Multi-substream parallel forwarding scheduling method for novel intelligent computing center network

Abstract

The invention discloses a multi-substream parallel forwarding scheduling method for a novel intelligent computing center network. The whole communication forwarding process is divided into three stages of data stream segmentation and distribution, parallel transmission and receiving end sorting and recombination. The cooperation of a data stream segmentation distribution strategy of a sending end and an out-of-order data packet sorting and recombination operation of a receiving end is comprehensively considered, an optimal distribution ratio and a corresponding optimal forwarding path are determined, and the optimal distribution ratio and the corresponding optimal forwarding path are mapped to a physical network to complete parallel transmission of multiple substreams. According to the method, the problem of failure of a traditional network scheduling method caused by new characteristics of small flow quantity and large single flow bandwidth of distributed training service data flow of an AI large model in a novel intelligent calculation center is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer network data transmission, and specifically, is a forwarding scheduling method based on multi-substream parallel transmission for ultra-large bandwidth business data streams in a new intelligent computing center network. Background Art

[0002] The New Intelligent Computing Center (NICC) for distributed training of large AI models is a new type of data center that deploys high-performance distributed computing clusters, specifically for artificial intelligence services, and provides powerful computing power and storage resources. Traditional data centers use CPUs to perform computing tasks, while the intelligent computing center network is mainly used to carry AI large model training services. Its GPU computing power has higher computing performance than the CPU. In addition, in the distributed training scenario of large AI models, common parallel modes such as data parallelism, model parallelism, and tensor parallelism require the computing nodes in the cluster to perform a large number of collective communication operations after completing the current round of computing tasks to quickly synchronize the calculation results to other nodes for the next round of calculations. Complex collective communications will generate a very large amount of communication data, and the bandwidth demand for network forwarding will increase significantly. The network of the intelligent computing center is facing unprecedented challenges.

[0003] As the scale of new intelligent computing centers continues to expand, the network of intelligent computing centers will develop in the direction of integrating multiple network modes, such as the coexistence of network modes such as MTN, InfiniBand and RoCE. Different network modes are heterogeneous, and there are significant differences in communication performance and characteristics. The completion time of communication is determined by the completion time of the slowest node. Scenarios where multiple network modes coexist are more likely to cause long-tail delay problems and slow down task completion time. During the distributed training of large AI models, the communication time accounts for too long, and the services provided by the network are difficult to match the speed of high-performance computing resources, which will cause a large amount of computing resources to be idle and waiting, directly affecting the parameter update of large model training and a new round of iterative calculations. The network is becoming a bottleneck that limits the efficiency and effectiveness of large model training.

[0004] Different from the business traffic of traditional data centers, the business data flow of the new intelligent computing center has the significant characteristics of a small number of flows and a large single-flow bandwidth. To deal with the elephant flow scheduling problem, the solution of single-flow single-path scheduling can be summarized as follows: first, divide the elephant flow from the data flow, and then reroute the elephant flow separately by comprehensively considering the influencing factors such as the available load of the path and the distribution degree of the large flow. The above scheme does reduce the impact of the elephant flow on network load balancing in traditional data centers, but to deal with the business needs of small flows and super-large flows in the new intelligent computing center, hash routing based on flow granularity may lead to uneven utilization of network links, cause network congestion, and significantly increase network forwarding delay. Another type of scheduling scheme is to split the elephant flow and complete the communication forwarding of the elephant flow in combination with multi-path parallel transmission. The existing scheme based on data packets for multi-path scheduling and forwarding of operation units can give full play to the advantages of multi-path parallel transmission and effectively shorten the transmission delay, but it will cause serious packet disorder problems and introduce additional sorting and reorganization delays. The disorder problem is difficult to solve, and the effect of multi-path parallel scheduling by data packets may sometimes even be worse than that of the single-path scheme.

[0005] In view of the gap between the new intelligent computing center network and business needs, the multi-path parallel transmission scheduling scheme can make full use of the rich parallel path resources in the new intelligent computing center network. However, the existing elephant flow multi-path parallel scheduling method has the following problems that need to be solved urgently:

[0006] 1. Elephant flows usually contain a large amount of data, which needs to be segmented and transmitted in parallel through multiple paths. The granularity of elephant flow segmentation has a significant impact on the efficiency of multi-path parallel transmission. If the segmentation granularity is too large, it may not be able to fully utilize the parallel transmission capabilities of multiple paths, and it is more likely to cause congestion during transmission in the network; while if the segmentation granularity is too small, it may cause serious disorder problems, increase the complexity of data processing at the receiving end, generate a large amount of additional sorting and reorganization overhead, and reduce transmission efficiency. Therefore, the selection of the segmentation granularity of elephant flows is a key issue, and a clear solution is urgently needed to segment elephant flows.

[0007] 2. When the elephant flow is transmitted in parallel through multiple paths, due to differences in performance such as latency and bandwidth, the data may take longer to reach the receiving end on some paths, resulting in long-tail latency problems and slowing down the overall transmission efficiency. In addition, in an environment with large path differences, inappropriate path selection will lead to a surge in the amount of out-of-order data, exacerbating the overhead problem of out-of-order processing, causing the performance of multi-path parallel transmission to decrease instead of increase. In the scenario where multiple network modes coexist in the new intelligent computing center, how to select the optimal transmission path for the split elephant flow, so as to achieve a technical solution to reduce long-tail latency and alleviate data disorder, is not supported by the existing solutions.

[0008] 3. The current research on multi-path parallel scheduling methods focuses on avoiding disorder as much as possible, but the existing schemes cannot completely avoid the situation where packets arrive out of order. Therefore, it is crucial to comprehensively consider the split distribution strategy of the sender and the sorting and reorganization scheme of the receiver, and weigh the increase or decrease of the parallel transmission delay of the elephant flow on the link and the sorting and reorganization delay of the receiver, so as to find a scheduling scheme that minimizes the completion time of the elephant flow.

[0009] In summary, based on the ultra-large bandwidth business data flow communication scenario of the new intelligent computing center, it is urgent to design a new multi-path parallel forwarding efficient scheduling mechanism. Summary of the invention

[0010] The purpose of the present invention is to provide a multi-sub-flow parallel forwarding scheduling method for a new intelligent computing center network, which is used to solve the problem that the existing scheduling strategy is difficult to meet the communication needs when the new intelligent computing center network copes with a small number of flows and super-large flows in the distributed training of large AI models.

[0011] The technical solution to achieve the purpose of the present invention is: a multi-substream parallel forwarding scheduling method for a new intelligent computing center network, which performs multi-substream parallel forwarding scheduling on the AI ​​large model distributed training service data flow in the new intelligent computing center, including:

[0012] In the data stream segmentation and distribution phase, the service data stream is first segmented, and the data packets are sequentially placed into fixed-length containers of pre-set lengths for standby use, and a continuous sequence number is sequentially assigned to each fixed-length container as a unique identifier; the sender distributes the segmented fixed-length containers to the corresponding sub-streams in a round-robin manner according to the configuration information issued by the central controller;

[0013] In the parallel transmission stage, the configuration information is mapped to the physical network, and the fixed-length container is used as the smallest operation unit for path selection. Different fixed-length containers can choose different sub-flow paths for forwarding, thus realizing parallel transmission with sub-flow as the granularity.

[0014] In the sorting and reorganization phase at the receiving end, after each round of parallel transmission, the receiver sorts and reorganizes the out-of-order fixed-length containers in the cache and submits them to the upper layer; it also feeds back information to the central controller; and finally clears the corresponding cache space to wait for the next round of multi-substream parallel transmission.

[0015] Furthermore, before the current service data flow communication is forwarded, the central controller integrates and sends configuration information and cache management information, including:

[0016] After receiving the service request, the sender uploads the key information contained in the request to the central controller;

[0017] The central controller obtains the physical network status in real time, determines the number of available sub-flows in the network for data forwarding from the sender to the receiver, and the link performance indicators corresponding to each sub-flow path, including the current available physical bandwidth and propagation delay;

[0018] The acquired physical network status information is used as input, with the optimization goals of minimizing the total delay of data flow communication forwarding and maximizing the utilization of network resources, to find the best diversion ratio of the current business data flow and the corresponding optimal forwarding path, to achieve the optimal balance between parallel transmission delay and sorting and reorganization delay, and to integrate them into configuration information and cache management information, which are sent from the central controller to the data forwarding plane to control parallel forwarding scheduling.

[0019] Furthermore, the data stream segmentation and distribution stage includes:

[0020] The sender divides the service data stream requested for communication forwarding. The sending node records the data packets in order into fixed-length containers of a preset length to the maximum extent possible while ensuring the integrity of the data packets, and assigns a continuous sequence number as a unique identifier to each fixed-length container in order.

[0021] After completing the splitting operation, the sending node determines the optimal load corresponding to each sub-stream based on the configuration information sent by the central controller, and takes the maximum number of fixed-length containers that the receiving end cache can accommodate as the basic indicator. It synchronizes the transmission completion time of multiple sub-streams in each round as much as possible, calculates the total number of fixed-length containers transmitted in each round, and distributes the split fixed-length containers to the corresponding sub-streams in rounds for parallel transmission.

[0022] Furthermore, the parallel transmission stage includes:

[0023] The segmented fixed-length container is used as the smallest operation unit for path selection. All data packets belonging to the same fixed-length container will be scheduled to the same sub-flow path for forwarding, ensuring that the data packets in the fixed-length container arrive at the receiving end in order. Different fixed-length containers are mapped to the physical network according to the configuration information sent by the central controller, and different paths can be selected for transmission, realizing parallel transmission of multiple sub-flows.

[0024] Furthermore, the receiving end sorting and reorganization stage includes:

[0025] After receiving the cache management information sent by the central controller, the receiving end opens up an independent space for the current business data flow to classify and cache the disordered fixed-length containers; the disordered fixed-length containers arriving at the receiving end are cached by the receiving end. If a certain fixed-length container is the first fixed-length container on the substream that arrives at the receiving end and needs to be cached, the receiving end creates a new cache queue for it in the corresponding independent space for temporary storage; otherwise, the fixed-length container is directly placed in the corresponding cache queue for temporary storage; wait for the last fixed-length container of the same round to arrive and be confirmed, and then perform unified sorting and reorganization operations;

[0026] After each round of parallel transmission, the receiver sorts the random fixed-length containers in the cache using the sequence number as the key, submits them to the upper layer after reorganization, and feeds back information to the central controller; finally, the corresponding cache space is cleared to wait for the next round of multi-substream parallel transmission;

[0027] The central controller notifies the sender that it can start the next round of parallel transmission until the entire business data stream is fully forwarded.

[0028] Furthermore, the key information in the service request includes: the sending end node , receiving node , length of elephant flow , Maximum communication delay .

[0029] Furthermore, a constrained multi-objective optimization algorithm is used to balance the increase and decrease of parallel transmission delay and sorting and reorganization delay, determine the best split ratio of the current business data flow, and find the corresponding optimal forwarding path;

[0030] The constrained multi-objective optimization algorithm is specifically: By The sum of integers is a fixed value The row vector of is used as the initial population, and different sub-flow division and forwarding path matching schemes are tried in multiple selection, crossover and mutation processes. The results after multiple iterations are controlled to continuously approach the optimal balance between the multi-sub-flow parallel transmission delay and the sorting and reorganization delay. Under the conditions of minimizing the total communication forwarding delay of the business data flow and maximizing the network resource utilization as the key goals, the optimal diversion ratio of the current business data flow and the corresponding optimal forwarding path are found, and finally the optimal configuration of the number of fixed-length containers on each virtual sub-flow path is determined;

[0031] Furthermore, the constraint processing method is specifically as follows: define the violation value of each individual ; The calculation method is ,in is a row vector that satisfies , Indicates the total number of fixed-length containers after the business data stream is split. Indicates the number of fixed-length containers that can be transmitted by each available sub-stream, is the number of available subflows in the network for data forwarding from the sender to the receiver; if the individual does not meet the constraint conditions, the constraint violation value is the absolute value of the true constraint function value; if the individual meets the constraint conditions, the constraint violation value is 0; the constraint violation value is normalized to obtain the constraint violation degree, and the value range of the constraint violation degree is The multi-objective optimization evolutionary algorithm uses constraint violation as the standard to distinguish feasible solutions from infeasible solutions. Individuals with a constraint violation of 0 correspond to feasible solutions, and individuals with a constraint violation of non-zero correspond to infeasible solutions.

[0032] Furthermore, the process of the sender segmenting the service data flow requesting communication forwarding is as follows:

[0033] Sending Node First, the service request information is forwarded to the central controller, and the information needs to include the data stream waiting for communication forwarding , Defined as a four-tuple ,in, and Represents data streams The sending node and the receiving node, Indicates the length of the data stream, It is the longest communication forwarding time that the data flow can accept;

[0034] Sending Node Before distributing the data The original data is split, and each data packet is sequentially placed into multiple fixed-length containers of the same length for standby according to the position order in the original data stream; when the current fixed-length container is full or the remaining space is insufficient to place a complete data packet, the data packet is switched to the next fixed-length container; a continuous sequence number is sequentially assigned to each fixed-length container as a unique identifier; the length of the fixed-length container is set in advance to , satisfying the conditions .

[0035] Furthermore, the specific implementation of distributing the segmented fixed-length containers to the corresponding sub-streams in rounds for parallel transmission is as follows:

[0036] The total number of fixed-length containers transmitted in each round is the number of caches on the receiving end. As a reference indicator, under the premise that the cache can accommodate, arrange as many fixed-length containers as possible for multi-substream parallel transmission according to the diversion ratio in the configuration information, and calculate the total number of data packets transmitted in each round by the following formula ,

[0037]

[0038]

[0039]

[0040] in Indicates the number of fixed-length containers that can be transmitted by each available sub-stream. is the number of available sub-flows in the network for data forwarding from the sender to the receiver, is a positive integer variable;

[0041] The sending node is based on the configuration deployment information issued by the central controller and , determine the optimal load situation of each sub-flow in each round, distribute the segmented fixed-length containers to the corresponding sub-flows in rounds, and prepare for subsequent mapping to the physical network for parallel transmission of multiple sub-flows.

[0042] Compared with the existing scheduling scheme, the present invention has the following advantages:

[0043] (1) The present invention makes full use of the abundant parallel paths within the new intelligent computing center network. These paths are balanced to provide more efficient transmission services for the elephant flow and meet the communication needs of the AI ​​large model distributed training business. Multi-path parallelism can make full use of the available bandwidth in the network and effectively shorten the delay of data forwarding on the physical link. Although it will introduce additional delays in the sorting and reorganization operations at the receiving end, it can control the diversion situation and weigh the increase and decrease of delay to find the optimal solution, truly giving play to the advantages of parallel transmission and greatly shortening the total communication forwarding delay of the elephant flow. In addition, multi-path can avoid the fragmentation of network resources caused by single-flow and single-path, alleviate the problem of elephant flows waiting in line at the sending end because there is no complete and sufficient remaining bandwidth, and effectively improve the utilization rate of network bandwidth resources.

[0044] (2) The present invention proposes to use sub-streams as the scheduling granularity to realize elephant flow transmission. Scheduling by sub-stream solves the problems of traditional scheduling by stream, such as too large granularity, inflexibility, uneven network load, long communication time, and fragmentation and waste of network resources. Compared with the existing method of scheduling by data packet, the scheme of scheduling by sub-stream has a larger granularity, which alleviates the problems of severe disorder of data packets in parallel transmission, which are difficult to handle, and long reassembly time. Scheduling elephant flows by sub-streams achieves a trade-off between the parallel transmission delay of elephant flows on the link and the sorting and reassembly delay of the receiving end, which can effectively shorten the completion time of elephant flow communication.

[0045] (3) The present invention introduces the idea of ​​merging multiple modes of future networks. Considering that the network modes of different sub-streams may be different, there may be large differences in performance such as latency. The latency of each sub-stream is independently analyzed and calculated, which is widely applicable to various network architectures that may exist in the intelligent computing center now and in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the control plane parallel forwarding scheduling processing flow of an embodiment of the present invention.

[0047] Figure 2 It is a schematic diagram of the parallel forwarding and sorting processing flow of the data forwarding plane according to an embodiment of the present invention.

[0048] Figure 3 It is a schematic diagram of the cooperative relationship between the segmentation and distribution and the sorting and reorganization strategies in an embodiment of the present invention.

[0049] Figure 4 It is a flowchart of a multi-objective evolutionary optimization algorithm of a central controller according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In view of the new demand characteristics of the elephant flow communication of the intelligent computing business, the scheduling mechanism needs to comprehensively consider the coordination of the data flow segmentation and distribution at the sending end and the data packet sorting and reorganization operations at the receiving end, find the best diversion ratio of the elephant flow and map it to the physical network to complete the parallel transmission of multiple sub-flows according to the optimal forwarding path, and achieve the two core goals of minimizing the completion time of the elephant flow communication and maximizing the utilization of network resources. In this way, the advantages of multi-path parallel forwarding can be truly utilized, and the delay of data forwarding through network communication during distributed training of large AI models can be effectively reduced, thereby improving the utilization of intelligent computing resources and providing vital network communication service guarantees for the sustainable development of large AI models.

[0051] This invention uses the new intelligent computing center network as the application scenario, optimizes the scheduling and forwarding of AI large model distributed training business data flow, and proposes an efficient scheduling mechanism for parallel forwarding of multiple sub-flows, which divides the entire communication forwarding process into three stages: data flow segmentation and distribution, parallel transmission, and receiving end sorting and reorganization. The data flow segmentation and distribution strategy at the sending end and the coordination of the out-of-order data packet sorting and reorganization operation at the receiving end are comprehensively considered to determine the optimal elephant flow split ratio and map it to the physical network to complete the parallel transmission of multiple sub-flows.

[0052] The multi-substream parallel forwarding scheduling method provided by the present invention solves the problem of failure of traditional network scheduling methods caused by the new characteristics of large-scale AI model distributed training business data flow, small number of flows, and large single-flow bandwidth in the new intelligent computing center. The present invention makes full use of network resources, balances network load, effectively shortens network transmission delay, and alleviates the current dilemma of the network becoming a bottleneck for AI large model training. In order to deepen the understanding of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments of the specification.

[0053] Embodiment 1

[0054] See also Figure 1 and Figure 2 , the embodiment of the present invention provides an overall architecture of a multi-substream parallel forwarding scheduling mechanism in a new intelligent computing center network, which can effectively reduce the communication forwarding delay of the elephant flow, improve the utilization of network resources, and better meet the communication needs of the AI ​​large model distributed training business data flow. To a certain extent, it can make up for the gap between the current new intelligent computing center network and business needs, improve the efficiency and effect of large model training, and help the sustainable development of the AI ​​industry. In addition, the solution has the ability to adapt to the fusion of multiple modes of future networks.

[0055] In order to improve the effective bandwidth of the new intelligent computing center network, the core idea of ​​the embodiment of the present invention is to realize efficient parallel transmission of multiple sub-streams. For this purpose, the entire elephant flow communication forwarding process is divided into three stages: data flow segmentation and distribution stage, parallel transmission stage, and receiving end sorting and reorganization stage. The specific operations of each stage are as follows.

[0056] In the data flow segmentation and distribution phase, the sending node It is the main work object. First, the service request information needs to be forwarded to the central controller, and the information needs to include the data stream waiting for communication forwarding , Defined as a four-tuple .in, and Represents data streams The sending node and the receiving node, Indicates the length of the data stream, It is the longest communication forwarding time that the data flow can accept. The original data of data stream F needs to be split before distribution. Each data packet will be placed in multiple fixed-length containers of the same length in sequence according to the position order in the original data stream for standby. During the process, when the current fixed-length container is full or there is insufficient remaining space to place a complete data packet, the data packet will be placed in the next fixed-length container. A continuous sequence number is assigned to each fixed-length container in sequence as a unique identifier. This identifier is used by the receiving end to confirm the order and integrity so that it can be correctly reassembled into the original data stream. It is known that the data flow characteristics of the communication service of the large model training of the intelligent computing center are small number of flows and large single-flow bandwidth, that is, The amount of data will be very large, so using a fixed-length container for segmentation will not cause a lot of space waste. At the same time, compared with the variable-length segmentation solution, fixed-length containers are more convenient for sorting and reorganization at the receiving end, which can reduce unnecessary overhead and thus shorten the overall communication forwarding time. In addition, fixed-length containers can effectively avoid the polarization problem caused by variable segmentation lengths, that is, there will be no extreme situation where some data blocks are large and others are small after segmentation, which helps to balance the network load. The length of the fixed-length container is set in advance to , satisfying the conditions The sending node determines the optimal load situation corresponding to each sub-flow according to the configuration deployment information sent by the central controller, and distributes the divided fixed-length containers to the corresponding sub-flows in rounds for parallel transmission.

[0057] In the parallel transmission stage, in order to facilitate the analysis of data flow In multiple independent sub-streams available In order to reduce the latency of parallel transmission, we can uniformly model complex network topologies such as CLOS, FatTree, Dragonfly, etc. as directed graphs. ,in Indicates the network The set of terminal nodes, It represents the set of available subflows between all nodes in the network. Defined as From the sending node to the receiving node of A collection of available substreams. Considering that there are likely to be multiple network modes coexisting in the future intelligent computing center, the network modes of multiple substreams used in a data flow communication may be different, and they will have great differences in performance such as latency. This is directly reflected in the selection of different substreams when transmitting data of the same length. , the delay of parallel transmission The difference may be very large, so it is necessary to analyze and calculate the delay of each sub-flow independently. By sending delay and propagation delay It consists of two parts, and the formula is:

[0058]

[0059] Transmission delay is the time required for data to be transmitted within the network device, that is, the time from the first bit of data to the last bit of data. Transmission delay is determined by the length of the data and the transmission rate. Propagation delay refers to the time required for data to propagate a certain distance in the physical link in the form of electromagnetic waves. Propagation delay is closely related to the physical length of the channel and the propagation speed of the signal on the physical medium, but has nothing to do with the size of the data. For a certain network, it can be considered that the substream The corresponding propagation delay fluctuates within a certain range, using the symbol express.

[0060] In the receiving end sorting and reorganization phase, the main work is done by the receiving end node After sorting the out-of-order fixed-length containers by serial number and completing the reorganization, feedback is submitted to the central controller to complete the communication forwarding. Although the sending node distributes the fixed-length containers in order according to the serial number, due to the use of multiple sub-streams for parallel transmission, the path network mode corresponding to each sub-stream may be different, and there will be different delay characteristics. There are also differences in the real-time congestion conditions of the paths, which may cause the order in which the fixed-length containers actually arrive at the receiving end to be confused. The out-of-order fixed-length containers will be cached by the receiving end, waiting for the last fixed-length container of the same round to arrive and be confirmed before the sorting and reorganization operations are uniformly performed. The caching mechanism of the receiving node has a great impact on the operation of the stream reorganization stage. Considering the periodic characteristics of the large-model distributed training business, the data streams that its business requests need to forward often burst synchronously, and the receiving nodes in the intelligent computing center network often face many-to-one communication situations. In order to clearly distinguish the fixed-length containers after the different data streams forwarded by different sending nodes are split, the receiving end will open up an independent space for the corresponding data stream for classification caching after receiving the cache management information sent by the central controller. The detailed cache operation is: if the data packet is the first data packet to arrive at the receiving end on the sub-flow, the receiving end creates a new cache queue for it in the corresponding independent space for temporary storage; otherwise, the data packet is directly placed in the corresponding cache queue for temporary storage.

[0061] Combine the following Figure 1 and Figure 2 The specific steps of the embodiment of the present invention are described in detail:

[0062] Step 1: After receiving the Elephant Flow forwarding service request, the sender sends the key information in the request, such as the sender node , receiving node , length of elephant flow , Maximum communication delay Wait, upload to the central controller.

[0063] Step 2: The central controller obtains the physical network status in real time and determines the available senders in the network. To the receiver The number of available sub-flows for data forwarding and the currently remaining available physical bandwidth, propagation delay, and other information of the links corresponding to each sub-flow path.

[0064] Step 3: Taking the physical network status information obtained in step 2 as input, the central controller calculates the optimal diversion ratio of the current business request and configures the number of fixed-length load containers on the corresponding virtual path with the goal of minimizing the total delay of data flow communication forwarding and maximizing network resource utilization.

[0065] Step 4: The sender splits the original elephant flow, assembles the data packets into fixed-length containers of pre-set lengths in sequence while ensuring the integrity of the data packets, and assigns a continuous sequence number as a unique identifier to each fixed-length container in sequence. In the subsequent elephant flow forwarding process, the fixed-length container is used as the smallest unit for transmission in the network.

[0066] In step 5, the sending node performs mapping deployment of the physical network according to the configuration information issued in step 3, and performs parallel transmission in a round-by-round manner with sub-streams as the granularity. The total number of fixed-length containers transmitted in each round is based on the receiving end cache as a reference indicator. Under the premise that the cache can accommodate, each round distributes as many fixed-length containers as possible according to the diversion ratio in the configuration information for multi-sub-stream parallel transmission.

[0067] Step 6: The central controller sends cache management information to the corresponding receiving end. After receiving the information, the receiving end opens up an independent space for the data stream to classify and cache the out-of-order fixed-length containers.

[0068] Step 7: After each round of parallel transmission, the receiver sorts and reorganizes the random fixed-length containers in the cache and submits them to the upper layer. The receiver needs to feedback information to the central controller, which notifies the sender to start the next round of parallel transmission after receiving the information, until the entire business data stream is fully forwarded.

[0069] The specific operation of step 2 is described in detail below. To represent the elephant flow requesting communication forwarding, Can be defined as a four-tuple .in, and Represent the sending node and receiving node of the data flow respectively, Indicates the length of the data stream, is the longest communication forwarding time that the data flow can accept. Assume that When a communication forwarding request is made, the central controller obtains the Sliver streams are available to the sender node To the receiving node For data forwarding, use Indicates In the embodiment of the present invention, it is assumed that The sub-flows are independent of each other. The path used corresponds to the link .link The currently available physical bandwidth is defined as ,link Actual load data flow The number of fixed-length containers is . It can be represented by a tuple as In the scenario where multiple network modes coexist in the intelligent computing center, the links of these sub-flows may be built using different network modes, and they may have different performance indicators such as latency and bandwidth. arrive The corresponding propagation delays of the links may be very different. In order to better fit the actual situation of the intelligent computing center network, the model uses To estimate the link The propagation delay of the link It has nothing to do with the data allocated to it and can be obtained through the feedback information of the receiving end.

[0070] In traditional Ethernet, the flow is used as the smallest unit for scheduling and forwarding. The classic algorithm is the ECMP equal-cost multi-path routing algorithm, which hashes the data flows with the same five-tuple to the same path for transmission. It is known that the business data flow characteristics of AI large model training in the new intelligent computing center are small number of flows and large single-flow bandwidth. In this scenario, scheduling by flow is prone to polarization, that is, some elephant flows are concentrated and hashed to the same path. It is easy to cause a small number of paths in the entire network to be overloaded, while the link bandwidth of other paths is not effectively utilized, and the network load is seriously uneven. Another existing scheduling method is to select paths for data packets in the data flow one by one, and use the granularity of packets for balancing. Different data packets of the same flow can choose different paths for transmission. This packet-based scheduling method has the smallest granularity in theory, and can make full use of the available remaining bandwidth in the network for forwarding, but it will cause out-of-order problems when the data packets arrive at the receiving end. Severe out-of-order problems will cause a significant decrease in throughput, and require a large amount of additional out-of-order reorganization work that consumes cache and CPU resources. The out-of-order problem is an important reason for the poor performance of packet scheduling and the difficulty in supporting large-scale network traffic forwarding.

[0071] In order to solve the above problem, a more reasonable granularity is selected for scheduling and forwarding. The embodiment of the present invention designs a scheme for scheduling by sub-flow. First, step 4 divides the elephant flow. The specific operation is as follows: the sending end node After receiving the service request, the data packets are recorded in sequence into a fixed-length container with a preset length to the maximum extent possible under the premise of ensuring the integrity of the data packets. When the length of the fixed-length container is exceeded, the data packets are recorded into the next fixed-length container. The sender will assign a continuous sequence number as a unique identifier to each fixed-length container in sequence. This identifier is used by the receiver to confirm the order and integrity of the fixed-length container received so as to correctly reassemble the original data stream. In the parallel transmission process, the fixed-length container is used as the smallest operation unit for path selection to ensure that all data packets belonging to the same fixed-length container will be scheduled to the same sub-flow path for forwarding, and the data packets in the fixed-length container arrive at the receiving end in order. Different fixed-length containers can select different paths for transmission according to the configuration information issued by the central controller, realize multi-sub-flow parallelism, and improve the effective bandwidth of the network. Due to the difference in transmission delays of different sub-flows, there may be disorder between different fixed-length containers at the receiving end. It is only necessary to sort and reorganize the fixed-length containers. Compared with processing data packet disorder, unnecessary overhead can be reduced, thereby shortening the overall communication forwarding time and truly giving play to the advantages of multi-path parallelism.

[0072] Considering that if there is disorder between fixed-length containers, they need to be cached and sorted at the receiving end. Once there are too many fixed-length containers in the cache, it is likely to cause serious cache congestion and affect the global communication forwarding performance. Therefore, we will perform multi-substream parallel transmission in rounds and select an appropriate total number of fixed-length container transmissions for each round to ensure that the receiving end can process the disordered fixed-length containers in the cache in time after each round, while balancing the additional computing cost and time cost that may be caused by too many rounds. In step 5, the maximum number of fixed-length data packets that can be accommodated in the receiving end cache is As the basic indicator, the transmission completion time of multiple sub-flows in each round is synchronized as much as possible. The total number of packets transmitted in each round can be calculated by the following formula :

[0073]

[0074]

[0075]

[0076] Step 6 is the operation of caching the out-of-order fixed-length containers of the receiving end according to the sub-stream scheduling scheme designed in the embodiment of the present invention. The out-of-order fixed-length containers arriving at the receiving end will be cached by the receiving end, waiting for the last fixed-length container of the same round to arrive and be confirmed before performing the sorting and reorganization operation uniformly. The caching mechanism of the receiving end has a great impact on the operation of the stream reorganization stage. Considering the periodic characteristics of the large-model distributed training business, the data streams that need to be forwarded by its business requests often burst synchronously, and the receiving nodes in the intelligent computing center network often face many-to-one communication situations. In order to clearly distinguish different data streams forwarded by different sending end nodes, the receiving end opens up an independent space for each data stream for classified caching according to the cache management information issued by the central controller. The detailed caching operation is: if a fixed-length container is the first fixed-length container on the sub-stream that arrives at the receiving end and needs to be cached, the receiving end creates a new cache queue for it in the corresponding independent space for temporary storage; otherwise, the fixed-length container is directly placed in the corresponding cache queue for temporary storage. After each round of parallel transmission, the receiver sorts and reorganizes the out-of-order fixed-length containers in the cache and submits them to the upper layer, and feeds back information to the central controller, and finally clears the corresponding cache space to wait for the next round of fixed-length containers for multi-substream parallel transmission.

[0077] Embodiment 2

[0078] See also Figure 2 , an embodiment of the present invention provides a method for coordinating and cooperating between the sending-end splitting and distribution and the receiving-end sorting and reorganization strategies in a multi-substream parallel forwarding scheduling mechanism. The multi-substream parallel forwarding scheduling mechanism designed by the present invention provides efficient services for the new intelligent computing center network under the background of AI large-model distributed training business. Reducing the delay of data communication and forwarding between parallel computing devices through the network is crucial for large-model distributed training. Communication delay is not only related to the efficiency and cost of large-model training, but also affects the synchronous update of model parameters between nodes. The update lag of some nodes leads to a decrease in the convergence speed and final quality of the large model. Therefore, how to perform optimal sub-stream division and allocate a reasonable fixed-length container load to each sub-stream to minimize the communication forwarding delay of the elephant stream is the core optimization goal in the multi-substream parallel forwarding scheduling mechanism.

[0079] The embodiment of the present invention designs a multi-substream parallel forwarding scheduling mechanism to optimize the communication forwarding performance of the network facing ultra-large bandwidth service data streams. Total communication delay It consists of three parts: segmentation and distribution delay , Parallel transmission delay and sorting and reassembly delay , the formula is described as . The split distribution delay Less affected by different diversion operations The impact of can also be basically ignored, so in the multi-subflow parallel forwarding scheduling mechanism, The decisive factor is the parallel transmission delay and sorting and reassembly delay Because multiple sub-streams and multiple paths can make full use of the available bandwidth in the network, balance the network load, and improve the utilization of network resources, the effective bandwidth of the network is significantly increased, and the transmission delay of the data stream is reduced. In addition, by By analyzing the calculation formula, it can be roughly concluded that selecting more sub-streams to perform reasonable parallel transmission of the same data stream can further shorten the data stream transmission delay. . Sorting and reorganization delay It is the time it takes for the receiving end to sort the possibly disordered data packets received and reassemble them into a correct and complete original data stream. Qualitative analysis shows that the more sub-streams are actually used when the data stream is transmitted in parallel, the more serious the disorder problem will be when it reaches the receiving end. The receiving end will inevitably spend more time to do more complex sorting and reorganization to obtain the original data stream, that is, The larger the value of . In simple terms, It is a variable that is positively correlated with the total number of sub-flows. In practical applications, in addition to the number of sub-flows, the delay of the sorting and reorganization phase It will also be affected by other factors, but this model is mainly established to explore the optimal scheduling strategy for parallel transmission of multiple sub-streams of ultra-large bandwidth data streams. It is prudent to simplify the model by assuming that other minor influencing factors are exactly the same.

[0080] In summary, for different sub-stream division schemes of the same data stream, parallel transmission delay and sorting and reassembly delay These are two factors that influence and restrict each other and jointly determine the business data flow of the request forwarding. The total communication delay (minimize the total delay of data flow communication forwarding). In the embodiment of the present invention, an optimization algorithm is designed in the central controller to find the parallel transmission delay by trying different sub-flow division schemes. and sorting and reassembly delay The optimal balance is explored, and the cooperation between the sending end segmentation and distribution and the receiving end sorting and reorganization strategy is explored to achieve the total communication completion delay of the business data flow. Minimum, network resource utilization The biggest key decision goal. It should be noted that the central controller optimization algorithm designed in the embodiment of the present invention allows the decision result to degenerate into a special case of single-path scheduling, that is, for a specific service request and actual network status, when the algorithm minimizes the task completion time and the output corresponds to the optimization result, there is only one sub-flow path with load in the configuration information, the network will directly select the path according to the configuration information and complete the communication forwarding of the service data flow in a single-flow single-path manner.

[0081] Embodiment 3

[0082] See also Figure 3 , Figure 4 , an embodiment of the present invention provides a multi-objective optimization method for a central controller of a multi-substream parallel forwarding scheduling mechanism to determine the optimal diversion ratio. The constrained NSGA-II algorithm can determine the optimal number of divided substreams based on the real-time acquisition of the number of available substreams in the physical network that can be used for data forwarding from the sender to the receiver, as well as the current available physical bandwidth of the link corresponding to each substream path, the propagation delay and other information, and allocate a reasonable load to each substream, effectively shortening the communication forwarding delay of the elephant flow and improving the network resource utilization of the intelligent computing center. The constraint condition limits the total number of fixed-length containers transmitted by all substreams to be consistent with the number of fixed-length containers after the original elephant flow is split, which can ensure that the receiving end receives complete and non-repetitive data. The constrained NSGA-II algorithm specifically includes the following steps:

[0083] Step 1: Enter the preset data and the obtained network parameters, including the fixed-length container length , the length of the original elephant flow to be transmitted , the number of available sub-flows wait.

[0084] Step 2: Adjust the parameter settings of the optimization algorithm, including the number of iterations, population size, crossover and mutation operation distribution indicators, etc.

[0085] Step 3: Initialize the population, i.e. randomly generate a population of size The initial population of .

[0086] Step 4: Calculate the constraint violation degree of the solution and distinguish feasible solutions from infeasible solutions according to the rules

[0087] Step 5: Perform fast non-dominated sorting, following the principle of comparing the advantages and disadvantages of solutions by distinguishing feasible solutions from infeasible solutions under constraints.

[0088] Step 6: Calculate the crowding distance for each individual in the non-dominated layer according to the sorting results, and select appropriate individuals to form a new parent population based on the crowding comparison operator.

[0089] Step 7: Use a genetic algorithm to perform selection, crossover and mutation operations on the parent population to obtain the offspring population.

[0090] Step 8: merge the parent population with the offspring population to form a population of size The new population is formed and suitable individuals are selected to form the next parent generation.

[0091] Repeat steps 4 to 8 in this way until the number of iterations reaches the preset maximum value, then stop the loop, draw the corresponding Pareto frontier, and the algorithm program ends.

[0092] The following is a detailed description of how to initialize the population in step 3. The individuals in the population are The row vector consists of elements, which can be expressed as .in is the number of available subflows in the network for data forwarding from the sender to the receiver. Each element are all non-negative integers, corresponding to the number of fixed-length containers that need to be transmitted after segmentation allocated to each available sub-stream. Indicates that the available substream has not been assigned a data stream transmission task. In order to ensure that the receiving end receives complete and non-duplicate data, the constraints must be met when randomly generating the initial population. Therefore, the total number of fixed-length containers transmitted by all available substreams is The number of fixed-length containers to be split from the original data stream In summary, the population initialization operation can be specifically described as: randomly generating By The sum of integers is a fixed value The row vector of .

[0093] In optimization problems, constraints can be handled in a variety of ways, among which the penalty function method and the method of distinguishing feasible solutions from infeasible solutions are commonly used. The penalty function method needs to construct a penalty function according to the characteristics of the constraints, add the penalty function to the objective function, transform the constraint problem into an unconstrained problem, and then solve it using conventional methods. However, the penalty function method has the inherent defect that the penalty factor is difficult to set. The method of distinguishing feasible solutions from infeasible solutions is a processing technique based on the comparison principle. The feasible solution and the infeasible solution are distinguished according to whether they meet the constraints. The solution that meets the constraints is a feasible solution, and the solution that does not meet the constraints is defined as an infeasible solution. It can handle most constrained optimization problems on the basis of making minor changes to the unconstrained optimization algorithm, so that the optimal solution search moves closer to the feasible domain, and then quickly moves closer to the Pareto frontier.

[0094] Based on this, the present invention defines the violation value of each individual in step 4 . The calculation method is ,in is a row vector that satisfies . According to the calculation principle, if the individual does not meet the constraint conditions, the constraint violation value is the absolute value of the true constraint function value; if the individual meets the constraint conditions, the constraint violation value is 0. The constraint violation value is normalized to obtain the constraint violation degree, and the value range of the constraint violation degree is The constraint violation degree is used as the criterion to distinguish feasible solutions from infeasible solutions. Individuals with a constraint violation degree of 0 correspond to feasible solutions, and individuals with a constraint violation degree not equal to 0 correspond to infeasible solutions.

[0095] In step 5, in the constrained NSGA-II multi-objective optimization that distinguishes feasible solutions from infeasible solutions, the principle of comparing the advantages and disadvantages of the solutions of the fast non-dominated sorting is improved accordingly: (1) In any case, the feasible solution is unconditionally better than the infeasible solution. (2) The comparison between feasible solutions is still carried out according to the rules of the original unconstrained optimization algorithm. (3) The comparison between infeasible solutions follows the principle of giving priority to solutions with smaller constraint violations. The specific operation of the fast non-dominated sorting is as follows. Each solution is assigned two parameters: and ,in Indicates dominance The number of solutions to a solution, Indicates that The number of solutions dominated by a solution. First, we calculate the number of solutions dominated by a solution by traversing each solution. and It is easy to find that The solution is located on the first non-dominated layer. All solutions on the first non-dominated layer can be represented by the set For the set The solution in will be dominated by the solution Subtract 1. At this point, the new The solution is the individuals on the second non-dominated layer. The liberation of the collection By repeating the above steps, we can find all the non-dominated layer sets in order of rank and give each set a non-dominated rank.

[0096] In step 6, the crowding distance must be defined first. For each objective function, find two solutions that are adjacent to the function value corresponding to the current solution, and calculate the function difference between the two solutions. The result of summing up all the calculated differences is the crowding distance of the current solution. In particular, for individuals located at the boundary, due to the lack of adjacent solutions, the crowding distance cannot be calculated using the above definition and is set to infinity by default. From the above definition, it can be seen that the smaller the crowding distance, the more crowded the solution is. In order to ensure the uniformity of the distribution of solutions in the approximate Pareto solution set, a crowding comparison operator is designed based on the two attributes of the non-dominated rank and crowding distance of each individual in the population. It can be expressed as a partial order relationship:

[0097]

[0098] That is, solutions with lower rank values ​​are preferred. If two solutions have the same rank value, the solution with a larger distance value takes precedence over the solution with a smaller distance value.

[0099] The multi-substream parallel forwarding scheduling mechanism provided by the present invention serves the new intelligent computing center network in the context of AI large-model distributed training business. Reducing the delay of data forwarding through the network in parallel computing devices is crucial for large-model distributed training. The central controller provided in this embodiment determines the optimal diversion ratio optimization method, with minimizing the network forwarding delay of AI large-model training business data flow and maximizing the network resource utilization of the intelligent computing center as the two major optimization goals. The constrained NSGA-II multi-objective evolutionary optimization algorithm is used to determine the optimal number of divided substreams, and a reasonable load is allocated to each substream. This is a key step to ensure the efficient operation of the multi-substream parallel forwarding scheduling mechanism and that the new intelligent computing center network can meet the needs of large-model distributed training business.

[0100] The optimal split ratio of the elephant flow of the service request is determined according to the optimization result of the central controller, and the distribution plan of the sender is designed accordingly and mapped to the physical network to complete the parallel transmission of multiple sub-flows. For a service request, the total communication delay of the elephant flow under the multi-sub-flow parallel scheduling mechanism provided by the present invention is Delay distribution by slicing , Parallel transmission delay , Sorting and reorganization delay It consists of three parts. The parallel operation of multiple sub-streams can make full use of the available bandwidth of the network and effectively shorten the delay of data forwarding on the physical link. Although it will introduce additional delays for sorting and reorganization operations at the receiving end, the multi-objective optimization algorithm of the central controller can find the optimal solution by controlling the diversion situation and weighing the increase and decrease of delay to minimize the completion time of the elephant flow. In addition, multiple sub-streams can avoid the fragmentation of network resources caused by a single flow, alleviate the problem of elephant flows waiting in queue at the sending end because there is not enough remaining available bandwidth, and effectively improve the utilization of network resources.

[0101] The above description is only an embodiment of the present invention, which is of exemplary nature and is intended to illustrate the technical solution of the present invention, but the protection scope of the present invention is not limited thereto. Any technician in the field can make appropriate changes or modifications within the technical scope disclosed in the present invention. Such changes or modifications should be included in the protection scope of the present invention without violating the purpose of the present invention and the claims.

Claims

1. A multi-subflow parallel forwarding scheduling method for a new intelligent computing center network, characterized by: The distributed training service data stream of AI large models in the new intelligent computing center is forwarded and scheduled in parallel with multiple sub-streams, including: In the data stream segmentation and distribution phase, the service data stream is first segmented, and the data packets are sequentially placed into fixed-length containers of pre-set lengths for standby use, and a continuous sequence number is sequentially assigned to each fixed-length container as a unique identifier; the sender distributes the segmented fixed-length containers to the corresponding sub-streams in a round-robin manner according to the configuration information issued by the central controller; In the parallel transmission stage, the configuration information is mapped to the physical network, and the fixed-length container is used as the smallest operation unit for path selection. Different fixed-length containers select different sub-flow paths for forwarding, thus realizing parallel transmission with sub-flow as the granularity. In the sorting and reorganization phase at the receiving end, after each round of parallel transmission, the receiver sorts and reorganizes the out-of-order fixed-length containers in the cache and submits them to the upper layer; it also feeds back information to the central controller; and finally clears the corresponding cache space to wait for the next round of multi-substream parallel transmission.

2. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 1 is characterized in that: Before forwarding the current business data flow communication, the central controller integrates and sends configuration information and cache management information, including: After receiving the service request, the sender uploads the key information contained in the request to the central controller; The central controller obtains the physical network status in real time, determines the number of available sub-flows in the network for data forwarding from the sender to the receiver, and the link performance indicators corresponding to each sub-flow path, including the current available physical bandwidth and propagation delay; The acquired physical network status information is used as input, with the optimization goals of minimizing the total delay of data flow communication forwarding and maximizing the utilization of network resources, to find the best diversion ratio of the current business data flow and the corresponding optimal forwarding path, to achieve the optimal balance between parallel transmission delay and sorting and reorganization delay, and to integrate them into configuration information and cache management information, which are sent from the central controller to the data forwarding plane to control parallel forwarding scheduling.

3. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 1 is characterized in that: The data stream segmentation and distribution stage includes: The sender divides the service data stream requested for communication forwarding. The sending node records the data packets in order into fixed-length containers of a preset length to the maximum extent possible while ensuring the integrity of the data packets, and assigns a continuous sequence number as a unique identifier to each fixed-length container in order. After completing the splitting operation, the sending node determines the optimal load corresponding to each sub-stream based on the configuration information sent by the central controller, and takes the maximum number of fixed-length containers that the receiving end cache can accommodate as the basic indicator. It synchronizes the transmission completion time of multiple sub-streams in each round as much as possible, calculates the total number of fixed-length containers transmitted in each round, and distributes the split fixed-length containers to the corresponding sub-streams in rounds for parallel transmission.

4. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 1 is characterized in that: The parallel transmission stage includes: The segmented fixed-length container is used as the smallest operation unit for path selection. All data packets belonging to the same fixed-length container will be scheduled to the same sub-flow path for forwarding, ensuring that the data packets in the fixed-length container arrive at the receiving end in order. Different fixed-length containers are mapped to the physical network according to the configuration information sent by the central controller, and different paths are selected for transmission to achieve parallel transmission of multiple sub-flows.

5. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 1 is characterized in that: The receiving end sorting and reorganization stage includes: After receiving the cache management information sent by the central controller, the receiving end opens up an independent space for the current business data flow to classify and cache the disordered fixed-length containers; the disordered fixed-length containers arriving at the receiving end are cached by the receiving end. If a certain fixed-length container is the first fixed-length container on the substream that arrives at the receiving end and needs to be cached, the receiving end creates a new cache queue for it in the corresponding independent space for temporary storage; otherwise, the fixed-length container is directly placed in the corresponding cache queue for temporary storage; wait for the last fixed-length container of the same round to arrive and be confirmed, and then perform unified sorting and reorganization operations; After each round of parallel transmission, the receiver sorts the random fixed-length containers in the cache using the sequence number as the key, submits them to the upper layer after reorganization, and feeds back information to the central controller; finally, the corresponding cache space is cleared to wait for the next round of multi-substream parallel transmission; The central controller notifies the sender that it can start the next round of parallel transmission until the entire business data stream is fully forwarded.

6. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 2 is characterized in that: The key information in the service request includes: the sending node , receiving node , length of elephant flow , Maximum communication delay .

7. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 2 is characterized in that: A constrained multi-objective optimization algorithm is used to balance the increase and decrease of parallel transmission delay and sorting and reorganization delay, determine the best diversion ratio of the current business data flow, and find the corresponding optimal forwarding path; The constrained multi-objective optimization algorithm is specifically: By The sum of integers is a fixed value The row vector of is used as the initialization population, and different sub-flow division and forwarding path matching schemes are tried in multiple selection, crossover and mutation processes. The results after multiple iterations are controlled to continuously approach the optimal balance between the parallel transmission delay of multiple sub-flows and the sorting and reorganization delay. Under the conditions of minimizing the total communication forwarding delay of the business data flow and maximizing the network resource utilization as the key goals, the optimal diversion ratio of the current business data flow and the corresponding optimal forwarding path are found, and finally the optimal configuration of the number of fixed-length containers on each virtual sub-flow path is decided.

8. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 7 is characterized in that: The constraint processing method is specifically as follows: Define the violation value for each individual ; The calculation method is ,in is a row vector that satisfies , Indicates the total number of fixed-length containers after the business data stream is split. Indicates the number of fixed-length containers that can be transmitted by each available sub-stream, is the number of available subflows in the network for data forwarding from the sender to the receiver; if the individual does not meet the constraint conditions, the constraint violation value is the absolute value of the true constraint function value; if the individual meets the constraint conditions, the constraint violation value is 0; the constraint violation value is normalized to obtain the constraint violation degree, and the value range of the constraint violation degree is The multi-objective optimization evolutionary algorithm uses constraint violation as the standard to distinguish feasible solutions from infeasible solutions. Individuals with a constraint violation of 0 correspond to feasible solutions, and individuals with a constraint violation of non-zero correspond to infeasible solutions.

9. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 3 is characterized in that: The process of the sender segmenting the service data flow requesting communication forwarding is as follows: Sending Node First, the service request information is forwarded to the central controller, and the information needs to include the data stream waiting for communication forwarding , Defined as a four-tuple ,in, and Represents data streams The sending node and the receiving node, Indicates the length of the data stream, It is the longest communication forwarding time that the data flow can accept; Sending Node Before distributing the data The original data is split, and each data packet is sequentially placed into multiple fixed-length containers of the same length for standby according to the position order in the original data stream; when the current fixed-length container is full or the remaining space is insufficient to place a complete data packet, the data packet is switched to the next fixed-length container; a continuous sequence number is sequentially assigned to each fixed-length container as a unique identifier; the length of the fixed-length container is set in advance to , satisfying the conditions .

10. The multi-subflow parallel forwarding scheduling method for a novel intelligent computing center network according to claim 3 is characterized in that: The specific implementation of distributing the segmented fixed-length containers to the corresponding sub-streams in rounds for parallel transmission is as follows: The total number of fixed-length containers transmitted in each round is the number of caches on the receiving end. As a reference indicator, under the premise that the cache can accommodate, arrange as many fixed-length containers as possible for multi-substream parallel transmission according to the diversion ratio in the configuration information, and calculate the total number of data packets transmitted in each round by the following formula , ； ； ； in Indicates the number of fixed-length containers that can be transmitted by each available sub-stream. is the number of available sub-flows in the network for data forwarding from the sender to the receiver, is a positive integer variable; The sending node is based on the configuration deployment information issued by the central controller and , determine the optimal load situation of each sub-flow in each round, distribute the segmented fixed-length containers to the corresponding sub-flows in rounds, and prepare for subsequent mapping to the physical network for parallel transmission of multiple sub-flows.

Images