A transmission control and intelligent scheduling system for integrated chips

Abstract

The application relates to the technical field of integrated circuits and computer networks, and particularly discloses a transmission control and intelligent scheduling system for an integrated chip, which is provided with a dual-mode decision and dynamic switching mechanism for each routing node; the node continuously monitors a local running state and calculates dynamic characteristic parameters containing instantaneous values, first-order trends and second-order accelerations; when the parameters match pre-stored abnormal characteristic sets and the load exceeds a threshold value, the node is immediately atomized and switched to a predefined safe scheduling strategy loaded from a shared storage area; otherwise, the node generates a scheduling decision according to adaptive decision logic continuously optimized based on historical performance feedback; all the nodes perform data packet forwarding control according to their current execution strategies; and the system also periodically realizes federated global knowledge evolution according to quality evaluation results of adaptive decisions of the nodes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of integrated circuits and computer network technology, in particular to a transmission control and intelligent scheduling system for integrated chips. BACKGROUND

[0002] With the continuous progress of integrated circuit technology, the scale and complexity of integrated chips are constantly improving, and network-on-chip has become a key interconnection infrastructure for connecting numerous processor cores, memory units and special accelerators. In order to cope with dynamic and variable application loads and optimize overall performance, intelligent scheduling methods based on machine learning, especially reinforcement learning, are introduced into the routing and flow control of network-on-chip. These methods allow each network node (router) to learn and optimize data forwarding decisions based on local observation history, aiming to reduce communication delay and improve throughput. However, this distributed, locally experience-driven learning paradigm has inherent flaws: without effective global coordination and stabilization mechanisms, the independent exploratory decisions of individual agents (routers) may produce unpredictable synergistic effects under extreme or non-stationary loads.

[0003] The present application aims to solve the problem of global collapse caused by the coordination disorder of individual agents in the existing distributed machine learning-based network-on-chip transmission control method. This problem is manifested as follows: when the network load changes dramatically, the intelligent routing decisions made by individual nodes based on local historical experience, aiming to optimize long-term statistical indicators (such as average delay), may produce conflicting or resonant effects due to the lack of global transient state awareness and coordination, leading to ineffective oscillation and convergence of traffic on limited network resources, and triggering self-enhancing congestion or deadlock in the entire network, resulting in a sharp decline in network throughput or even zero. Traditional methods cannot predict and prevent such chain reaction disasters caused by intelligent decision-making in a short period of time, nor can they guide the system to safely and quickly recover after the fact. Therefore, how to embed a "fusing" mechanism that can real-time perceive abnormal risks and automatically trigger deterministic protection strategies at the critical point while maintaining its global learning and evolution ability has become a core technical problem that needs to be solved in this field. SUMMARY

[0004] The purpose of the present application is to provide a transmission control and intelligent scheduling system for integrated chips to solve the problems in the above background.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A transmission control and intelligent scheduling system for integrated chips, comprising:

[0007] The state monitoring and feature extraction module is configured to collect local running state data of each routing node in the integrated chip and calculate a set of dynamic feature parameters based on the local running state data.

[0008] The policy matching and signal generation module is configured to match the dynamic feature parameters with an abnormal state feature set pre-stored in a shared storage area of the integrated chip and generate a policy selection signal according to a matching result.

[0009] The security policy loading module is configured to load a pre-defined security scheduling policy corresponding to the current matching result from the shared storage area when the policy selection signal indicates to enable security scheduling and set the security scheduling policy as a current execution policy of the corresponding routing node.

[0010] The adaptive decision generation module is configured to generate a scheduling decision according to a locally continuously optimized adaptive decision logic when the policy selection signal indicates to maintain regular scheduling and set the scheduling decision as the current execution policy of the corresponding routing node.

[0011] The data packet forwarding control module is configured to perform forwarding control on an arrived data packet according to the current execution policy of the routing node.

[0012] The global policy aggregation and update module is configured to periodically aggregate parameter update information generated based on the adaptive decision logic from each routing node and perform fusion update on global policy parameters stored in the shared storage area according to the aggregated information.

[0013] As a further scheme of the present application, the generation of the set of dynamic feature parameters specifically includes:

[0014] The current queue length of each virtual channel in the routing node is obtained as original state data.

[0015] Based on the original state data, a first-order change rate of each queue length in a preset time window is calculated to obtain a load change trend parameter.

[0016] A second-order change rate of the load change trend parameter in the time window is calculated to obtain a load acceleration state parameter.

[0017] The original state data, the load change trend parameter and the load acceleration state parameter are combined to generate the dynamic feature parameters.

[0018] As a further scheme of the present application, the generation of the policy selection signal specifically includes:

[0019] Each component of the dynamic feature parameters is compared with a preset threshold range of a corresponding component in the abnormal state feature set one by one to generate a preliminary matching state of the components.

[0020] Counting the components satisfying the matching condition in the preliminary matching state, and comparing the count value with a preset first-level quantity threshold to generate a preliminary matching signal;

[0021] Performing logical AND operation on the preliminary matching signal and a local comprehensive load level of the current routing node, wherein the local comprehensive load level is obtained by calculating the sum of the lengths of the virtual channel queues;

[0022] According to the result of the logical AND operation, a final strategy selection signal is generated; when the operation result is true, the strategy selection signal indicates to enable the safe scheduling; when the result of the logical AND operation is false, the strategy selection signal indicates to maintain the regular scheduling.

[0023] As a further scheme of the present application, the loading of the predefined safe scheduling strategy corresponding to the current matching result from the shared storage area specifically includes:

[0024] Based on the strategy selection signal and the dynamic characteristic parameters, an abnormal type identifier is generated;

[0025] According to the abnormal type identifier, a set of predefined safe scheduling basic rules and their corresponding adjustable parameter sets are indexed and read from the shared storage area;

[0026] Based on the local running state data of the current routing node, the key parameter values in the adjustable parameter set are calculated and filled in real time to generate a safe scheduling strategy instance adapted to the current time;

[0027] The safe scheduling strategy instance is set as the current execution strategy of the corresponding routing node.

[0028] As a further scheme of the present application, the generation of the abnormal type identifier specifically includes:

[0029] The preliminary abnormal level information encoded in the strategy selection signal is parsed;

[0030] According to the preliminary abnormal level information, at least two core characteristic components are selected from the dynamic characteristic parameters as key discriminators;

[0031] According to a preset weight proportion, the selected key discriminators are weighted and summed to obtain a comprehensive abnormal score;

[0032] The comprehensive abnormal score is compared with a set of pre-stored abnormal level threshold intervals, and according to the interval into which it falls, a final abnormal type identifier is determined and output.

[0033] As a further scheme of the present application, the generation of the scheduling decision according to the self-adaptive decision logic of the local continuous optimization specifically includes:

[0034] The decision execution effect data is formed by collecting link transmission delay and queue emptying rate obtained by forwarding data packets according to past scheduling decisions in a historical time window corresponding to the routing node;

[0035] The contribution degree and the deviation degree of the decision execution effect data to the preset global optimization target are calculated to generate a quality evaluation result of the current decision;

[0036] According to the quality evaluation result, an adjustment direction and amplitude instruction of a key weight parameter in the local decision rule are generated;

[0037] The key weight parameter is updated by using the adjustment instruction, and a scheduling decision is generated by using the updated decision rule in combination with the current local running state data.

[0038] As a further scheme of the application, the generation of the quality evaluation result of the current decision specifically includes:

[0039] The average value of the link transmission delay and the average value of the queue emptying rate are extracted from the decision execution effect data as a first evaluation vector;

[0040] The first evaluation vector and the preset global optimization target reference vector are compared component by component to calculate a contribution degree score reflecting the degree of achievement;

[0041] The fluctuation variance of the first evaluation vector in the preset time window is calculated, and the fluctuation variance is compared with the allowed fluctuation range to obtain a deviation distance reflecting stability;

[0042] The contribution degree score and the deviation distance are weighted and combined according to the preset contribution weight and the deviation weight to generate a comprehensive quality evaluation result.

[0043] As a further scheme of the application, the method for forwarding control of the data packet according to the current execution strategy of the routing node specifically includes:

[0044] The current execution strategy is analyzed to extract a set of output port selection rules and corresponding dynamic adjustment parameters contained therein;

[0045] According to the set of output port selection rules, the basic output port pointing is calculated in combination with the service type identifier carried by the data packet;

[0046] The basic output port pointing is dynamically corrected in priority weight according to the dynamic adjustment parameters and the real-time local load information of the routing node to generate a final output port pointing;

[0047] According to the final output port pointing, the data packet is scheduled to the corresponding output cache queue to complete the forwarding control.

[0048] As a further scheme of the present application: the global strategy parameter stored in the shared storage area is fused and updated according to the information after the aggregation, and specifically comprises:

[0049] Receiving parameter update information from each routing node, the parameter update information containing key weight parameters of decision rules and corresponding quality evaluation results;

[0050] According to the quality evaluation results attached in each node information, a dynamic credibility weight is calculated for the key weight parameters of each node;

[0051] Based on the dynamic credibility weight, the key weight parameters provided by all nodes are weighted and averaged to generate a new generation of global key weight parameter set;

[0052] The new generation of global key weight parameter set is used to replace the original global strategy parameter in the shared storage area, and the fusion update is completed.

[0053] The beneficial effects of the present application are:

[0054] (1) By continuously calculating the high-order dynamic characteristics (including instantaneous value, first derivative and second derivative) of the local state of each node, and quickly matching with the pre-stored abnormal pattern library, when the matching degree and the local load level exceed the threshold value at the same time, the node can atomically switch to the globally verified and deterministic safe scheduling strategy within nanosecond time. This integrated design of "monitoring-matching-fusing" implants a hardware-level safety guardrail for the distributed learning system, so that when the network encounters unknown or high-risk traffic patterns, it can immediately downgrade from the exploratory optimization mode to the safe mode that guarantees connectivity and basic service quality, thereby ensuring the safety of system functions and services in the worst case and greatly enhancing the robustness of complex chip networks.

[0055] (2) By constructing a federated hierarchical learning and global knowledge aggregation framework based on quality evaluation, the balance between local adaptive optimization and global experience sharing is realized, and the learning efficiency and performance of the entire chip are improved. BRIEF DESCRIPTION OF DRAWINGS

[0056] The present application will be further described below in conjunction with the accompanying drawings.

[0057] Figure 1 is a system block diagram of the present application. DETAILED DESCRIPTION

[0058] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work are within the protection scope of the present application.

[0059] Please refer to Figure 1 As shown in the drawings, the present application is a transmission control and intelligent scheduling system for integrated chips, comprising:

[0060] The state monitoring and feature extraction module is used for collecting local running state data of each routing node in the integrated chip, and calculating a set of dynamic feature parameters based on the local running state data;

[0061] The policy matching and signal generation module is used for matching the dynamic feature parameters with an abnormal state feature set pre-stored in the shared storage area of the integrated chip, and generating a policy selection signal according to the matching result;

[0062] The security policy loading module is used for loading a pre-defined security scheduling policy corresponding to the current matching result from the shared storage area when the policy selection signal indicates to enable security scheduling, and setting the security scheduling policy as the current execution policy of the corresponding routing node;

[0063] The adaptive decision generation module is used for generating a scheduling decision according to its locally continuously optimized adaptive decision logic when the policy selection signal indicates to maintain the regular scheduling, and setting the scheduling decision as the current execution policy of the corresponding routing node;

[0064] The data packet forwarding control module is used for performing forwarding control on the arrived data packet according to the current execution policy of the routing node;

[0065] The global policy aggregation update module is used for periodically aggregating parameter update information generated based on the adaptive decision logic from each routing node, and fusing and updating the global policy parameters stored in the shared storage area according to the aggregated information.

[0066] In the state monitoring and feature extraction module, first, a state monitoring and feature extraction process needs to be performed by each routing node in the integrated chip. This process aims to convert the real-time, original running state inside the routing node into a set of feature parameters that can comprehensively and deeply characterize the instantaneous dynamics of itself and the local network, for subsequent decision-making. The specific implementation steps are as follows:

[0067] The routing node acquires its local running state data. The local running state data mainly refers to the states of multiple independent queues inside the routing node for temporarily storing data packets to be forwarded. Specifically, each routing node contains multiple virtual channels, and each virtual channel corresponds to a first-in-first-out data cache queue. In this step, the routing node acquires the queue length of each virtual channel in the current clock cycle by reading the value of its internal hardware register. The queue length is represented by the number of data packets currently stored in the queue and is a non-negative integer. For example, a routing node with 8 virtual channels will acquire 8 queue length values, which together constitute the original state data.

[0068] Based on the acquired original state data, the routing node calculates a load change trend parameter. This parameter is used to quantify the change speed of the load of each queue in the short term. In specific implementation, the routing node maintains a preset length of time window, which contains 128 consecutive clock cycles. For each virtual channel, the routing node calculates the first-order change rate of the queue length value in the current time window. The calculation process is as follows: record the queue length of the queue in the current clock cycle and 127 cycles ago, subtract the two values, and then divide the difference by the number of 128 clock cycles corresponding to the time window. The result obtained is the average length change amount of the queue per clock cycle in the current window, which is the load change trend parameter of the queue. Repeat the calculation for all virtual channels to obtain a set of load change trend parameters.

[0069] The routing node further calculates a load acceleration state parameter. This parameter is used to represent the acceleration of the load change, that is, the change speed of the change trend itself. The calculation process is based on the load change trend parameters of each queue obtained in the previous step. The routing node uses the same length (128 clock cycles) of time window as in the previous step, but applies it to the sequence of load change trend parameters. For each virtual channel, calculate the difference between the current load change trend parameter value and the historical value 127 cycles ago, and then divide the difference by the number of 128 clock cycles. The result obtained is the average change rate of the load change trend of the queue in the recent time window, that is, the load acceleration state parameter. This parameter can indicate whether the queue load is accelerating, growing at a constant speed, or decelerating.

[0070] The routing node combines the three types of parameters described above to generate the final dynamic characteristic parameter. The specific way of combination is to splice the three values corresponding to each virtual channel: the current queue length (original state data), the load change trend parameter, and the load acceleration state parameter, into a numerical vector in a fixed order. For example, for the aforementioned 8-virtual-channel routing node, a characteristic vector containing 24 numerical components will be generated. This dynamic characteristic parameter integrates the instantaneous value, first-order dynamics, and second-order dynamics of the state, providing multi-dimensional and accurate quantitative input for subsequent judgment of network behavior patterns. This characteristic vector will be directly output for matching with the pre-stored abnormal state characteristic set.

[0071] In the policy matching and signal generation module, after the routing node generates the dynamic characteristic parameter, the matching and signal generation process needs to be performed. The purpose of this process is to compare the dynamic characteristic parameter representing the current network instantaneous state with the pre-defined characteristic pattern representing the known abnormal situation, and accordingly generate a clear control instruction to determine which scheduling mode should be adopted subsequently. The specific implementation steps are as follows:

[0072] Perform a one-by-one comparison of each component of the dynamic characteristic parameter. The abnormal state characteristic set is pre-stored in the shared memory area of the integrated chip, which contains multiple characteristic templates, each corresponding to a pre-defined abnormal state pattern. Each characteristic template consists of multiple threshold ranges with the same dimension as the dynamic characteristic parameter. For example, for a dynamic characteristic parameter containing 24 components, the corresponding characteristic template also contains 24 threshold ranges. When matching, the first component of the dynamic characteristic parameter is compared with the threshold range of the first component in the characteristic template. If the former falls within the numerical interval defined by the latter, it is determined that the component matches, recorded as logical "true"; otherwise, recorded as logical "false". In this way, the comparison of all 24 components is completed one by one, and finally a sequence of 24 logical values (true or false) is generated, which is the preliminary matching state of the group of components.

[0073] Count the preliminary matching state and compare it with the first-level quantity threshold to generate the primary matching signal. The specific operation is to count the number of logical values "true" in the logical value sequence obtained above. This count value reflects the degree of agreement between the current state and the specific abnormal characteristic template. The pre-set first-level quantity threshold is an integer, for example, 18. If the count value is greater than or equal to 18, it is considered that the current state preliminarily agrees with the abnormal pattern, and an effective primary matching signal (e.g., logical level "high" or numerical value "1") is generated; if the count value is less than 18, it is considered that the matching standard has not been reached, and an ineffective primary matching signal (e.g., logical level "low" or numerical value "0") is generated. This step introduces fault tolerance into the matching, allowing individual characteristic components to be unmatched due to noise fluctuations.

[0074] The primary match signal is logically ANDed with the local comprehensive load level of the current routing node. The local comprehensive load level is calculated by reading the current queue length of all virtual lanes of the routing node, and summing up these length values. A threshold value, for example 64, is set for the load level for logical judgment. When the local comprehensive load level is greater than or equal to 64, the load level condition is determined to be satisfied, and is converted to logical "true"; otherwise, it is converted to logical "false". The rule of logical AND operation is that the output result is "true" only when both inputs (validity of the primary match signal and truth of the load level condition) are "true"; the output result is "false" as long as one of the inputs is "false".

[0075] According to the Boolean result of the logical AND operation, a final strategy selection signal is generated. The signal is a binary control signal. Specifically, when the result of the logical AND operation is "true", it indicates that both the abnormal feature pattern and the heavy load pressure are detected, and the strategy selection signal generated at this time is coded as an "enable safe scheduling" instruction. When the result of the logical AND operation is "false", it indicates that at least one of the two conditions does not satisfy, and the strategy selection signal generated at this time is coded as a "maintain regular scheduling" instruction. The signal will be directly output for controlling the selection and loading of the next stage scheduling strategy.

[0076] In the safe strategy loading module, when the strategy selection signal indicates to enable safe scheduling, the routing node will perform a safe strategy loading process. The process aims to retrieve and instantiate a corresponding deterministic scheduling strategy from a pre-stored knowledge base according to the specific type and severity of the detected abnormal situation, to replace the original adaptive decision logic, thereby ensuring the reliability and stability of the network in abnormal state. The specific implementation steps are as follows:

[0077] Based on the strategy selection signal and the dynamic feature parameters, an accurate abnormal type identifier is generated. The strategy selection signal is a multi-bit digital signal. For example, a 3-bit signal can represent multiple states, in which specific bit combinations are encoded to represent the preliminary abnormal level: '001' represents primary abnormality, '010' represents intermediate abnormality, and '011' represents high-level abnormality. The routing node is internally provided with a decoding logic circuit to extract specific bits from the signal and convert them into an integer value representing the preliminary abnormal level according to a predefined encoding table, denoted as , where The value of may be 1, 2, or 3. The dynamic feature parameters are a vector containing multiple components, denoted as: According to the preliminary abnormal level , the core feature component index that needs to be focused on is determined from a pre-set selection mapping table. For example, when When, select index as and The component, that is , ;when When, select index as , and The component, that is , , These were selected The weight will serve as a key criterion for judgment.

[0078] The selected key discrimination criteria are weighted and summed to calculate a comprehensive anomaly score. Each selected core feature component... It is assigned a preset weighting coefficient. These weighting coefficients satisfy the normalization condition, meaning the sum of all weighting coefficients is 1. The comprehensive anomaly score SS is calculated using the following formula: ;

[0079] in, It is currently based on level The number of core feature components selected. For example, when , And weight , When the time is right, the calculation formula is: Weighting coefficients The determination is based on historical data analysis and simulation verification to ensure the scoring... It can sensitively and stably reflect the severity of the corresponding anomaly level.

[0080] The calculated comprehensive anomaly score The anomaly type is determined by comparing it with a set of pre-stored anomaly level threshold ranges. Pre-stored thresholds are located in the shared storage area. There are several threshold intervals, each interval consisting of a lower limit value. and upper limit value Define, where And the interval is according to Arranged in ascending order. For example, consider three intervals: correspond , correspond , correspond The calculated score Compare with these intervals in turn. If satisfy: For the last interval, the condition is: This identifier is an integer value, which points to the security scheduling policy to be invoked.

[0081] According to the final determined exception type identifier , the corresponding predefined security scheduling base rule and its adjustable parameter set are indexed and read from the shared memory area of the integrated chip. A policy lookup table is stored in the shared memory area, which is indexed by the exception type identifier . Each table entry contains two parts: one part is the base rule descriptor , which defines the fixed logical framework of routing, arbitration priority adjustment, etc.; the other part is the adjustable parameter set , which contains several parameter names and their corresponding base values or calculation formula identifiers.

[0082] Based on the local running state data of the current routing node, the key parameter values in the adjustable parameter set are calculated and filled in real time to generate a security scheduling policy instance adapted to the current time. The local running state data mainly refers to the current queue length of each virtual channel, denoted as , where is the channel index. For each parameter marked as dynamically calculated in , the calculation formula associated with it is used to perform calculation using real-time data. For example, a parameter named "output port arbitration bias" , its calculation formula can be defined as: ; where is the maximum value of all , is the minimum value, is the sum of , is a preset scaling factor (e.g. 2.0). The routing node calculates , and in real time, and then substitutes them into the formula to obtain the current value of . After all dynamic calculations are completed, the parameter set and the base rule descriptor together constitute a fully instantiated security scheduling policy instance adapted to the current state.

[0083] The generated security scheduling policy instance is set as the current execution policy of the corresponding routing node. The routing node is internally provided with a policy execution context, which contains the currently effective routing and scheduling logic. In this step, the routing node loads the newly generated security scheduling policy instance into the context, making it immediately effective, overriding any adaptive decision logic that may have been in use previously. From this point on, the node's subsequent forwarding control of all arriving data packets will strictly follow the rules and parameters defined by this security scheduling policy instance, until the policy selection signal changes again.

[0084] In the adaptive decision generation module, when the policy selection signal indicates maintaining regular scheduling, the routing node will generate scheduling decisions based on its locally continuously optimized adaptive decision logic. This process is a closed-loop optimization process based on historical performance feedback, aiming to continuously fine-tune its decision rules through local learning to adapt to dynamically changing network traffic patterns without global intervention, and continuously approach the global optimization goal. The specific implementation steps are as follows:

[0085] The decision execution effect data is formed by collecting the link transmission delay and queue emptying rate obtained by the corresponding routing node after forwarding data packets according to past scheduling decisions within a historical time window. The length of the historical time window is pre-set to 1024 consecutive clock cycles. Within this window, every time a data packet is successfully forwarded to the next hop node, the routing node records the time it takes from the data packet entering the node's queue to leaving the node, which is the single link transmission delay. At the same time, the routing node calculates the queue emptying rate of each virtual channel every 64 clock cycles, which is calculated by counting the total number of data packets successfully forwarded from the queue in the past 64 cycles, and then dividing the total number by 64. Finally, at the end of a complete 1024-cycle time window, the routing node will obtain a set of sampling values of the link transmission delay, for example, 16, and a set of sampling values of the queue emptying rate, for example, 16. These two sets of sampling values together constitute the decision execution effect data for the scheduling decisions executed within the historical window.

[0086] Next, the average value of the link transmission delay and the average value of the queue emptying rate are extracted from the decision execution effect data as the first evaluation vector. The specific calculation process is as follows: the arithmetic mean of all recorded link transmission delay sampling values is calculated to obtain a delay average value; the arithmetic mean of all recorded queue emptying rate sampling values is calculated to obtain an emptying rate average value. These two average values are combined in order to form a first evaluation vector containing two components, denoted as (delay average value, emptying rate average value).

[0087] The contribution score of the first evaluation vector to the preset global optimization target is calculated. The global optimization target benchmark vector is pre-stored locally in the routing node, which contains two components: one is the ideal link transmission delay target value expected to be achieved, such as 10 clock cycles; the other is the ideal queue emptying rate target value expected to be achieved, such as 0.8 data packets per cycle. When calculating the contribution score, each component is processed respectively. For the delay average component, the ratio of its value to the ideal delay target value is calculated. If the delay average value is less than or equal to the ideal target value, the contribution score of this component is 1.0; if it is greater than the target value, the contribution score is the ideal target value divided by the delay average value. For the emptying rate average component, the ratio of its value to the ideal emptying rate target value is calculated. If the emptying rate average value is greater than or equal to the ideal target value, the contribution score of this component is 1.0; if it is less than the target value, the contribution score is the emptying rate average value divided by the ideal target value. Finally, the contribution scores of the two components are arithmetically averaged to obtain the final contribution score, which ranges from 0 to 1. The larger the value, the higher the positive contribution to the global optimization target.

[0088] The fluctuation variance of the first evaluation vector within the preset time window is calculated, and compared with the allowed fluctuation range to obtain the deviation distance reflecting stability. The calculation of the fluctuation variance is performed for the two components of the first evaluation vector respectively. Taking the original delay sample values based on which the delay average value is calculated as an example, the variance of this set of sample values relative to their arithmetic average (i.e. the first component of the first evaluation vector) is calculated. The specific calculation process is: first, calculate the difference between each sample value and the average value, square each difference value, then sum all the squared values, and finally divide the sum by the number of sample values minus 1. The same variance calculation is performed on the emptying rate sample values. After obtaining the two variance values, they are compared with the predefined allowed fluctuation range respectively. The allowed fluctuation range for the delay variance may be set to 4.0 (squared clock cycles); for the emptying rate variance, it may be set to 0.04 (squared data packets per cycle). The calculation method of the deviation distance is: divide the actual calculated delay variance by the allowed upper limit of the delay variance, divide the actual calculated emptying rate variance by the allowed upper limit of the emptying rate variance, then take the larger of the two ratios as the deviation distance. A distance value greater than 1 indicates that the fluctuation exceeds the allowed range, and a value equal to or less than 1 indicates that the fluctuation is within the acceptable range.

[0089] Then, the contribution score and the deviation distance are weighted and combined according to the preset contribution weight and deviation weight to generate a comprehensive quality evaluation result. The contribution weight and the deviation weight are two positive numbers preset in advance, for example, the contribution weight is set to 0.7 and the deviation weight is set to 0.3, and the sum of the two is 1.0. The comprehensive quality evaluation result is the product of the contribution weight and the contribution score plus the product of the deviation weight and the reciprocal of the deviation distance. The reciprocal of the deviation distance is used to convert the relationship of "the smaller the distance, the more stable" into the evaluation logic of "the larger the value, the better" so as to be consistent with the direction of the contribution score. The final quality evaluation result is a numerical value, and the higher the value, the better the comprehensive performance of the past scheduling decisions in pursuing high performance and maintaining stability.

[0090] According to the calculated quality evaluation result, an adjustment direction and amplitude instruction for a key weight parameter in the local decision rule is generated. The local decision rule of the routing node depends on a set of key weight parameters, for example, weight values used to weigh the priorities of different output ports. The generation of the adjustment instruction is based on the following rules: compare the current quality evaluation result with the quality evaluation result of the last historical time window. If the current result is higher than the historical result, it is determined that the adjustment direction of the weight parameter in the last period is effective, and the same direction is continued to be fine-tuned this time, and the adjustment amplitude is a fixed basic step (for example, 0.01) multiplied by the difference proportion of the current and historical results. If the current result is lower than or equal to the historical result, it is determined that the adjustment direction of the last period may be invalid, and the adjustment is performed in the opposite direction this time, and the adjustment amplitude is also the basic step multiplied by the proportion of the result drop.

[0091] The key weight parameter in the local decision rule is updated by using the above adjustment instruction. The specific operation is: the adjustment amplitude indicated by the adjustment instruction (positive number indicates increase, negative number indicates decrease) is algebraically added to the current value of the corresponding weight parameter to obtain the updated weight parameter value. In order to ensure the rationality of the weight parameter, the updated value will be limited within a preset valid range, for example, between 0.0 and 1.0. If the added value exceeds the upper limit, the upper limit value is taken; if it is lower than the lower limit, the lower limit value is taken.

[0092] With the updated decision rule, combined with the current local running state data, the final scheduling decision is generated. The current local running state data refers to the current queue length of each virtual channel of the routing node, the state of the available output ports, and other information. The updated decision rule defines how to calculate a "preference score" based on these state information to select an output port. The process of generating a scheduling decision is to calculate the real-time preference score of each selectable output port according to the latest rule, and select the port with the highest score as the target port for this time packet forwarding. The decision is immediately set as the current execution strategy of the routing node at the current time, and is immediately applied to the forwarding control of the next arriving packet.

[0093] In the packet forwarding control module, after the current execution strategy of the routing node is determined, the forwarding control needs to be performed for each arriving packet according to the strategy. This process is responsible for converting the abstract rules and parameters in the strategy into deterministic output port selection and scheduling actions for specific packets. The specific implementation steps are as follows:

[0094] The current execution strategy of the routing node is parsed to extract the output port selection rule set and the corresponding dynamic adjustment parameters. The current execution strategy is stored in the local strategy register of the routing node in the form of a data structure. The parsing process includes reading two main parts in the data structure. The first part is the output port selection rule set, which is composed of multiple "condition-action" rules. For example, a rule may stipulate: "if the service type identifier of the packet is 'low delay', and the northbound output port is available, then prefer the northbound port". The rule set contains such mapping relationships for different service types and port states. The second part is the dynamic adjustment parameter, which is a set of variable values, such as a list containing 4 elements, each element corresponding to the current base weight value of an output port, which will be dynamically updated according to the network state.

[0095] According to the output port selection rule set, combined with the service type identifier carried by the current to-be-forwarded packet, the base output port pointing is calculated. The service type identifier of the packet is located in a specific field of its packet header, which is written by the sender when injecting the network, for example, the value 0 represents "best effort", 1 represents "low delay", and 2 represents "high bandwidth". After the routing node reads the identifier, it searches the output port selection rule set for a matching rule entry. For example, when the identifier is 1, the "low delay" rule is matched. Then, the routing node checks the current available state of the "northbound port" mentioned in the rule (such as whether it is blocked due to insufficient credit). If the port is available, it is determined as the preliminary calculation result, i.e. the base output port pointing; if it is not available, the next port is checked in the preset alternative order (such as east, south, west) according to the rule set, until an available port is found as the base output port pointing.

[0096] According to the dynamic adjustment parameter and the real-time local load information of the routing node, the dynamic modification of the priority weight of the basic output port pointing is performed, and the final output port pointing is generated. The real-time local load information mainly refers to the queue length of the virtual channel corresponding to each output port. The process of dynamic modification is as follows: for all available output ports (assuming there are 4), a comprehensive priority score is calculated. For the port determined by the basic output port pointing, the comprehensive priority score is equal to the basic weight value corresponding to it in the dynamic adjustment parameter list. For other available ports, the comprehensive priority score is equal to the basic weight value minus a penalty value positively related to the current queue length, for example, the penalty value is equal to the queue length of the port divided by 10. After the calculation is completed, the routing node compares the comprehensive priority scores of all available ports, and selects the port with the highest score as the final output port pointing. This step ensures that the path can be locally adjusted according to the real-time congestion situation while following the basic rules.

[0097] According to the final output port pointing, the data packet is scheduled to the corresponding output buffer queue, and the forwarding control is completed. The routing node is internally provided with a buffer queue associated with each output port. After the final output port pointing is determined, the switching logic of the routing node takes out the current data packet to be forwarded from the input buffer, and stores it at the tail of the buffer queue corresponding to the target output port. At the same time, the routing node checks whether it has the sending credit according to the flow control protocol (such as credit system) adopted by the output port. If it has, the data packet will be sent to the physical link connected with the downstream node in the subsequent clock cycle; if there is no credit for the time being, the data packet will wait in the queue until it obtains the credit. At this point, the forwarding control operation based on the current execution strategy for the specific data packet is completed.

[0098] In the global strategy aggregation update module, the integrated chip periodically performs the aggregation and fusion update process of the global strategy parameter. This process is responsible for collecting the local experience generated by the distributed routing node in the autonomous optimization process, and integrating it into shareable global knowledge in a coordinated and consistent manner, so as to improve the overall learning efficiency and consistency of the entire network on chip. The specific implementation steps are as follows:

[0099] The parameter update information from each routing node is received. This process is executed by a hardware unit responsible for coordination within the integrated chip, which is triggered at a preset aggregation period. Each routing node adjusts a set of key weight parameters in its local adaptive decision-making process, which are used to guide its decision-making such as port selection. For example, four weight values representing the initial priorities of the east, south, west, and north output ports. At the same time, the node calculates a quality evaluation result for evaluating the effect of its recent decision-making, which is a value between 0 and 1, the higher the value, the better the overall performance of the recent decision-making. When the aggregation period arrives, each routing node packages the current key weight parameter value (for example, a list containing 4 values) and its latest calculated quality evaluation result, and sends them to the designated coordination unit through the control channel of the on-chip network.

[0100] According to the quality evaluation results attached in the information of each node, a dynamic credibility weight is calculated for the key weight parameters sent by each node. After receiving the information of all nodes, the coordination unit first extracts the quality evaluation result in each information package. The purpose of calculating the dynamic credibility weight is to give greater influence to the nodes with better decision-making effects. The specific calculation process is as follows: add the quality evaluation results from all nodes to get a total sum. Then, divide the quality evaluation result of each node by the total sum, and the quotient value is the dynamic credibility weight of the parameter of the node. For example, if the quality evaluation results of three nodes are 0.8, 0.5 and 0.7 respectively, the total sum is 2.0, and their corresponding dynamic credibility weights are 0.4 (0.8 / 2.0), 0.25 (0.5 / 2.0) and 0.35 (0.7 / 2.0) respectively. To ensure that all nodes can participate in fusion, if the quality evaluation result of a node is zero, replace it with a very small positive number (such as 0.001) before participating in the calculation.

[0101] Based on the calculated dynamic credibility weights, a weighted average calculation is performed on the key weight parameters provided by all nodes to generate a new generation of global key weight parameter set. The weighted average is performed independently for each component of the key weight parameter. Assuming that the key weight parameter contains 4 components (corresponding to 4 ports). For the first component, the new generation of global value is calculated as follows: multiply the value of component A of node A by the dynamic credibility weight of node A, add the value of component B of node B multiplied by the dynamic credibility weight of node B, and so on. The same product accumulation operation is performed for all nodes, and the final accumulation sum is the value of the first component of the new generation of global parameter. The same calculation process is repeated for the second, third and fourth components, that is, the dynamic credibility weight of each node is used to weight average the corresponding parameter component provided by the node. Finally, a new generation of global key weight parameter set containing 4 average values is generated.

[0102] The generated new generation global key weight parameter set is used to replace the original global strategy parameters in the shared storage area, and the fusion update is completed. A special area is arranged in the shared storage area of the integrated chip for storing the global strategy parameters referenced by all routing nodes. After the weighted average calculation is completed, the coordination unit writes the obtained new generation parameter set to the storage area through the bus, directly covering the previously stored old parameter set. The update operation is designed as an atomic operation to ensure that there is no inconsistent parameter state during the update process. Once the update is completed, each routing node can read the global parameters fused with the latest and most effective experience of all network nodes in the subsequent initialization or strategy synchronization stage, thereby guiding the local decision logic to evolve in a more optimal direction.

[0103] The working principle of the application is as follows: in the on-chip network of the integrated chip, a dual-mode decision and dynamic switching mechanism is configured for each routing node: in the normal state, each node continuously optimizes its adaptive decision logic based on local historical performance feedback, evaluates the decision quality by calculating the transmission delay and queue emptying rate of historical data packets, and adjusts the weight parameters in the local decision rule to generate a scheduling decision for the current load; when the node monitors that its dynamic characteristic parameters match the abnormal state feature set pre-stored in the shared storage area, and the local comprehensive load exceeds the threshold, the safety mode is triggered immediately, the pre-defined safety scheduling strategy corresponding to the specific abnormal type is loaded and instantiated from the shared storage area to replace the adaptive decision logic. All nodes perform real-time port selection and forwarding control on the arriving data packets according to their currently effective decision mode. In addition, the system periodically collects the parameter update information generated by each node in the adaptive mode and the corresponding quality evaluation results, calculates the dynamic credibility weight based on the node performance, and generates a new generation of global strategy parameters through weighted average fusion, and updates to the shared storage area, thereby realizing the continuous convergence and co-evolution of distributed experience and global knowledge, and finally achieving the intelligent scheduling goal of balancing performance, stability and robustness under high dynamic load.

[0104] The above describes one embodiment of the application in detail, but the content described is only the preferred embodiment of the application and cannot be considered as limiting the scope of the implementation of the application. Any equivalent changes and improvements made in the scope of the application should still belong to the patent coverage of the application.

Claims

1. A transmission control and intelligent scheduling system for integrated chips, characterized by, The method comprises the following steps: a state monitoring and feature extraction module for collecting local running state data of each routing node in the integrated chip, and generating a set of dynamic feature parameters based on the local running state data; a policy matching and signal generation module for matching the dynamic feature parameters with an abnormal state feature set pre-stored in a shared storage area of the integrated chip, and generating a policy selection signal according to the matching result; a security policy loading module for loading a pre-defined security scheduling policy corresponding to the current matching result from the shared storage area when the policy selection signal indicates to enable security scheduling, specifically comprising: generating an abnormal type identifier based on the policy selection signal and the dynamic feature parameters; indexing and reading a set of pre-defined security scheduling basic rules and their corresponding adjustable parameter set from the shared storage area according to the abnormal type identifier; generating a security scheduling policy instance adapted to the current time based on real-time calculation and filling of key parameter values in the adjustable parameter set based on the local running state data of the current routing node; setting the security scheduling policy instance as the current execution policy of the corresponding routing node; an adaptive decision generation module for generating a scheduling decision according to its locally continuously optimized adaptive decision logic when the policy selection signal indicates to maintain regular scheduling, and setting the scheduling decision as the current execution policy of the corresponding routing node; a data packet forwarding control module for forwarding control of the arrived data packet according to the current execution policy of the routing node; a global policy aggregation update module for periodically aggregating parameter update information generated based on the adaptive decision logic from each routing node, and fusing and updating global policy parameters stored in the shared storage area according to the aggregated information, specifically comprising: receiving parameter update information from each routing node, which contains key weight parameters of decision rules and corresponding quality evaluation results; calculating a dynamic credibility weight for each node's key weight parameter according to the quality evaluation results attached in each node information; based on the dynamic credibility weight, performing weighted average calculation on the key weight parameters provided by all nodes to generate a new generation of global key weight parameter set; using the new generation of global key weight parameter set to replace the original global policy parameters in the shared storage area to complete fusion update.

2. The transmission control and intelligent scheduling system for integrated chips of claim 1, wherein, The method for generating a set of dynamic feature parameters specifically comprises: obtaining the current queue length of each virtual channel in the routing node as the original state data; based on the original state data, calculating the first-order change rate of each queue length within a preset time window to obtain a load change trend parameter; calculating the second-order change rate of the load change trend parameter within the time window to obtain a load acceleration state parameter; combining the original state data, the load change trend parameter and the load acceleration state parameter to generate the dynamic feature parameters.

3. The transmission control and intelligent scheduling system for integrated chips of claim 1, wherein, The method for generating a policy selection signal specifically comprises: comparing each component of the dynamic feature parameters with the preset threshold range of the corresponding component in the abnormal state feature set one by one to generate a preliminary matching state of a set of components; Counting the components satisfying the matching condition in the preliminary matching state, and comparing the count value with a preset first-level quantity threshold to generate a preliminary matching signal; Performing logical AND operation on the preliminary matching signal and a local comprehensive load level of the current routing node, wherein the local comprehensive load level is obtained by calculating the sum of the lengths of the virtual channel queues; According to the result of the logical AND operation, a final strategy selection signal is generated; when the operation result is true, the strategy selection signal indicates to enable the safe scheduling; when the result of the logical AND operation is false, the strategy selection signal indicates to maintain the regular scheduling.

4. The transmission control and intelligent scheduling system for integrated chips of claim 1, wherein, The generated abnormal type identifier specifically includes: Analyzing the strategy selection signal to obtain the preliminary abnormal level information encoded therein; According to the preliminary abnormal level information, selecting at least two core feature components from the dynamic characteristic parameters as key discriminators; According to the preset weight proportion, performing weighted summation calculation on the selected key discriminators to obtain a comprehensive abnormal score; Comparing the comprehensive abnormal score with a set of pre-stored abnormal level threshold intervals, and determining and outputting the final abnormal type identifier according to the interval into which it falls.

5. The transmission control and intelligent scheduling system for integrated chips of claim 1, wherein, The adaptive decision logic based on its local continuous optimization generates a scheduling decision, specifically including: Collecting the link transmission delay and queue emptying rate obtained after the data packet is forwarded according to the past scheduling decision within a historical time window of the corresponding routing node to form decision execution effect data; Calculating the contribution degree and deviation degree of the decision execution effect data to the preset global optimization target to generate a quality evaluation result of the current decision; According to the quality evaluation result, generating an adjustment direction and amplitude instruction for the key weight parameter in the local decision rule; Updating the key weight parameter by applying the adjustment instruction, and generating the scheduling decision by using the updated decision rule in combination with the current local running state data.

6. The transmission control and intelligent scheduling system for integrated chips of claim 5, wherein, The quality evaluation result of the current decision is generated, specifically including: Extracting the average value of the link transmission delay and the average value of the queue emptying rate from the decision execution effect data as a first evaluation vector; Performing component-by-component comparison between the first evaluation vector and a preset global optimization target reference vector to calculate a contribution degree score reflecting the degree of achievement; Calculating the fluctuation variance of the first evaluation vector within a preset time window, comparing the fluctuation variance with the allowed fluctuation range to obtain a deviation distance reflecting stability; According to the preset contribution weight and deviation weight, the contribution degree score and the deviation distance are weighted and combined to generate a comprehensive quality evaluation result.

7. The transmission control and intelligent scheduling system for integrated chips of claim 1, wherein, The method for forwarding control of the data packet according to the current execution strategy of the routing node specifically includes: Analyzing the current execution strategy to extract the output port selection rule set and the corresponding dynamic adjustment parameters contained therein; According to the output port selection rule set, in combination with the service type identifier carried by the data packet, the basic output port pointing is calculated; According to the dynamic adjustment parameters and the real-time local load information of the routing node, the basic output port pointing is dynamically modified in priority weight to generate the final output port pointing; According to the final output port pointing, the data packet is scheduled to the corresponding output buffer queue to complete the forwarding control.

Images