Transmission control and intelligent scheduling system for integrated chip

Abstract

The invention 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, and the system sets a dual-mode decision and dynamic switching mechanism for each routing node. Calculating dynamic characteristic parameters including an instantaneous value, a first-order trend and a second-order acceleration; when the parameter is matched with a pre-stored abnormal feature set and the load exceeds a threshold value, the node is immediately atomized and switched to a predefined security scheduling strategy loaded from a shared storage area, and otherwise, the node generates a scheduling decision according to self-adaptive decision logic continuously optimized based on historical performance feedback; all the nodes carry out data packet forwarding control according to the current execution strategy; and the system also periodically realizes federated global knowledge evolution according to the quality evaluation result of the self-adaptive decision of each node.

Description

Technical Field

[0001] This invention relates to the field of integrated circuits and computer network technology, and more specifically to a transmission control and intelligent scheduling system for integrated chips. Background Technology

[0002] With the continuous advancement of integrated circuit technology, the scale and complexity of integrated chips are constantly increasing, and on-chip networks have become a critical interconnect infrastructure connecting numerous processor cores, memory units, and dedicated accelerators. To cope with dynamically changing application loads and optimize overall performance, intelligent scheduling methods based on machine learning, especially reinforcement learning, have been introduced into routing and flow control in on-chip networks. These methods allow each network node (router) to autonomously learn and optimize data forwarding decisions based on local observation history, aiming to reduce communication latency and improve throughput. However, this distributed, locally experience-driven learning paradigm has inherent flaws: in the absence of effective global coordination and stabilization mechanisms, the independent exploratory decisions of each agent (router) may produce unpredictable synergistic effects under extreme or non-stationary loads.

[0003] This invention aims to address the global collapse problem caused by the independent exploration and decision-making of each node's intelligent agent in a dynamic environment, which is a consequence of uncoordinated global operation in existing on-chip network transmission control methods based on distributed machine learning. This problem manifests as follows: when network load fluctuates drastically, intelligent routing decisions made by each node based on local historical experience, aimed at optimizing long-term statistical indicators (such as average latency), may conflict or resonate due to a lack of awareness and coordination of the global transient state. This leads to ineffective oscillation and convergence of traffic on limited network resources, triggering self-reinforcing congestion or deadlock across the entire network, causing a sharp drop in network throughput, even to zero. Traditional methods are unable to predict and prevent such a disaster caused by the chain reaction of intelligent decisions in a short time, nor can they effectively guide the system to recover safely and quickly afterward. Therefore, embedding a "circuit breaker" mechanism into the distributed intelligent routing system—capable of real-time detection of abnormal risks and automatic triggering of deterministic protection strategies at critical points—while maintaining its global learning and evolution capabilities, has become a core technical challenge that urgently needs to be overcome in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a transmission control and intelligent scheduling system for integrated chips to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A transmission control and intelligent scheduling system for integrated chips, comprising: The status monitoring and feature extraction module is used to collect local operating status data of each routing node in the integrated chip, and calculate and generate a set of dynamic feature parameters based on the local operating status data. The strategy matching and signal generation module is used to match dynamic feature parameters with the abnormal state feature set pre-stored in the shared memory area of ​​the integrated chip, and generate a strategy selection signal based on the matching result. The security policy loading module is used to load the predefined security scheduling policy corresponding to the current matching result from the shared storage area when the policy selection signal indicates that security scheduling is enabled, and set the security scheduling policy as the current execution policy of the corresponding routing node. The adaptive decision generation module is used to generate a scheduling decision based on its locally continuously optimized adaptive decision logic when the policy selection signal indicates that the normal scheduling should be maintained, and set the scheduling decision as the current execution policy of the corresponding routing node. The packet forwarding control module is used to control the forwarding of arriving packets according to the current execution policy of the routing node. The global policy aggregation and update module is used to periodically aggregate parameter update information from each routing node based on adaptive decision logic, and to merge and update the global policy parameters stored in the shared storage area based on the aggregated information.

[0006] As a further aspect of the present invention: the generation of a set of dynamic feature parameters specifically includes: Obtain the current queue length of each virtual channel in the routing node as the raw state data; Based on the original state data, the first-order rate of change of the length of each queue within a preset time window is calculated to obtain the load change trend parameters. Calculate the second-order rate of change of the load change trend parameter within the time window to obtain the load acceleration state parameter; The original state data, load change trend parameters, and load acceleration state parameters are combined to generate dynamic characteristic parameters.

[0007] As a further aspect of the present invention: the generation of the strategy selection signal specifically includes: Each component of the dynamic feature parameter is compared one by one with the preset threshold range of the corresponding component in the abnormal state feature set to generate a preliminary matching state of a set of components. The components that meet the matching conditions in the initial matching state are counted, and the count value is compared with a preset first-level quantity threshold to generate a primary matching signal; The primary matching signal is logically ANDed with the local comprehensive load level of the current routing node, where the local comprehensive load level is obtained by calculating the sum of the lengths of each virtual channel queue; Based on the result of the logical AND operation, a final policy selection signal is generated; when the result is true, the policy selection signal indicates that safe scheduling is enabled; when the result of the logical AND operation is false, the policy selection signal indicates that normal scheduling is maintained.

[0008] As a further aspect of the present invention: the loading of the predefined security scheduling policy corresponding to the current matching result from the shared storage area specifically includes: Anomaly type identifiers are generated based on strategy selection signals and dynamic feature parameters. Based on the exception type identifier, index and read a set of predefined security scheduling basic rules and their corresponding adjustable parameter sets from the shared storage area; Based on the local operating status 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 strategy instance adapted to the current moment. Set the security scheduling policy instance to the current execution policy of the corresponding routing node.

[0009] As a further aspect of the present invention: the generation of the exception type identifier specifically includes: Analyze the strategy selection signal to obtain the preliminary anomaly level information encoded therein; Based on the preliminary anomaly level information, at least two core feature components are selected from the dynamic feature parameters as key discrimination criteria. Based on the preset weight ratio, the selected key discrimination criteria are weighted and summed to obtain a comprehensive anomaly score; The comprehensive anomaly score is compared with a set of pre-stored anomaly level threshold ranges. Based on the range it falls into, the final anomaly type identifier is determined and output.

[0010] As a further aspect of the present invention: the generation of scheduling decisions based on its locally continuously optimized adaptive decision logic specifically includes: Collect the link transmission delay and queue emptying rate of the corresponding routing node within the historical time window after forwarding data packets based on past scheduling decisions, and form decision execution effect data; The contribution and deviation of the decision execution effect data to the preset global optimization goal are calculated to generate the quality assessment results of the current decision. Based on the quality assessment results, instructions are generated regarding the direction and magnitude of adjustments to key weight parameters in the local decision-making rules; The application adjusts the instructions to update key weight parameters and uses the updated decision rules, combined with the current local running status data, to generate scheduling decisions.

[0011] As a further aspect of the present invention: the generation of the quality assessment result of the current decision specifically includes: From the decision execution effect data, the average value of link transmission delay and the average value of queue emptying rate are extracted as the first evaluation vector. The first evaluation vector is compared with the preset global optimization target benchmark vector component by component to calculate the contribution score that reflects the degree of achievement. Calculate the variance of the first evaluation vector within a preset time window, compare the variance with the allowable fluctuation range, and obtain the deviation distance that reflects stability. Based on the preset contribution weight and deviation weight, the contribution score and deviation distance are weighted and merged to generate a comprehensive quality assessment result.

[0012] As a further aspect of the present invention: the method for forwarding data packets according to the current execution policy of the routing node specifically includes: Analyze the current execution strategy and extract the output port selection rule set and corresponding dynamic adjustment parameters contained therein; Based on the rule set selected by the output port and combined with the service type identifier carried in the data packet, the basic output port direction is calculated; Based on the dynamically adjusted parameters and the real-time local load information of the routing nodes, the priority weight of the basic output port is dynamically corrected to generate the final output port. Based on the final output port, the data packet is scheduled to the corresponding output buffer queue to complete the forwarding control.

[0013] As a further aspect of the present invention: the step of fusing and updating the global policy parameters stored in the shared storage area based on the aggregated information specifically includes: Receive parameter update information from each routing node. The parameter update information includes the key weight parameters of the decision rules and the corresponding quality assessment results. Based on the quality assessment results attached to the information of each node, a dynamic credibility weight is calculated for the key weight parameters of each node. Based on dynamic credibility weights, a weighted average of the key weight parameters provided by all nodes is calculated to generate a new generation of global key weight parameter set. The new generation of global key weight parameters is used to replace the original global policy parameters in the shared storage area to complete the fusion update.

[0014] The beneficial effects of this invention are: (1) Each node continuously calculates the high-order dynamic features of its local state (including instantaneous values, first derivatives, and second derivatives) and quickly matches them with a pre-stored abnormal pattern library. When the matching degree and the local load level both exceed the threshold, the node can atomically switch to a globally verified, deterministic safe scheduling strategy within nanoseconds. This integrated "monitoring-matching-circuit breaking" design implants a hardware-level security barrier into the distributed learning system, enabling the network to immediately degrade from the exploratory optimization mode to a safe mode that ensures connectivity and basic service quality when encountering unknown or high-risk traffic patterns. This ensures the system's functional safety and service determinism in the worst-case scenario, greatly enhancing the robustness of complex on-chip networks.

[0015] (2) By constructing a federated hierarchical learning and global knowledge aggregation framework based on quality assessment, a balance between local adaptive optimization and global experience sharing is achieved, thereby improving the learning efficiency and performance of the entire chip. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, the present invention is a transmission control and intelligent scheduling system for integrated chips, comprising: The status monitoring and feature extraction module is used to collect local operating status data of each routing node in the integrated chip, and calculate and generate a set of dynamic feature parameters based on the local operating status data. The strategy matching and signal generation module is used to match dynamic feature parameters with the abnormal state feature set pre-stored in the shared memory area of ​​the integrated chip, and generate a strategy selection signal based on the matching result. The security policy loading module is used to load the predefined security scheduling policy corresponding to the current matching result from the shared storage area when the policy selection signal indicates that security scheduling is enabled, and set the security scheduling policy as the current execution policy of the corresponding routing node. The adaptive decision generation module is used to generate a scheduling decision based on its locally continuously optimized adaptive decision logic when the policy selection signal indicates that the normal scheduling should be maintained, and set the scheduling decision as the current execution policy of the corresponding routing node. The packet forwarding control module is used to control the forwarding of arriving packets according to the current execution policy of the routing node. The global policy aggregation and update module is used to periodically aggregate parameter update information from each routing node based on adaptive decision logic, and to merge and update the global policy parameters stored in the shared storage area based on the aggregated information.

[0020] In the state monitoring and feature extraction module, each routing node within the integrated chip first needs to perform a state monitoring and feature extraction process. This process aims to transform the real-time, raw operating state of the routing node into a set of feature parameters that comprehensively and deeply characterize the instantaneous dynamics of itself and the local network, for subsequent decision-making. The specific implementation steps are as follows: The routing node acquires its local operational status data. This local operational status data mainly refers to the status of multiple independent queues within the routing node used for temporary storage of data packets to be forwarded. Specifically, each routing node contains multiple virtual channels, and each virtual channel corresponds to a first-in, first-out (FIFO) data buffer queue. In this step, the routing node obtains the queue length of each virtual channel in the current clock cycle by reading the values ​​of its internal hardware registers. 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 status data.

[0021] Based on the acquired raw state data, the routing node calculates a load change trend parameter. This parameter quantifies the rate of change of load for each queue in the short term. Specifically, the routing node maintains a preset time window containing 128 consecutive clock cycles. For each virtual channel, the routing node calculates the first-order rate of change of its queue length within the current time window. The calculation process is as follows: record the queue length at the current clock cycle and 127 cycles ago, subtract the two values, and then divide the difference by the 128 clock cycles corresponding to the time window. The result is the average length change of the queue per clock cycle within the current window, and this change is the load change trend parameter for that queue. This calculation is repeated for all virtual channels to obtain a set of load change trend parameters.

[0022] The routing node further calculates the load acceleration state parameter. This parameter characterizes the acceleration of load changes, i.e., how fast the trend of change itself changes. The calculation process is based on the load change trend parameters of each queue obtained in the previous step. The routing node uses a time window of the same length (128 clock cycles) as in the previous step, but applies it to the load change trend parameter sequence. For each virtual channel, its current load change trend parameter value is calculated and compared with the historical value from 127 cycles ago. The two values ​​are subtracted, and the difference is divided by 128 clock cycles. The result is the average rate of change of the queue's load change trend within the most recent time window, i.e., the load acceleration state parameter. This parameter indicates whether the queue load is growing at an accelerating, constant, or decelerating rate.

[0023] The routing node combines the three types of parameters mentioned above to generate the final dynamic feature parameters. Specifically, the combination involves concatenating three values ​​corresponding to each virtual channel—the current queue length (raw state data), the load change trend parameter, and the load acceleration state parameter—in a fixed order into a numerical vector. For example, for the aforementioned 8-virtual-channel routing node, a feature vector containing 24 numerical components will be generated. This dynamic feature parameter integrates the instantaneous value of the state, first-order dynamics, and second-order dynamics, providing multi-dimensional and precise quantitative input for subsequent judgment of network behavior patterns. This feature vector will be directly output for matching against a pre-stored set of abnormal state features.

[0024] In the policy matching and signal generation module, after the routing node generates dynamic characteristic parameters, a matching and signal generation process needs to be executed. The purpose of this process is to compare the dynamic characteristic parameters, which characterize the current instantaneous state of the network, with predefined characteristic patterns representing known abnormal situations, and based on this, generate a clear control command to determine which scheduling mode should be adopted subsequently. The specific implementation steps are as follows: The process involves comparing each component of the dynamic feature parameter one by one. An abnormal state feature set is pre-stored in the shared memory area of ​​the integrated chip. This set contains multiple feature templates, each corresponding to a predefined abnormal state pattern. Each feature template consists of multiple threshold ranges with the same dimension as the dynamic feature parameter. For example, for a dynamic feature parameter with 24 components, the corresponding feature template also contains 24 threshold ranges. During matching, the first component of the dynamic feature parameter is compared with the threshold range of the first component in the feature template. If the former falls within the numerical range defined by the latter, the component is considered a match and recorded as logical "true"; otherwise, it is recorded as logical "false". This process continues, comparing all 24 components sequentially, ultimately generating a sequence of 24 logical values ​​(true or false). This sequence represents the initial matching state of the set of components.

[0025] The initial matching states are statistically analyzed and compared with a first-level threshold to generate a primary matching signal. Specifically, the obtained sequence of logic values ​​is counted, and the number of values ​​that are "true" is recorded. This count reflects the degree of agreement between the current state and a specific anomaly feature template. The preset first-level threshold is an integer, such as 18. If the count is greater than or equal to 18, the current state is considered to have initially matched the anomaly pattern, and a valid primary matching signal is generated (e.g., logic level "high" or the value "1"). If the count is less than 18, the matching standard is not met, and an invalid primary matching signal is generated (e.g., logic level "low" or the value "0"). This step introduces fault tolerance to the matching process, allowing individual feature components to fail to match due to noise fluctuations.

[0026] The primary matching 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 channels of the routing node, arithmetically adding these length values, and the sum is the local comprehensive load level. For logical judgment, a threshold, such as 64, needs to be set for this load level. When the local comprehensive load level is greater than or equal to 64, the load level condition is satisfied, and the result is converted to logical "true"; otherwise, it is converted to logical "false". The rule of the logical "AND" operation is: the output result is "true" only when both inputs (the validity of the primary matching signal and the authenticity of the load level condition) are both "true"; if either input is "false", the output result is "false".

[0027] The final strategy selection signal is generated based on the Boolean result of the logical AND operation. This signal is a binary control signal. Specifically, when the logical AND operation results in "true," it indicates that both abnormal characteristic patterns and heavy load pressure have been detected simultaneously. In this case, the generated strategy selection signal is encoded as an "Enable Safe Scheduling" instruction. When the logical AND operation results in "false," it indicates that at least one of the above two conditions is not met. In this case, the generated strategy selection signal is encoded as a "Maintain Regular Scheduling" instruction. This signal will be directly output to control the selection and loading of the scheduling strategy in the next stage.

[0028] In the security policy loading module, when the policy selection signal indicates that security scheduling is enabled, the routing node will execute the security policy loading process. This process aims to retrieve and instantiate a corresponding deterministic scheduling policy from a pre-stored knowledge base based on the specific type and severity of the detected anomalies, replacing the original adaptive decision-making logic, thereby ensuring the reliability and stability of the network under abnormal conditions. The specific implementation steps are as follows: Based on the policy selection signal and dynamic characteristic parameters, a precise anomaly type identifier is generated. The policy selection signal is a multi-bit digital signal. For example, a 3-bit signal can represent multiple states, where specific bit combinations are encoded to represent the initial anomaly level: '001' represents a primary anomaly, '010' represents a secondary anomaly, and '011' represents a high-level anomaly. The routing node has internal decoding logic circuitry that extracts specific bits from the signal and converts them into an integer value representing the initial anomaly level according to a predefined encoding table, denoted as [value missing]. ,in The value of can be 1, 2, or 3. The dynamic feature parameter is a vector containing multiple components, denoted as: According to the preliminary anomaly level The core feature component indices to be of interest are determined from a pre-defined 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.

[0029] 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: ; in, It is currently based on level The number of core feature components selected. For example, when , And weight , When, the calculation formula is: Weighting coefficient 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.

[0030] 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 It is an integer value used to precisely point to the security scheduling policy that needs to be invoked.

[0031] Based on the finally determined anomaly type identifier It indexes and reads the corresponding predefined security scheduling basic rules and their adjustable parameter sets from the shared memory area of ​​the integrated chip. A policy lookup table is stored in the shared memory area, identified by exception type. For indexing. Each table entry contains two parts: one part is the basic rule descriptor. One part defines a fixed logical framework for routing, arbitration priority adjustment, etc.; the other part is an adjustable parameter set. It contains several parameter names and their corresponding basic values ​​or calculation formula identifiers.

[0032] Based on the local operating status data of the current routing node, the adjustable parameter set is... The key parameter values ​​are calculated and filled in real time to generate a security scheduling policy instance adapted to the current moment. Local runtime status data mainly refers to the current queue length of each virtual channel, denoted as... ,in For channel indexes. Each parameter marked as requiring dynamic calculation is calculated using real-time data according to its associated calculation formula. For example, a parameter named "Output Port Arbitration Bias"... Its calculation formula may be defined as: ;in It is all The maximum value in, It is the minimum value. yes The sum, It is a preset scaling factor (e.g., 2.0). The routing nodes calculate this in real time. , and Then substitute into the formula to get The current value. The parameter set and basic rule descriptor after all dynamic calculations are completed. Together, they constitute a fully instantiated, safe scheduling policy instance adapted to the current state.

[0033] The generated security scheduling policy instance is set as the current execution policy of the corresponding routing node. The routing node internally maintains 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 this context, making it effective immediately and overriding any previously used adaptive decision logic. From this point onward, the node's subsequent forwarding control of all arriving packets will strictly adhere to the rules and parameters defined by this security scheduling policy instance until the policy selection signal changes again.

[0034] In the adaptive decision generation module, when the policy selection signal indicates that regular scheduling should be maintained, the routing node will generate a scheduling decision 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 without global intervention, to adapt to dynamically changing network traffic patterns and continuously approach the global optimization goal. The specific implementation steps are as follows: The system collects link transmission delays and queue emptying rates from corresponding routing nodes within a historical time window, based on past scheduling decisions, to form decision execution effect data. The historical time window is preset to a continuous 1024 clock cycles. Within this window, whenever a data packet is successfully forwarded to the next-hop node, the routing node records the time elapsed from when the packet enters its queue to when it leaves; this time is the single-transmission delay. Simultaneously, the routing node calculates the queue emptying rate for each virtual channel every 64 clock cycles by counting the total number of data packets successfully forwarded from the queue in the past 64 cycles and then dividing that total by 64. Finally, at the end of a complete 1024-cycle time window, the routing node obtains a set of sampled link transmission delays (e.g., 16) and a set of sampled queue emptying rates (e.g., 16). These two sets of sampled values ​​together constitute the decision execution effect data for the scheduling decisions executed within this historical window.

[0035] Next, the average link transmission delay and the average 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 the sampled link transmission delay values ​​of all records is calculated to obtain an average delay value; the arithmetic mean of the sampled queue emptying rate values ​​of all records is calculated to obtain an average emptying rate value. These two averages are combined sequentially to form a first evaluation vector containing two components, denoted as (average delay, average emptying rate).

[0036] Calculate the contribution score of the first evaluation vector to the preset global optimization objective. The global optimization objective baseline vector is pre-stored locally on the routing node. It contains two components: one is the desired ideal link transmission delay target value, for example, 10 clock cycles; the other is the desired ideal queue emptying rate target value, for example, 0.8 packets per cycle. When calculating the contribution score, each component is processed separately. For the delay average component, calculate its ratio to the ideal delay target value. If the delay average is less than or equal to the ideal target value, the contribution score for 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. For the emptying rate average component, calculate its ratio to the ideal emptying rate target value. If the emptying rate average is greater than or equal to the ideal target value, the contribution score for this component is 1.0; if it is less than the target value, the contribution score is the emptying rate average 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 objective.

[0037] The variance of the first evaluation vector within a preset time window is calculated and compared with the allowable fluctuation range to obtain the deviation distance reflecting stability. The variance calculation is performed separately for each of the two components of the first evaluation vector. Taking the original delay sample values ​​on which the delay average is based as an example, the variance of this set of sample values ​​relative to its arithmetic mean (i.e., the first component of the first evaluation vector) is calculated. Specifically, the difference between each sample value and the average is calculated, each difference is squared, all squared values ​​are summed, and the sum is divided 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 allowable fluctuation range. The allowable fluctuation range may be set to 4.0 (the square of a clock cycle) for delay variance and 0.04 (the square of a data packet per cycle) for emptying rate variance. The deviation distance is calculated as follows: divide the actual calculated delay variance by the upper limit of the allowable delay variance, and divide the actual calculated venting rate variance by the upper limit of the allowable venting 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 allowable range, while a value equal to or less than 1 indicates that the fluctuation is within the acceptable range.

[0038] Subsequently, based on preset contribution and deviation weights, the contribution score and deviation distance are weighted and combined to generate a comprehensive quality assessment result. The contribution weight and deviation weight are two pre-defined positive numbers; for example, a contribution weight of 0.7 and a deviation weight of 0.3 result in a sum of 1.0. The comprehensive quality assessment 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 transform the relationship of "smaller distance, more stable" into an assessment logic of "larger value, better," aligning with the direction of the contribution score. The final quality assessment result is a numerical value; a higher value indicates better overall performance in pursuing high performance and maintaining stability in the past scheduling decisions.

[0039] Based on the calculated quality assessment results, instructions are generated to adjust the direction and magnitude of key weight parameters in the local decision rules. The local decision rules of a routing node rely on a set of key weight parameters, such as weight values ​​used to weigh the priorities of different output ports. The adjustment instructions are generated based on the following rule: comparing the current quality assessment result with the quality assessment result of the previous historical time window. If the current result is higher than the historical result, the adjustment direction of the weight parameters in the previous period is deemed valid, and fine-tuning continues in the same direction, with the adjustment magnitude being a fixed base step size (e.g., 0.01) multiplied by the percentage difference between the current and historical results. If the current result is lower than or equal to the historical result, the adjustment direction in the previous period is deemed invalid, and adjustment is made in the opposite direction, with the adjustment magnitude also being the base step size multiplied by the percentage decrease in result.

[0040] The above adjustment instructions are applied to update the key weight parameters in the local decision rules. Specifically, the adjustment magnitude indicated by the instruction (positive numbers represent increases, negative numbers represent decreases) is algebraically added to the current value of the corresponding weight parameter to obtain the updated weight parameter value. To ensure the reasonableness of the weight parameters, the updated value is restricted to a preset valid range, such as between 0.0 and 1.0. If the added value exceeds the upper limit, the upper limit value is used; if it is lower than the lower limit, the lower limit value is used.

[0041] Using the updated decision rules and current local operational status data, a final scheduling decision is generated. The current local operational status data refers to information such as the current queue length and available output port status of each virtual channel of the routing node. The updated decision rules define how to calculate a "preference score" based on this status information to select the output port. The process of generating the scheduling decision involves calculating the instantaneous preference score for each available output port according to the latest rules and selecting the port with the highest score as the target port for forwarding this data packet. This decision is then set as the current execution strategy for the routing node at the current moment and immediately applied to the forwarding control of the next arriving data packet.

[0042] In the packet forwarding control module, after the current execution policy of the routing node is determined, forwarding control must be performed on each arriving packet according to this policy. This process is responsible for transforming the abstract rules and parameters in the policy into deterministic output port selection and scheduling actions for specific packets. The specific implementation steps are as follows: The process parses the currently active execution policy of the routing node, extracting the output port selection rule set and corresponding dynamic adjustment parameters. The current execution policy is stored as a data structure in the routing node's local policy register. The parsing process involves reading two main parts of this data structure. The first part is the output port selection rule set, which consists of multiple "condition-action" rules. For example, a rule might state: "If the service type of a packet is identified as 'low latency' and the northbound output port is available, then the northbound port is selected first." The rule set contains such mappings for different service types and port states. The second part is the dynamic adjustment parameters, a set of variable values, such as a list containing four elements, each corresponding to the current base weight value of an output port. These weight values ​​are dynamically updated based on network conditions.

[0043] The basic output port pointer is calculated based on the output port selection rule set and the service type identifier carried by the current data packet to be forwarded. The service type identifier of the data packet is a specific field in its packet header, written by the sender when injecting it into the network. For example, 0 represents "best effort", 1 represents "low latency", and 2 represents "high bandwidth". After reading this identifier, the routing node searches for a matching rule entry in the output port selection rule set. For example, when the identifier is 1, the aforementioned "low latency" rule is matched. Then, the routing node checks the current availability status of the "northbound port" mentioned in the rule (e.g., whether it is blocked due to insufficient credit). If the port is available, it is determined as the preliminary calculation result, i.e., the basic output port pointer; if it is unavailable, the next port is checked in the preset alternative order of the rule set (e.g., eastbound, southbound, westbound) until an available port is found as the basic output port pointer.

[0044] Based on dynamically adjusted parameters and real-time local load information of routing nodes, the priority weights of the basic output port pointers are dynamically adjusted to generate the final output port pointers. Real-time local load information mainly refers to the queue length of the virtual channel corresponding to each output port. The dynamic adjustment process 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 pointer, its comprehensive priority score is equal to its corresponding basic weight value in the dynamic adjustment parameter list. For other available ports, its comprehensive priority score is equal to its basic weight value minus a penalty value positively correlated with the current queue length; for example, the penalty value is equal to the port's queue length divided by 10. After calculation, 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 pointer. This step ensures that while adhering to basic rules, the path can be locally adjusted according to real-time congestion.

[0045] Based on the final output port, the data packet is scheduled to the corresponding output buffer queue, completing the forwarding control. Each routing node has an internal buffer queue associated with each output port. After determining the final output port, the routing node's switching logic retrieves the 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. Simultaneously, the routing node checks whether it has sufficient transmission credit based on the flow control protocol (such as credit-based) used by the output port. If it does, the data packet will be sent to the physical link connecting to the downstream node in a subsequent clock cycle; if it does not yet have sufficient credit, the data packet waits in the queue until it receives credit. Thus, the forwarding control operation for that specific data packet, based on the current execution policy, is completed.

[0046] In the global policy aggregation and update module, the integrated chip periodically executes the aggregation and fusion update process of global policy parameters. This process is responsible for collecting the local experience generated by distributed routing nodes during autonomous optimization and integrating it into shareable global knowledge in a coordinated manner, thereby improving the overall learning efficiency and consistency of the entire on-chip network. The specific implementation steps are as follows: The system receives parameter updates from each routing node. This process is executed by a coordinating hardware unit within the integrated chip, triggered at a preset aggregation cycle. During its local adaptive decision-making process, each routing node adjusts a set of key weight parameters guiding its port selection and other decisions. These parameters may include four weights representing the initial priorities of the four output ports: East, South, West, and North. Simultaneously, the node calculates a quality assessment result to evaluate the effectiveness of its recent decisions. This result is a value between 0 and 1, with higher values ​​indicating better overall performance. At the arrival of the aggregation cycle, each routing node packages its current key weight parameter values ​​(e.g., a list of four values) and its most recently calculated quality assessment result, and sends them to the designated coordinating unit via the on-chip network's control channel.

[0047] Based on the quality assessment results attached to the information from each node, a dynamic credibility weight is calculated for the key weight parameters sent by each node. After receiving information from all nodes, the coordination unit first extracts the quality assessment results from each information packet. The purpose of calculating the dynamic credibility weight is to give greater influence to nodes with better decision-making performance. The specific calculation process is as follows: The quality assessment results from all nodes are added together to obtain a sum. Then, the quality assessment result of each node is divided by this sum, and the quotient is the dynamic credibility weight of that node's parameter. For example, if the quality assessment results of three nodes are 0.8, 0.5, and 0.7 respectively, the 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 the fusion, if the quality assessment result of a certain node is zero, it is replaced with a very small positive number (such as 0.001) before participating in the calculation.

[0048] Based on the calculated dynamic credibility weights, a weighted average is calculated for 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. Assume the key weight parameter contains four components (corresponding to four ports). For the first component, its new generation global value is calculated as follows: multiply the value of this component from node A by node A's dynamic credibility weight, add the value of this component from node B by node B's dynamic credibility weight, and so on, performing the same multiplication and summation operation on all nodes. The final sum is the value of the first component in the new generation of global parameters. The same calculation process is repeated for the second, third, and fourth components, i.e., using the dynamic credibility weight corresponding to each node to weight and average its corresponding parameter component. Finally, a new generation of global key weight parameter set containing four average values ​​is generated.

[0049] The generated new-generation global key weight parameter set replaces the original global policy parameters in the shared storage area, completing the fusion update. The integrated chip's shared storage area has a dedicated region for storing global policy parameters for all routing nodes to reference. After completing the weighted average calculation, the coordination unit writes the resulting new-generation parameter set to this storage area via the bus, directly overwriting the previously stored old parameter set. This update operation is designed as an atomic operation to ensure that no parameter inconsistencies occur during the update process. Once the update is complete, each routing node can read these global parameters, which incorporate the latest and most effective experience from all network nodes, during subsequent initialization or policy synchronization phases, thereby guiding its local decision-making logic towards a better direction.

[0050] The working principle of this invention is as follows: In the on-chip network of an integrated chip, each routing node is configured with a dual-mode decision-making and dynamic switching mechanism. Under normal conditions, each node continuously optimizes its adaptive decision-making logic based on local historical performance feedback. It evaluates the decision quality by calculating the transmission delay and queue emptying rate of historical data packets and adjusts the weight parameters in its local decision-making rules accordingly, thereby autonomously generating scheduling decisions for the current load. When a node detects that its dynamic characteristic parameters match the abnormal state feature set pre-stored in the shared storage area, and the local comprehensive load exceeds a threshold, it immediately triggers a safe mode. It loads and instantiates a predefined safe scheduling policy corresponding to a specific abnormal type from the shared storage area to replace the adaptive decision-making logic. All nodes perform real-time port selection and forwarding control on arriving data packets according to their currently effective decision mode. In addition, the system periodically collects parameter update information and corresponding quality assessment results generated by each node in adaptive mode, calculates dynamic credibility weights based on node performance, and generates a new generation of global policy parameters through weighted average fusion, which are then updated to the shared storage area. This enables the continuous convergence and collaborative evolution of distributed experience into global knowledge, ultimately achieving the goal of intelligent scheduling that balances performance, stability, and robustness under high dynamic loads.

[0051] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A transmission control and intelligent scheduling system for integrated chips, characterized in that, include: The status monitoring and feature extraction module is used to collect local operating status data of each routing node in the integrated chip, and calculate and generate a set of dynamic feature parameters based on the local operating status data. The strategy matching and signal generation module is used to match dynamic feature parameters with the abnormal state feature set pre-stored in the shared memory area of ​​the integrated chip, and generate a strategy selection signal based on the matching result. The security policy loading module is used to load the predefined security scheduling policy corresponding to the current matching result from the shared storage area when the policy selection signal indicates that security scheduling is enabled, and set the security scheduling policy as the current execution policy of the corresponding routing node. The adaptive decision generation module is used to generate a scheduling decision based on its locally continuously optimized adaptive decision logic when the policy selection signal indicates that the normal scheduling should be maintained, and set the scheduling decision as the current execution policy of the corresponding routing node. The packet forwarding control module is used to control the forwarding of arriving packets according to the current execution policy of the routing node. The global policy aggregation and update module is used to periodically aggregate parameter update information from each routing node based on adaptive decision logic, and to merge and update the global policy parameters stored in the shared storage area based on the aggregated information.

2. The transmission control and intelligent scheduling system for integrated chips according to claim 1, characterized in that, The generation of a set of dynamic feature parameters specifically includes: Obtain the current queue length of each virtual channel in the routing node as the raw state data; Based on the original state data, the first-order rate of change of the length of each queue within a preset time window is calculated to obtain the load change trend parameters. Calculate the second-order rate of change of the load change trend parameter within the time window to obtain the load acceleration state parameter; The original state data, load change trend parameters, and load acceleration state parameters are combined to generate dynamic characteristic parameters.

3. The transmission control and intelligent scheduling system for integrated chips according to claim 1, characterized in that, The generation strategy selection signal specifically includes: Each component of the dynamic feature parameter is compared one by one with the preset threshold range of the corresponding component in the abnormal state feature set to generate a preliminary matching state of a set of components. The components that meet the matching conditions in the initial matching state are counted, and the count value is compared with a preset first-level quantity threshold to generate a primary matching signal; The primary matching signal is logically ANDed with the local comprehensive load level of the current routing node, where the local comprehensive load level is obtained by calculating the sum of the lengths of each virtual channel queue; Based on the result of the logical AND operation, a final policy selection signal is generated; when the result is true, the policy selection signal indicates that safe scheduling is enabled; when the result of the logical AND operation is false, the policy selection signal indicates that normal scheduling is maintained.

4. The transmission control and intelligent scheduling system for integrated chips according to claim 1, characterized in that, The loading of the predefined security scheduling policy corresponding to the current matching result from the shared storage area specifically includes: Anomaly type identifiers are generated based on strategy selection signals and dynamic feature parameters. Based on the exception type identifier, index and read a set of predefined security scheduling basic rules and their corresponding adjustable parameter sets from the shared storage area; Based on the local operating status 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 strategy instance adapted to the current moment. Set the security scheduling policy instance to the current execution policy of the corresponding routing node.

5. A transmission control and intelligent scheduling system for integrated chips according to claim 4, characterized in that, The generated exception type identifier specifically includes: Analyze the strategy selection signal to obtain the preliminary anomaly level information encoded therein; Based on the preliminary anomaly level information, at least two core feature components are selected from the dynamic feature parameters as key discrimination criteria. Based on the preset weight ratio, the selected key discrimination criteria are weighted and summed to obtain a comprehensive anomaly score; The comprehensive anomaly score is compared with a set of pre-stored anomaly level threshold ranges. Based on the range it falls into, the final anomaly type identifier is determined and output.

6. The transmission control and intelligent scheduling system for integrated chips according to claim 1, characterized in that, The process of generating scheduling decisions based on locally continuously optimized adaptive decision-making logic specifically includes: Collect the link transmission delay and queue emptying rate of the corresponding routing node within the historical time window after forwarding data packets based on past scheduling decisions, and form decision execution effect data; The contribution and deviation of the decision execution effect data to the preset global optimization goal are calculated to generate the quality assessment results of the current decision. Based on the quality assessment results, instructions are generated regarding the direction and magnitude of adjustments to key weight parameters in the local decision-making rules; The application adjusts the instructions to update key weight parameters and uses the updated decision rules, combined with the current local running status data, to generate scheduling decisions.

7. A transmission control and intelligent scheduling system for integrated chips according to claim 6, characterized in that, The quality assessment results for generating the current decision specifically include: From the decision execution effect data, the average value of link transmission delay and the average value of queue emptying rate are extracted as the first evaluation vector. The first evaluation vector is compared with the preset global optimization target benchmark vector component by component to calculate the contribution score that reflects the degree of achievement. Calculate the variance of the first evaluation vector within a preset time window, compare the variance with the allowable fluctuation range, and obtain the deviation distance that reflects stability. Based on the preset contribution weight and deviation weight, the contribution score and deviation distance are weighted and merged to generate a comprehensive quality assessment result.

8. A transmission control and intelligent scheduling system for integrated chips according to claim 1, characterized in that, The method for forwarding data packets according to the current execution policy of the routing node specifically includes: Analyze the current execution strategy and extract the output port selection rule set and corresponding dynamic adjustment parameters contained therein; Based on the rule set selected by the output port and combined with the service type identifier carried in the data packet, the basic output port direction is calculated; Based on the dynamically adjusted parameters and the real-time local load information of the routing nodes, the priority weight of the basic output port is dynamically corrected to generate the final output port. Based on the final output port, the data packet is scheduled to the corresponding output buffer queue to complete the forwarding control.

9. A transmission control and intelligent scheduling system for integrated chips according to claim 1, characterized in that, The process of fusing and updating global policy parameters stored in the shared storage area based on the aggregated information specifically includes: Receive parameter update information from each routing node. The parameter update information includes the key weight parameters of the decision rules and the corresponding quality assessment results. Based on the quality assessment results attached to the information of each node, a dynamic credibility weight is calculated for the key weight parameters of each node. Based on dynamic credibility weights, a weighted average of the key weight parameters provided by all nodes is calculated to generate a new generation of global key weight parameter set. The new generation of global key weight parameters is used to replace the original global policy parameters in the shared storage area to complete the fusion update.

Images