

Patent Information

Application Number
CN202411108118.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-15
Estimated Expiration
2044-08-13

AI Technical Summary

Technical Problem

When facing dynamic environments and complex tasks, traditional distributed computing and communication systems have low resource utilization, insufficient computing efficiency, poor system robustness, and lack independent learning and adaptability.

Method used

Adopting a distributed computing and communication system with brain-like consciousness, by configuring a state management module for consciousness on each computing node, combining efficient communication protocols and distributed state control, the independent learning and adaptation between computing nodes is realized, and the computing strategy is dynamically adjusted. , improve resource utilization and computing efficiency, and enhance system robustness.

Benefits of technology

It significantly improves resource utilization, computing efficiency and system robustness, realizes flexible response and self-regulation to dynamic environments, and improves the intelligence level of the overall system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118972387A_ABST
    Figure CN118972387A_ABST
Patent Text Reader

Abstract

The invention provides a distributed computing and communication system with brain-like consciousness and an implementation method thereof. The distributed computing and communication system comprises a plurality of computing nodes and a consciousness state management module, each computing node comprises a computing unit, a storage unit and a communication unit; the communication unit integrally forms an on-chip routing network NoC; the computing unit and the storage unit are connected with nodes of an on-chip routing network NoC, and collaboration and data sharing among the computing nodes are completed through an efficient communication protocol; a consciousness state management module is configured on each computing node to form a CTM consciousness management system, namely a CTM brain model, so that the nodes cooperate with each other to dynamically adjust a computing strategy to complete a complex computing task. Each computing node is provided with a consciousness state management module, the consciousness state management module can sense the real-time state of the node and receive input from other nodes, the modules can adjust the computing strategy of the node in real time by analyzing the data, and formation and dynamic adjustment of consciousness are achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of distributed computing and communication technology, and in particular, to a distributed computing and communication system with brain-like consciousness and an implementation method thereof. Background Art

[0002] Distributed computing and communication systems play a key role in modern information technology. Through parallel computing of numerous computing node clusters, they can significantly improve the system computing power and are widely used in big data processing, cloud computing, artificial intelligence training, and distributed database management. Traditional distributed systems usually rely on fixed task allocation and resource management strategies. Although they can improve computing efficiency to a certain extent, they still have the following shortcomings when facing dynamic environments and complex tasks:

[0003] 1. Low resource utilization: Due to the lack of flexibility in task and resource allocation, it is difficult for the system to achieve optimal resource configuration under different task loads, resulting in low resource utilization.

[0004] 2. Insufficient computing efficiency: When processing large-scale data and complex computing tasks, traditional distributed systems often cannot fully utilize the capabilities of each computing node due to the lack of an efficient coordination mechanism, affecting the overall computing efficiency.

[0005] 3. Poor system robustness: The existing system lacks effective self-adjustment and recovery mechanisms when dealing with uncertainties such as node failures and network fluctuations, resulting in insufficient system stability and robustness.

[0006] 4. Lack of autonomous learning and adaptability: Traditional distributed systems lack perception and learning capabilities and are unable to dynamically adjust computing strategies based on real-time data and environmental changes, limiting the intelligence level of the system.

[0007] In order to overcome the above problems, a new research direction is to propose a distributed computing and communication system with consciousness and self-regulation capabilities. As a new computing model, the Conscious Turing Machine (CTM) introduces a consciousness state management module to enable computing nodes to perceive their own state and its role in the system, thereby achieving autonomous learning and adaptation. The application of the CTM brain-like model is expected to significantly improve the resource utilization, computing efficiency and system robustness of distributed systems. Summary of the invention

[0008] In view of the defects in the prior art, the object of the present invention is to provide a distributed computing and communication system with brain-like consciousness and its implementation method.

[0009] A distributed computing and communication system with brain-like consciousness provided according to the present invention includes: a plurality of computing nodes and a consciousness state management module;

[0010] Each computing node includes a computing unit, a storage unit and a communication unit; the communication unit as a whole constitutes an on-chip routing network NoC; the computing unit and the storage unit are connected to the nodes of the on-chip routing network NoC, and the collaboration and data sharing between the computing nodes are completed through the communication protocol;

[0011] In addition, the consciousness state management module is configured on each computing node to form the consciousness management system of CTM, that is, the CTM brain-like model, so that each node can cooperate with each other to dynamically adjust the computing strategy to complete the computing task.

[0012] Preferably, the computing node includes a long-term memory module LTM or a short-term memory module STM defined in the CTM;

[0013] The on-chip routing network is a hierarchical multi-layer routing structure, including a hybrid hierarchical routing architecture based on Torus and Tree.

[0014] Preferably, the node of the on-chip routing network NoC refers to the Local end node of the router Router, which is the basic component unit of the NoC. The other four nodes of the router, namely the four nodes in the east, west, south and north, are connected to other routers respectively. Adjacent routers realize bidirectional multi-channel communication, and the whole constitutes a distributed parallel on-chip routing communication network.

[0015] Preferably, the communication protocols include asynchronous communication protocols of message queues, point-to-point communication protocols, multicast communication protocols and dynamic compression coding.

[0016] Preferably, the computing unit is used to perform distributed computing tasks, including neural network inference computing for independent operation on an autonomous robot;

[0017] The storage unit is used to store computing data and status information. Each computing node has its corresponding storage unit, which can quickly read data.

[0018] The communication unit is used to exchange data with each router node of the NoC.

[0019] Preferably, the consciousness state management module is used to implement the dynamic adjustment function of the CTM, perceive and adjust the state of the computing node and the task processing strategy.

[0020] Preferably, the consciousness state management module includes: a perception unit, a decision unit and a feedback unit;

[0021] The sensing unit is used to collect and analyze real-time status data of computing nodes;

[0022] The decision unit is used to adjust the task processing strategy of the computing node according to the data provided by the perception unit;

[0023] The feedback unit is used to record and evaluate the effect of the adjustment and further optimize the decision-making process.

[0024] Preferably, the consciousness state management module can autonomously learn and adjust computing strategies according to real-time data and environmental changes, thereby optimizing system performance;

[0025] The on-chip routing network can not only realize the uplink and downlink unicast or multicast communication required for communication between consciousness state management modules, but also realize fast dark channel connection between modules.

[0026] Preferably, the distributed computing and communication system adopts distributed state control, and determines the consistency and data integrity of each computing node when processing distributed tasks based on distributed state consistency and resource consistency.

[0027] According to a method for implementing a distributed computing and communication system with brain-like consciousness provided by the present invention, the distributed computing and communication system with brain-like consciousness is adopted, and the implementation method includes the following steps:

[0028] Step S1: Build an on-chip routing network of corresponding scale according to the number of computing nodes and the task scale;

[0029] Step S2: Connect the computing unit and the storage unit to the NoC routing node to enable parallel communication between the computing units, and realize collaboration and data sharing between the computing nodes through an efficient communication protocol;

[0030] Step S3: A consciousness state management module is configured on each computing node, forming a consciousness management system of the CTM as a whole;

[0031] Step S4: Utilize the consciousness state management module to combine historical data, current tasks and expected goals to dynamically adjust the computing unit weight information to complete the computing task.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. Each computing node of the present invention is equipped with a consciousness state management module, which can perceive the real-time state of the node and accept input from other nodes. By analyzing these data, the module can adjust the computing strategy of the node in real time to realize the formation and dynamic adjustment of "consciousness".

[0034] 2. The present invention adopts efficient communication protocols, including a multi-level converged topology on-chip routing network, an asynchronous communication protocol based on a message queue, a point-to-point unicast communication protocol, a point-to-multipoint broadcast communication and a dynamic compression encoding protocol, to achieve fast data exchange and collaborative work between CTM computing nodes, and reduce communication delays and overhead.

[0035] 3. The present invention adopts resource lock control to control the concurrent access of distributed computing nodes to shared resources; and realizes high-performance distributed scheduling by analyzing the dependency graph between computing tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0037] Figure 1 It is a schematic diagram of the overall structure of the distributed computing and communication system of the present invention;

[0038] Figure 2 It is a schematic diagram of the structure of a single computing node; it shows the computing unit, storage unit, and communication unit.

[0039] Figure 3 It is a schematic diagram of the internal structure of the consciousness state management module;

[0040] Figure 4 It is a flow chart of communication between computing nodes in the system of the present invention; it demonstrates an efficient communication protocol.

[0041] Figure 5 It is a distributed resource lock control mechanism;

[0042] Figure 6 It is a task dependency graph analysis diagram;

[0043] Figure 7 It is a schematic diagram of the overall operation of the consciousness management system. DETAILED DESCRIPTION

[0044] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0045] The distributed system of the present invention is composed of multiple parallel computing nodes, each of which includes a computing unit, a storage unit, a communication unit, and a consciousness state management module. The nodes cooperate with each other to complete complex computing tasks in parallel. Through an efficient communication protocol, the distributed system can achieve seamless collaboration and data sharing between computing nodes. Each node has autonomous learning and adaptation capabilities, and can adjust computing strategies according to real-time data and environmental changes. Specifically, the present invention is based on the CTM brain-like model, and by introducing a consciousness state management method, the entire system can achieve distributed management and computing characteristics similar to the brain, so that each computing node in the distributed system has the ability to perceive, self-adjust and optimize. Consciousness state management combines historical data, current tasks and expected goals, and dynamically adjusts the computing model, thereby improving the perception ability and robustness of the system. In addition, the system ensures the efficiency of parallel distributed computing through an efficient on-chip network communication system, and ensures the consistency and data integrity of each node when processing distributed tasks through distributed state control and task dependency graph management.

[0046] According to the present invention, a distributed computing and communication system with brain-like consciousness is provided, such as Figure 1 As shown, it includes: multiple computing nodes and a consciousness state management module. Each computing node includes a computing unit, a storage unit and a communication unit; the communication unit as a whole constitutes an on-chip routing network NoC, Network On Chip; the computing unit, the storage unit and the node of the on-chip routing network NoC are connected to complete the collaboration and data sharing between computing nodes through an efficient communication protocol; and the consciousness state management module is configured on each computing node to constitute the consciousness management system of CTM, so that each node can cooperate with each other to dynamically adjust the computing strategy to complete complex computing tasks.

[0047] Wherein, the computing node is a long-term memory module LTM or a short-term memory module STM defined in CTM (Conscious Turing Machine). The on-chip routing network adopts a 2D Mesh topology, and realizes lower communication delay and QoS management mechanism based on virtual channels. The node of the on-chip routing network NoC refers to the Local end node of the router Router, the basic component unit of NoC, and the other four nodes of the router, namely the four nodes in the east, west, south and north, are connected to other routers respectively, and adjacent routers realize two-way multi-channel communication, forming a distributed parallel on-chip routing communication network as a whole. The efficient communication protocol includes asynchronous communication protocol, point-to-point communication protocol, multicast communication protocol and dynamic compression coding technology of message queue.

[0048] The computing unit is used to perform distributed computing tasks, which may be independently operated neural network reasoning calculations on autonomous robots; the storage unit is used to store computing data and status information, and each computing node has its corresponding storage unit to realize fast data reading, presenting an overall "near memory computing" operation mode; the communication unit is used to exchange data with other nodes, specifically each router node of the NoC; the consciousness state management module is used to realize the dynamic adjustment function of CTM, perceive and adjust the state and task processing strategy of the computing node. The consciousness state management module includes: a perception unit, a decision unit and a feedback unit. The perception unit is used to collect and analyze the real-time status data of the computing node; the decision unit is used to adjust the task processing strategy of the computing node according to the data provided by the perception unit; the feedback unit is used to record and evaluate the effect of the adjustment, and further optimize the decision-making process. The consciousness state management module can autonomously learn and adjust the computing strategy according to real-time data and environmental changes, thereby optimizing system performance. The on-chip routing network can not only realize the uplink and downlink unicast or multicast communication required for communication between consciousness state management modules, but also realize the fast connection of dark channels between modules.

[0049] The distributed computing and communication system with brain-like consciousness described in the present invention adopts distributed state control to ensure the consistency and data integrity of each computing node when processing distributed tasks from the aspects of distributed state consistency and resource consistency.

[0050] The method for implementing the distributed computing and communication system with brain-like consciousness provided by the present invention comprises the following steps:

[0051] Step S1: Build a Network On Chip (NoC) of corresponding scale according to the number of computing nodes (task scale).

[0052] Step S2: Connect the computing unit and the storage unit to the NoC routing node to realize parallel communication between the computing units, and realize collaboration and data sharing between the computing nodes through an efficient communication protocol.

[0053] Step S3: A consciousness state management module is configured on each computing node, which constitutes the consciousness management system of the CTM as a whole.

[0054] Step S4: Utilize the consciousness state management module in combination with historical data, current tasks and expected goals to dynamically adjust the calculation module weight information to complete the calculation task.

[0055] The core of the present invention is to use the consciousness state management module to perceive the real-time state of the computing node and its position and role in the system. Combining historical data, current tasks and expected goals, the consciousness state management module can dynamically adjust the computing module. By adopting efficient communication protocols and distributed resource lock control, coordination and data consistency between computing nodes are ensured.

[0056] Furthermore, the distributed computing and communication system with brain-like consciousness and its implementation method of the present invention are described in detail as follows in conjunction with the accompanying drawings:

[0057] like Figure 1 As shown in FIG. 1 , the overall architecture of the distributed computing and communication system with brain-like consciousness of the present invention is shown. The system includes multiple computing nodes, which are divided into STM and LTM. Each node has a computing unit, a storage unit, a communication unit and a consciousness management module. The communication unit as a whole constitutes an on-chip routing network NoC, such as Figure 1 As shown on the right, NoC forms a communication connection between computing nodes to achieve data exchange and task collaboration. Specifically, the on-chip routing network implements a hierarchical parallel distributed architecture. In order to improve the scalability of the hierarchical interconnection architecture, the present invention designs a hybrid hierarchical on-chip network routing architecture, which adopts a hierarchical multi-layer routing structure, mainly based on the hybrid hierarchical routing architecture of Torus and Tree. The hierarchical architecture can be customized and expanded into multiple layers, in which the top layer adopts a bidirectional Torus structure to achieve the scalable flexibility of routing, and each Torus route corresponds to a low-level partition; the lower layer to the bottom layer adopts a Tree structure to interconnect more computing cores to achieve more efficient and convenient communication speed and obtain higher bandwidth. In the specific design, the communication node in the hierarchical architecture is Tile, and each Tile contains the upper-layer Torus route, the Tree route connected to it, and the four computing cores connected to the Tree route. The upper-layer Torus route communicates with different Torus routes through the four ports of East, West, South, and North (EWSN), and then communicates with the underlying Tree route through the Local end node. The four ports of East, West, South, and North of Tree communicate with the computing core.

[0058] The single computing node architecture of the present invention is as follows: Figure 2 and Figure 3 As shown in the figure, the computing, storage, communication unit and consciousness state management module in a single computing node are respectively shown. The computing unit mainly completes the reasoning operation of the neural network, that is, vector-matrix multiplication, and performs specific tasks on the autonomous robot, such as face recognition, map navigation, etc. The storage unit stores the weight information data of the neural network and the communication information Chunk that the computing unit needs to transmit to other units, which is encoded as

[0059] <address,t,gist,weight,intensity>

[0060] Among them, address is the address of the LTM processor that generates the information block, t is the time when the block is generated, gist is the brain language information encoded in the LTM module, and is also the core communication information for realizing the alignment of multiple LTM modules, weight is the weight of the processor to measure the importance of the current brain language, and intensity is used to record the cumulative amount of weights in the competing communication path. The initial time t0 is |ht| and weight respectively. In the consciousness management system based on CTM, external information with different characteristics is processed by different LTM modules and shared among LTM modules. The communication unit is the basic component of NoC, Router, which adopts the wormhole routing mechanism to divide the complete data packet Packet into multiple Flits and transmit them separately. When the computing node generates the data packet to be sent, it transmits it to the connected Router node and encodes the corresponding target location information in the data. The Flit data needs to make a routing decision in each Router according to the current location and target location to determine its output port in the current Router. The NoC of the present invention adopts multiple virtual channels. Flits of multiple channels may compete for the same downstream output port, and Round_Robin arbitration is required. The data that succeeds in arbitration can be allocated to the output port and the downstream virtual channel. The entire communication is executed cyclically according to the above process until the complete Packet arrives at the destination node.

[0061] The composition of the consciousness state management module is as follows: Figure 3 As shown in Figure 1, they have different configurations, and different LTM computing nodes have different capabilities. Specifically, Figure 3 (a) shows "perceptual coding LTM" - obtaining information from the outside world and converting it into "brain language" through coding, such as Figure 3 (b) "Information Storage and Recall LTM" - transforms the input data according to the stored information and outputs the transformation result; Figure 3 (c) shows the "behavior output LTM" - output action instructions based on input data. There is great flexibility in the specific implementation form, for example: multimodal image-text association network (CLIP) can be used in visual perception LTM to implement image encoding; language large model can be used in "storage recall" LTM to realize the function of calculating output "brain language" from input "brain language", etc.

[0062] The efficient communication protocols adopted by the present invention include asynchronous communication protocols, unicast and multicast communications, and dynamic compression coding technologies. Specifically, the asynchronous communication protocol is described as follows: Due to the influence of layered multi-cores, the communication between distributed computing cores is highly dependent on the clock, and the clock synchronization problem needs to be considered. In response to this situation, the present invention adopts a global asynchronous local synchronous (GALS) data communication solution, and there are mainly three methods to solve the problem of secure and reliable data transmission between independent clock domains: Pausible-clock generators, asynchronous FIFO, and boundary synchronization. The present invention divides the distributed on-chip routing network architecture into multiple local synchronization domains, and uses asynchronous FIFOs for local synchronization between different domains, eliminating clock coupling between multi-cores, reducing the difficulty of interconnection and expansion of the system, and getting rid of the strict clock synchronization requirements between chips, such as Figure 4 (a) As shown. For FIFOs with unlimited depth, the present invention uses a loop counting method to determine whether the FIFO memory is empty or full, and considers the asynchronous synchronization problem at the same time, requiring that the FIFO code changes from one value to another adjacent value each time, and only one bit changes. To meet the current conditions, the present invention uses Gray code to re-encode the read and write addresses to ensure that only one adjacent bit is different when the value changes. The operation of converting conventional binary code to Gray code is as follows:

[0063]

[0064] By converting binary to Gray code and introducing a higher bit for judgment, loop counting is implemented to judge whether the FIFO is empty or full. If the higher bits are the same, it means that no loop is performed, and the write address and read address are directly judged. If the difference is the FIFO depth, it means that the FIFO is full, and if they are equal, it means that the FIFO is empty. By using a global asynchronous and locally synchronous communication solution, the difficulty of interconnection and expansion of distributed systems can be reduced, and the strict clock synchronization requirements between distributed computing units can be eliminated, avoiding system function failure caused by errors in some clock domains.

[0065] In terms of unicast and multicast communications, the communication of data packets in NoC is mainly carried out in a serial manner. For each routing node, only one data flit is allowed to pass through in one clock cycle. The advantage of this is to better control the data flow and avoid data packet confusion errors during the transmission process as much as possible, while saving resource overhead. This data transmission method is called data unicast. As the most basic function of the on-chip network, data unicast can meet most traffic conditions. However, in the distributed communication system based on CTM, there is a broadcast communication from STM to all LTMs. Therefore, the present invention considers supporting multicast communication and transmitting data packets from the same source node to different destination nodes in parallel. In more detail, for each routing node, data flits are allowed to be copied and forwarded to multiple output ports in the same clock cycle. This data propagation method can reduce the delay of data transmission and improve transmission efficiency, thereby improving the overall communication performance of NoC. Traditional multicast is divided into two categories: path-based data multicast and topology-based data multicast. At present, there are also algorithms that combine the two traditional methods and use the idea of ​​partitioning for data multicast, such as Figure 4 The present invention takes into account the flexibility of the top-level Torus architecture and hopes to take advantage of the shortest communication path as much as possible. The multicast communication supported in the architecture is carried out in a topology-based manner. Figure 4 As shown in (b), path-based data multicast searches for Hamiltonian circuits in the topology, hoping to connect all routes with one path. This method first needs to determine a suitable Hamiltonian circuit and sort the routing nodes according to the path. The data packets that need to be multicast are copied and forwarded according to the sequence number of the destination node and the sequence number of the source node, traversing the entire path until all destination nodes receive the data packets. The advantage of this method is that it avoids deadlock, and the data packets forward data in ascending or descending order according to the node sequence number of the Hamiltonian circuit. The present invention modifies the data packet format, the arbitrator arbitration method, and the port output regulations, and is designed to copy and forward data at the Torus routing node that needs to branch, and forward it to the corresponding distributed computing unit at the lower tree routing at the same time. While supporting multicast communication, NoC also supports the original unicast communication, and changes the multicast address by setting the corresponding flag bit mask, thereby achieving simultaneous support for unicast and multicast.

[0066] In terms of dynamic compression coding protocols, by selecting appropriate data compression coding algorithms, the communication bandwidth pressure of NoC can be alleviated to a certain extent. In data-intensive and communication-intensive applications, cache misses or long-term blocking of multiple data packets will put tremendous pressure on the underlying network, thereby preventing data packets from being transmitted from multiple cache hierarchies to the requesting core, reducing the overall forwarding efficiency of NoC. In order to make the best use of the limited resources of NoC, it is necessary to save resource overhead by compressing pulses. The information transmitted between LTMs has a similar encoding format, the data content gap is small, and the dynamic floating range of data values ​​is limited, so fine-grained differential data compression can be performed. Figure 4 (c) shows a data packet with 5 Flits and a total of 20 bytes. After differential compression, the data packet can be compressed to 12 bytes without changing the header and base Flits, which can effectively reduce the pressure on bandwidth resources in the case of large-scale data. Fine-grained differential compression uses the FlitZip algorithm, which is hardware-friendly. The algorithm divides the data Flits into several sub-blocks, determines the base block in the sub-block as the differential basis, uses several subtractors to perform differential operations to obtain the encoded sub-blocks, and integrates them to form a compressed data packet with a higher compression ratio. Decoding uses several adders to perform the reverse operation. FlitZip can effectively reduce data Flit data, increase the compression ratio, and reduce bandwidth requirements and forwarding delays.

[0067] In a distributed computing and communication architecture, when multiple scheduling units perform task scheduling, resource conflicts and state control failures often occur. Distributed control is responsible for ensuring the scheduling state and resource coordination consistency state of the distributed system, and ensuring resource coordination and correct task submission during the distributed module computing process. At the same time, the present invention uses coarse-grained time steps as decision-making and learning task operation cycles, and controls the operating state of the overall system and the LTM task thread by setting a timer.

[0068] The following describes distributed state consistency and distributed resource consistency respectively:

[0069] In terms of distributed state consistency, decision-making and learning tasks are completed by multiple modules in a distributed manner. This process is often called a distributed transaction. The distributed transaction adopted by the present invention provides a mechanism to incorporate the task sequence involving module collaboration into an execution control unit. The transaction can only be committed if all operations are executed correctly; any failure of any operation will cause the entire transaction to roll back to the previous state, that is, all operations are canceled.

[0070] In terms of distributed resource consistency, the distributed lock adopted by the present invention is deployed in multiple LTM threads in a distributed environment to achieve multi-threaded distributed mutual exclusion, such as Figure 5 As shown. In order to achieve distributed mutual exclusion, consider setting up a distributed lock monitor in the distributed system. Each thread can see the resource lock status. When the resource lock does not exist, the mark can be set. When the resource lock is set, other threads can only wait for the thread that owns the resource lock to complete execution and release the resource lock before setting the resource lock and accessing the shared resource. The resource lock is a mark that allows threads to access shared resources without interfering with each other in the scenario where multiple threads access the same resource at the same time, thereby ensuring the validity and correctness of the operation. At the same time, in order to ensure that multiple threads can see the resource lock, the resource lock is stored in a common storage to enable multiple threads to concurrently access the same critical resource. Only one thread can access the shared resource at the same time to ensure data consistency during the operation of the distributed system.

[0071] The present invention intends to adopt distributed multi-module computing technology to adapt to the execution mode of the CTM brain-like model, automatically generate a communication graph model between multiple modules according to the decision learning logic, and support distributed parallel computing and single-cast and multicast communications in the brain-like behavior decision-making architecture.

[0072] In traditional distributed computing, a computing task is generally called a Job. A Job has multiple Tasks, and each Task can be run independently (cross-machine) or collaboratively (single-machine) on a set of servers. The number of Tasks a Job is split into determines the parallelism (Granularity or Parallel) of the distributed computing. In the distributed system of the present invention, a single decision or learning task can be regarded as a Job, and the Job is jointly participated by multiple LTM modules, that is, the LTM module is regarded as a Task. Therefore, in the distributed system, multiple Jobs (i.e., multiple decision and learning tasks) are running at the same time, and each Job is completed by multiple Tasks in a collaborative and distributed manner. The core of distributed computing is to maintain the topological diagram of the logical relationship between Jobs and Tasks, and to allow these tasks to be run in parallel, and it is also necessary to minimize the communication and data access overhead brought by the distributed network. Utilizing the single broadcast communication mechanism between multiple LTM modules, the present invention implements a distributed computing system based on multiple LTM modules (i.e., Tasks) through NoC. The execution of each module may be heterogeneous. In order to ensure performance, distributed computing technology needs to schedule tasks to all nodes as much as possible and avoid slow performance problems caused by the "barrel effect".

[0073] The distributed computing architecture used for this purpose automatically associates the corresponding functional modules according to the decision learning relationship between multiple modules, generates task labels, and generates a directed acyclic graph (DAG) according to the dependency relationship between tasks, such as Figure 6As shown, each node T0-T6 is a basic computing Task (such as LTM module brain language information generation), where Job 1 (such as STM generation and action LTM decision or learning) includes tasks T0, T1, T4, T5, and Job2 includes tasks T1, T2, T3, T6. Each directed edge represents a dependency. In this example, tasks T0-T3 can be executed simultaneously, with a maximum parallelism of 4, which means that 4 nodes can execute tasks at the same time. Task T4 needs to wait for the completion of the two previous subtasks T0 and T1 before it can start computing, and T5 needs to wait for the completion of tasks T4 and T2 before it can start computing.

[0074] In this computing mode, the scheduling unit is responsible for generating a job into a series of tasks and storing them in the task pool according to their task dependencies. When it comes to related tasks, it allocates enable signals, and the corresponding modules perform calculations and return results. Multiple jobs are distributed and scheduled by different scheduling units. When the same task or task conflict is detected, the resource lock mechanism is automatically started to wait, so as to complete the related computing jobs in a distributed manner, that is, the computing tasks of the computing units in the distributed system.

[0075] In addition, the dynamic awareness management mechanism based on CTM, such as Figure 1 The left side shows the overall architecture of the CTM brain-like model, where STM is a short-term memory module; LTM is a long-term memory module; Up Tree is the up-tree connecting LTM to STM; DownTree is the down-tree connecting STM to LTM; Links is a dark channel for direct communication established between LTMs in the form of associative learning; Input obtains environmental information collected by sensors and sends it to the specified LTM processor, and Output transmits the behavior output, that is, the decision result, to the actuator to act on the environment; Chunk is the information block transmitted between STM and each LTM, and the core communication information is collectively referred to as "brain language". Each LTM corresponds to different functional divisions of the brain and realizes different functions. Multiple LTMs transmit multiple "brain language" information to STM through the up-tree; when the information correlation between two LTMs is high enough, a dark channel will be established to skip the complex competition process and communicate directly; the semantic consistency of "brain language" between different LTMs is achieved through the "cooperative learning" mechanism. Based on the "competitive decision-making" mechanism, STM decides which LTM's "brain language" to publish from the information received from several LTMs. The published "brain language" is defined as "brain-like consciousness". The brain language is then broadcast to each LTM through the descending tree, thus forming a feedback loop and forming a continuous content update on STM. The operation of this model is as follows Figure 7(a) As shown in Figure 2. At each moment, all the “brain words” published on the STM are regarded as “state vectors”. The information on the STM is dynamically updated over time, forming a “chain” of “state” migration, as shown in Figure 2. Figure 7 (b) is shown. The change of "state" can be regarded as a Markov chain, that is, all LTMs take the current state as input, and their output determines the next state. Based on the semantic coding representation of "brain-like consciousness", an interpretable "brain-like consciousness stream" can be output, and then the content scope of the "brain language" published on STM can be limited through additional constraints to achieve the controllability of "brain-like consciousness".

[0076] The present invention aims to improve the computing efficiency, resource utilization and robustness of distributed systems, and is suitable for application scenarios such as autonomous robots and artificial intelligence devices. The brain-like distributed computing and communication system of the present invention is suitable for processing large-scale complex tasks, such as autonomous robots, artificial intelligence devices, etc. The present invention realizes high-performance computing through parallel distributed computing and communication, and realizes efficient collaboration and self-regulation capabilities between nodes through CTM, which can significantly improve the computing efficiency of the overall system, reduce computing delays, and enhance the flexibility and adaptability of the system.

[0077] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0078] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A distributed computing and communication system with brain-like consciousness, characterized in that: include: Multiple computing nodes and consciousness state management modules; Each computing node includes a computing unit, a storage unit and a communication unit; the communication unit as a whole constitutes an on-chip routing network NoC; the computing unit and the storage unit are connected to the nodes of the on-chip routing network NoC, and the collaboration and data sharing between the computing nodes are completed through the communication protocol; In addition, the consciousness state management module is configured on each computing node to form the consciousness management system of CTM, that is, the CTM brain-like model, so that each node can cooperate with each other to dynamically adjust the computing strategy to complete the computing task.

2. The distributed computing and communication system with brain-like consciousness according to claim 1, characterized in that: The computing node includes a long-term memory module LTM or a short-term memory module STM defined in the CTM; The on-chip routing network is a hierarchical multi-layer routing structure, including a hybrid hierarchical routing architecture based on Torus and Tree.

3. The distributed computing and communication system with brain-like consciousness according to claim 1, characterized in that: The node of the on-chip routing network NoC refers to the Local end node of the router Router, which is the basic component unit of NoC. The other four nodes of the router, namely the four nodes in the east, west, south and north, are connected to other routers respectively. Adjacent routers realize bidirectional multi-channel communication, and the whole constitutes a distributed parallel on-chip routing communication network.

4. The distributed computing and communication system with brain-like consciousness according to claim 1, characterized in that: The communication protocols include asynchronous communication protocols of message queues, point-to-point communication protocols, multicast communication protocols and dynamic compression coding.

5. The distributed computing and communication system with brain-like consciousness according to claim 1, characterized in that: The computing unit is used to perform distributed computing tasks, including neural network inference computing for independent operation on the autonomous robot; The storage unit is used to store computing data and status information. Each computing node has its corresponding storage unit, which can quickly read data. The communication unit is used to exchange data with each router node of the NoC.

6. The distributed computing and communication system with brain-like consciousness according to claim 1, characterized in that: The consciousness state management module is used to implement the dynamic adjustment function of the CTM, perceive and adjust the state of the computing node and the task processing strategy.

7. The distributed computing and communication system with brain-like consciousness according to claim 6, characterized in that: The consciousness state management module includes: a perception unit, a decision unit and a feedback unit; The sensing unit is used to collect and analyze real-time status data of computing nodes; The decision unit is used to adjust the task processing strategy of the computing node according to the data provided by the perception unit; The feedback unit is used to record and evaluate the effect of the adjustment and further optimize the decision-making process.

8. The distributed computing and communication system with brain-like consciousness according to claim 1, characterized in that: The consciousness state management module can autonomously learn and adjust computing strategies according to real-time data and environmental changes, thereby optimizing system performance; The on-chip routing network can not only realize the uplink and downlink unicast or multicast communication required for communication between consciousness state management modules, but also realize fast dark channel connection between modules.

9. The distributed computing and communication system with brain-like consciousness according to claim 1, characterized in that: The distributed computing and communication system adopts distributed state control, and determines the consistency and data integrity of each computing node when processing distributed tasks based on the distributed state consistency and resource consistency.

10. A method for implementing a distributed computing and communication system with brain-like consciousness, characterized in that: The distributed computing and communication system with brain-like consciousness as claimed in any one of claims 1 to 9 is adopted, and the implementation method includes: Step S1: Build an on-chip routing network of corresponding scale according to the number of computing nodes and the task scale; Step S2: Connect the computing unit and the storage unit to the NoC routing node to enable parallel communication between the computing units, and realize collaboration and data sharing between the computing nodes through an efficient communication protocol; Step S3: A consciousness state management module is configured on each computing node, forming a consciousness management system of the CTM as a whole; Step S4: Utilize the consciousness state management module to combine historical data, current tasks and expected goals to dynamically adjust the computing unit weight information to complete the computing task.

Citation Information

Patent Citations

  • Chip expansion method for hundred million-level neuron brain-like computer

    CN112269751A

  • Network-on-chip routing communication method for brain-like processor and network-on-chip

    CN112468401A

  • Distributed cloud computing system and method based on human brain simulation

    CN112698944A

  • Situation awareness method based on brain-like calculation

    CN113242267A

  • System using planning information to modify operation of a digital network

    US20050223088A1

Cited By

  • Brain-like behavior decision calculation model of autonomous robot

    CN119692382A