Deep learning model parameter synchronization-oriented intelligent calculation fusion network multipath transmission method
By using ant colony algorithm and active in-band network telemetry technology in intelligent computing converged networks, the congestion links are dynamically selected and replaced, and the multi-channel transmission of distributed deep learning model parameters is realized, solving the problem of path congestion in traditional networks , improves training efficiency and load balancing.
Patent Information
- Application Number
- CN202410885351.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-03
- Publication Date
- 2024-11-15
AI Technical Summary
In the process of parameter synchronization of distributed deep learning models, traditional network protocols are prone to path congestion and affect training efficiency. The existing technology rerouting mechanism and multiple transmission scheme have the problem of unoptimized path selection and aggravating network load imbalance.
By using ant colony algorithm in intelligent computing converged network, combining active in-band network telemetry technology INT detection packets and multiple transmission mechanism, dynamically replace the congested links, generate a set of multiple transmission paths, and in the switch The node adds forwarding rules to implement polling and forwarding of data packets.
It effectively reduces the possibility of path congestion, improves the speed of parameter synchronization of distributed deep learning models, meets the needs of diversified services, and improves the load balancing of the network.
Smart Images

Figure CN118972313A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer application technology, and in particular to an intelligent computing fusion network multi-channel transmission method for deep learning model parameter synchronization. Background Art
[0002] In recent years, in order to improve the processing speed of computing-intensive tasks such as distributed deep learning, data center networks have widely adopted the Ethernet-based Remote Direct Memory Access protocol (Remote Direct Memory Access over Converged Ethernet, RoCEv2) to replace the traditional TCP / IP protocol, thereby taking advantage of its high bandwidth and low latency to improve the parameter synchronization speed between nodes where distributed deep learning model training tasks are located.
[0003] Distributed deep learning mainly uses the Ring-all-Reduce (RAR) architecture for training in data center networks using the RoCEv2 protocol. In this architecture, each node only receives parameters from its left neighbor and sends parameters to its right neighbor. Only when all nodes have completed parameter synchronization for this batch will the next batch of training begin. The amount of parameter data that needs to be transmitted for each batch of parameter synchronization is extremely large. During this period, the link bandwidth occupancy rate in the network will increase significantly, which is likely to cause congestion on a certain link, prolong the communication delay between the nodes where the task is located, and thus affect the speed of distributed deep learning model parameter synchronization. Therefore, an efficient load balancing mechanism is needed to reduce the occurrence of network congestion.
[0004] However, traditional networks are unable to meet the diversified and differentiated network services of the future. The introduction of intelligent computing converged network theory effectively solves the problems currently faced by traditional networks. The intelligent computing converged network proposes a "three-layer three-domain" system architecture, innovatively introduces knowledge space, and provides differentiated and fine-grained network services for different business scenarios through intelligent scheduling strategies, thereby effectively meeting diversified service needs.
[0005] The prior art proposes an intra-network rerouting mechanism Conweave for the RoCE protocol. The mechanism deploys relevant components on the source Top of Rack (TOR) switch and the destination TOR switch respectively, periodically monitors the network status, and when it is found that the path between the two is congested, randomly selects some sample paths between the two. If these sample paths contain a non-congested path, the source routing technology is used to switch the flow to the path for transmission. If these sample paths are not available, it is considered that the network is in a highly congested state at this time, and the search for a new path for rerouting is abandoned.
[0006] The disadvantages of the above scheme include: the distributed deep learning parameter synchronization process requires a large bandwidth. When path congestion occurs, if it is only switched to a single path for transmission, it is very likely to cause path congestion again, affecting the transmission speed of the distributed deep learning parameters. In addition, this method selects a new replacement path in an approximately random manner, rather than selecting the best replacement path based on the specific status information of each path, which is very likely to aggravate the imbalance of the network load.
[0007] The prior art also proposes a software-based end-to-end multiplexing mechanism Maestro for the RoCE protocol. Multiple virtual network cards with different IP addresses are created on the server side, and these virtual network cards are bound to the same physical RoCE network card. When the sending server sends a certain flow, the sending server will split it into multiple small flows, which are transmitted through different virtual network cards in different paths without repeated links, and a dedicated memory is allocated on the receiving server for sorting between multiple small flows.
[0008] The disadvantages of the above scheme include: although the scheme realizes simultaneous multi-path transmission for the same traffic, the multi-path path it selects is an end-to-end path without duplicate links. When a core link in the data center network is congested and the load of the upstream aggregation link is also high, selecting the path according to this method is very likely to cause congestion of the aggregation link or the inability to discover more multi-path resources. Summary of the invention
[0009] An embodiment of the present invention provides an intelligent computing fusion network multi-path transmission method for deep learning model parameter synchronization, aiming to solve the problem that the distributed deep learning model parameter synchronization process in the RAR architecture occupies a large bandwidth, easily causes path congestion, and affects the deep learning model training efficiency.
[0010] In order to achieve the above object, the present invention adopts the following technical scheme.
[0011] A multi-channel transmission method for intelligent computing fusion network for deep learning model parameter synchronization, comprising:
[0012] The computing nodes in the intelligent computing fusion network periodically send active in-band network telemetry technology INT detection messages to the network to collect network status information through INT detection messages;
[0013] The controller in the intelligent computing fusion network determines whether the transmission path of the current task is congested by reading the information collected by the INT detection data packet and the deployment information of the distributed deep learning task of the full-ring RAR architecture. When congestion is detected, the ant colony algorithm is called to select the optimal path combination;
[0014] The controller uses the optimal path combination to replace the corresponding congested link in the original transmission path, generates a multi-path transmission path set between task hosts, and adds new forwarding rules to all switch nodes involved in the multi-path transmission path set.
[0015] Preferably, the computing nodes in the intelligent computing fusion network periodically send active in-band network telemetry technology INT detection messages to the network, and collect network status information through the INT detection messages, including:
[0016] The controller in the intelligent computing fusion network periodically sends INT detection messages to the network, and collects network status information through INT detection messages. The forwarding path of the INT detection message is specified by the source route. By setting the TYPE field of the link layer header to 0x1111, it is specified that the subsequent source routing header is the source routing header. The Last_header field of the last source routing header is set to 1, indicating that it is followed by an IP header. The protocol field of the IP header is set to 233 to indicate the number of subsequent INT headers, and the initial value is 0. When the message is forwarded in the network, the corresponding source routing header is removed for each hop switch, and the corresponding INT header is added after the INT_c header. Finally, all the source routing headers of the INT detection message received at the receiving end are removed, and the corresponding INT header is added;
[0017] The receiving end receives the INT detection message and sends a message carrying the switch serial number, port number, queue depth, bandwidth occupancy and delay information to the controller.
[0018] Preferably, the controller determines whether the transmission path of the current task is congested by reading the information collected by the INT detection data packet and the deployment information of the distributed deep learning task of the full-ring RAR architecture. When congestion is detected, the ant colony algorithm is called to select the optimal path combination, including:
[0019] The controller reads the information collected by the INT detection data packet and the deployment information of the distributed deep learning task of the RAR architecture to determine whether the transmission path of the model parameters between the hosts where the current task is located is congested. When congestion is detected, the controller records the specific congested link and adds all the non-congested paths between the source node of the congested link and the next node of the destination node to the candidate path set, and calls the ant colony algorithm to select the optimal path combination in the candidate path set.
[0020] The processing process of the ant colony algorithm includes:
[0021] 1. When the controller detects that the transmission path of the model parameters between the hosts where the current task is located is congested, the current network status information, candidate path set and iteration number are input into the ant colony algorithm;
[0022] 2. Initialize the pheromone level and ant colony of each path in the candidate path set;
[0023] 3. The ants in the ant colony calculate the path cost of each path, and calculate the probability of being selected based on the pheromone level and cost of each path, and sum them up to get the total selection probability P sum , where the path cost C i The link cost is obtained by summing the costs of all links in the path. The calculation formula is shown in Formula 2-1, where u and v represent the source node and destination node of the link respectively. is the bandwidth utilization of the link, c uv is the link bandwidth, is the maximum queue depth allowed by the port corresponding to switch v, l v is the actual queue depth of the port corresponding to switch v at this time, d uv is the link delay, The maximum delay allowed for the task;
[0024]
[0025] Path Q i The formula for calculating the probability of being selected is shown in Formula 2-2, where N i For path Q i pheromone concentration;
[0026]
[0027] 4. The ants traverse all candidate paths. When the probability of the path being selected is greater than or equal to the randomly generated path in (0, P sum ), add it to your own path combination until the total probability of the selected path combination is greater than 0.65*P sum When , the traversal ends (lines 6-12), and the total cost of the path set M selected by the ant is calculated (line 13). The calculation formula of the total cost is shown in formula 2-3, where maxT f_M is the maximum delay of the entire path after replacing the congested link with the selected path combination, maxT RAR is the maximum transmission delay between hosts of the current task, max T M is the maximum path delay in the selected path combination, minT M is the minimum path delay in the selected path combination;
[0028]
[0029] 5. Update the pheromone level of each path in the selected path combination according to the total cost of the selected path combination. The pheromone volatility is 0.2. The specific calculation formula is shown in Formula 2-4 (line 14);
[0030] N i =N i *0.8+1 / C sum Formula 2-4
[0031] 6. Repeat steps 3-5 until the maximum number of iterations is reached;
[0032] 7. Select the path combination with the minimum total cost selected by the ants in the ant colony after the last iteration as the optimal path combination.
[0033] Preferably, the controller uses the optimal path combination to replace the corresponding congested link in the original transmission path, generates a multi-path transmission path set between task hosts, and adds new forwarding rules to all switch nodes involved in the multi-path transmission path set, including:
[0034] The controller replaces the corresponding congested link in the original path with the optimal path combination, generates a multi-path transmission path set between task hosts, traverses all switch nodes involved in the multi-path transmission path set, clears the original forwarding rules for the destination host, and then adds new forwarding rules to all switch nodes: if a switch node has only a single next hop, the IPv4 forwarding flow table is sent down normally; if there are multiple next hops, a multi-path forwarding flow table is sent down, so that the data packets sent to the task destination host through the switch are forwarded between multiple next hops in a polling manner;
[0035] When the multi-path start node receives a data packet whose destination IP address is the task destination host, a hash operation is performed based on the data packet's source IP address, destination IP address, source port, and destination port to obtain the flow identifier to which it belongs, and then the data packet sequence number value stored in the register corresponding to the flow identifier is updated, and then the total number of next hops is modulo the data packet sequence number value to obtain the next hop path identifier, and then the flow table is matched according to the path identifier to forward the data packet.
[0036] It can be seen from the technical solution provided by the above-mentioned embodiments of the present invention that the present invention designs a multi-channel transmission mechanism for synchronization of deep learning model parameters of the RAR architecture, and adds the mechanism to the knowledge space of the intelligent computing fusion network. When the intelligent computing fusion network processes distributed deep learning tasks using the RAR architecture, the mechanism can be called to effectively speed up the synchronization speed of model parameters of this type of tasks, thereby better meeting user service needs.
[0037] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0039] Figure 1 A processing flow chart of a method for intelligent computing fusion network multi-channel transmission for deep learning model parameter synchronization provided by an embodiment of the present invention;
[0040] Figure 2 A source end INT detection data message structure diagram provided by an embodiment of the present invention;
[0041] Figure 3 A diagram of the structure of a destination INT detection datagram message provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0042] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.
[0043] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.
[0044] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.
[0045] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.
[0046] The embodiment of the present invention designs a multi-path load balancing mechanism within the network under the framework of the intelligent computing fusion network, aiming to solve the problem that the distributed deep learning model parameter synchronization process in the RAR architecture occupies a large bandwidth, which easily causes path congestion and affects the training efficiency of the deep learning model. The present invention mainly realizes related functions by using the controller of the mapping adaptation layer in the intelligent computing fusion network and the physical computing nodes and switches of the fusion network layer.
[0047] The processing flow of a method for intelligent computing fusion network multi-channel transmission for deep learning model parameter synchronization provided by an embodiment of the present invention is as follows: Figure 1 As shown in the figure, the control plane periodically monitors the network status, selects the optimal path combination based on the collected network status information, and then generates a flow table and sends it to the corresponding switch. The data plane forwards the deep learning model parameters according to the updated flow table.
[0048] At present, deep learning models widely use RAR distributed architecture for training. The same deep learning model is deployed on different computing nodes for training at the same time, and the model parameters generated by the training of different computing nodes need to be periodically sent to other nodes participating in the model training for synchronization. The node-to-node flow association characteristics of the RAR architecture require that the node that completes the parameter transfer first needs to wait until all nodes have completed the parameter transfer before starting the next batch of parameter transfer until the deep learning model training is completed. This process is the deep learning model parameter synchronization process, and the purpose of the present invention is to speed up the completion of this process by providing an intelligent computing network multi-channel transmission method.
[0049] The processing flow of a method for intelligent computing fusion network multi-channel transmission for deep learning model parameter synchronization provided by an embodiment of the present invention includes the following processing steps:
[0050] Step S10: Periodic monitoring of network status.
[0051] The computing nodes in the converged network layer of the intelligent computing converged network periodically send INT detection messages to the network by using active in-band network telemetry technology (INT). The INT detection messages collect network status information, thereby avoiding the congestion perception distortion problem that occurs in traditional end-to-end networks. The forwarding path of the INT detection message is specified by the source routing technology, and its data packet format is as follows: Figure 2 As shown in the figure, by setting the TYPE field of the link layer header to 0x1111, it is specified that the next header is the source routing header, and then the Last_header field of the last source routing header is set to 1, indicating that the next header is the IP header, and then the protocol field of the IP header is set to 233 to indicate that the next header is the number of INT headers (INT_c), and the initial value is 0. When the message is forwarded in the network, each time it passes through a hop switch, the corresponding source routing header is removed, and the corresponding INT header is placed after the INT_c header. The structure of the INT detection message finally received at the receiving end is as follows Figure 3 As shown, all source routing headers are removed and the corresponding INT headers are added.
[0052] The receiving end receives the above-mentioned INT detection message and sends a message carrying the switch serial number, port number, queue depth, bandwidth occupancy and delay information to the mapping adaptation layer controller in the intelligent computing fusion network. After receiving the message, the controller stores the information carried by the message in a file in the corresponding JSON format.
[0053] The receiving and sending ends of the INT detection message are both physical computing nodes of the fusion network layer in the intelligent computing fusion network, such as servers.
[0054] Step S20: Selecting the optimal path combination.
[0055] The controller reads the information collected by the INT detection data packet and the deployment information of the distributed deep learning task of the RAR architecture to determine whether the transmission path of the model parameters between the hosts where the current task is located is congested. When congestion is detected, the specific congested link is recorded, and all the uncongested paths between the source node of the congested link and the next node of the destination node are added to the candidate path set, and then the ant colony algorithm is called to select the optimal path combination in the candidate path set.
[0056] The overall workflow of the ant colony algorithm is shown in Table 1:
[0057] 1. When the controller detects that the transmission path of the model parameters between the hosts where the current task is located is congested, it will input the current network status information, candidate path set and number of iterations into the ant colony algorithm.
[0058] 2. Initialize the pheromone level and ant colony of each path in the candidate path set (line 1).
[0059] 3. The ants in the ant colony calculate the path cost of each path, and calculate the probability of being selected based on the pheromone level and cost of each path, and then sum them up to get the total selection probability P sum (Lines 4-5). Where the path cost C i The link cost is obtained by summing the costs of all links in the path. The calculation formula is shown in Formula 2-1, where u and v represent the source node and destination node of the link respectively. is the bandwidth utilization of the link, c uv is the link bandwidth, is the maximum queue depth allowed by the port corresponding to switch v, l v is the actual queue depth of the port corresponding to switch v at this time, d uv is the link delay, The maximum delay allowed for the task.
[0060]
[0061] Path Q i The formula for calculating the probability of being selected is shown in Formula 2-2, where N i For path Q i pheromone concentration.
[0062]
[0063] 4. Then the ant traverses all candidate paths, and when the probability of the path being selected is greater than or equal to the randomly generated one in (0, P sum ), add it to your own path combination until the total probability of the selected path combination is greater than 0.65*P sum When , the traversal ends (lines 6-12), and the total cost of the path set M selected by the ant is calculated (line 13). The calculation formula of the total cost is shown in formula 2-3, where maxT f_M is the maximum delay of the entire path after replacing the congested link with the selected path combination, maxT RAR is the maximum transmission delay between hosts of the current task, max T M is the maximum path delay in the selected path combination, minT M is the minimum path delay in the selected path combination.
[0064]
[0065] 5. Then, the pheromone level of each path in the selected path combination is updated according to the total cost of the selected path combination. The pheromone volatility is 0.2. The specific calculation formula is shown in Formula 2-4 (line 14).
[0066] N i =N i *0.8+1 / C sum Formula 2-4
[0067] 6. Repeat steps 3-5 until the maximum number of iterations is reached
[0068] 7. Select the path combination with the minimum total cost selected by the ants in the ant colony after the last iteration as the optimal path combination.
[0069] Table 1 Multi-path selection algorithm
[0070]
[0071]
[0072] Step S30: Update the corresponding switch flow table.
[0073] The present invention implements the functions of issuing, updating and forwarding data packets of the switch flow table based on the BMV2 programmable switch and the P4 programming language.
[0074] Since the optimal path combination selected by the ant colony algorithm is only a part of the model parameter transmission path between hosts, after the ant colony algorithm generates the optimal path combination, the optimal path combination is used to replace the corresponding congested link in the original path to generate a multi-path transmission path set between task hosts. Then, all switch nodes involved in the multi-path transmission path set are traversed, and after clearing the original forwarding rules for the destination host, new forwarding rules are added to all switch nodes: if a switch node has only a single next hop, the IPv4 forwarding flow table is sent normally; if there are multiple next hops, a multi-path forwarding flow table is sent, so that the data packet sent to the task destination host through the switch is forwarded between multiple next hops in a polling manner. Among them, the specific process of multi-path forwarding is as follows: when the multi-path start node receives a data packet with a destination IP address as the task destination host, a hash operation is performed according to the source IP address, destination IP address, source port and destination port of the data packet to obtain the flow identifier to which it belongs, and then the data packet sequence number value stored in the register corresponding to the flow identifier is updated, and then the total number of next hops is modulo the data packet sequence number value to obtain the next hop path identifier, and then the data packet is forwarded according to the path identifier matching flow table.
[0075] In summary, the embodiment of the present invention implements a multi-path transmission mechanism for distributed deep learning model parameter synchronization. When a link in a certain path is congested, the present invention selects multiple paths between the source node of the congested link and the next node of the destination node for data transmission, reducing path changes, thereby reducing the possibility of path changes causing congestion again; in addition, when selecting multi-path paths, the present invention adds the characteristics of the RoCE protocol and the RAR architecture into the consideration indicators, which is more targeted.
[0076] The multi-path transmission mechanism designed by the present invention for distributed deep learning model parameter synchronization is proposed for the full ring architecture and the RoCE protocol, taking into account the unique needs of the two, such as the difference in communication delay between hosts in the full ring architecture, the difference in delay between each path within the selected multi-path combination, etc.; at the same time, the result generated by the multi-path selection algorithm can be automatically converted into a flow table and sent to the corresponding switch. From this aspect, it is difficult to find other alternatives, but the specific details in the invention, such as the multi-path selection algorithm, can be replaced.
[0077] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.
[0078] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.
[0079] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.
[0080] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A multi-channel transmission method for intelligent computing fusion network for deep learning model parameter synchronization, It is characterized in that include: The computing nodes in the intelligent computing fusion network periodically send active in-band network telemetry technology INT detection messages to the network to collect network status information through INT detection messages; The controller in the intelligent computing fusion network determines whether the transmission path of the current task is congested by reading the information collected by the INT detection data packet and the deployment information of the distributed deep learning task of the full-ring RAR architecture. When congestion is detected, the ant colony algorithm is called to select the optimal path combination; The controller uses the optimal path combination to replace the corresponding congested link in the original transmission path, generates a multi-path transmission path set between task hosts, and adds new forwarding rules to all switch nodes involved in the multi-path transmission path set.
2. The method according to claim 1, It is characterized in that The computing nodes in the intelligent computing fusion network periodically send active in-band network telemetry technology INT detection messages to the network, and collect network status information through INT detection messages, including: The computing nodes in the intelligent computing fusion network periodically send INT detection messages to the network, and collect network status information through INT detection messages. The forwarding path of the INT detection message is specified by the source route. By setting the TYPE field of the link layer header to 0x1111, it is specified that the subsequent source routing header is the source routing header. The Last_header field of the last source routing header is set to 1, indicating that it is followed by an IP header. The protocol field of the IP header is set to 233 to indicate the number of subsequent INT headers, and the initial value is 0. When the message is forwarded in the network, the corresponding source routing header is removed for each hop switch, and the corresponding INT header is added after the INT_c header. Finally, all the source routing headers of the INT detection message received at the receiving end are removed, and the corresponding INT header is added; The receiving end receives the INT detection message and sends a message carrying the switch serial number, port number, queue depth, bandwidth occupancy and delay information to the controller.
3. The method according to claim 2, It is characterized in that The controller in the intelligent computing fusion network determines whether the transmission path of the current task is congested by reading the information collected by the INT detection data packet and the deployment information of the distributed deep learning task of the full-ring RAR architecture. When congestion is detected, the ant colony algorithm is called to select the optimal path combination, including: The controller in the intelligent computing fusion network reads the information collected by the INT detection data packet and the deployment information of the distributed deep learning task of the RAR architecture to determine whether the transmission path of the model parameters between the hosts where the current task is located is congested. When congestion is detected, the specific congested link is recorded, and all non-congested paths between the source node of the congested link and the next node of the destination node are added to the candidate path set, and the ant colony algorithm is called to select the optimal path combination in the candidate path set; The processing process of the ant colony algorithm includes:
1. When the controller detects that the transmission path of the model parameters between the hosts where the current task is located is congested, the current network status information, candidate path set and iteration number are input into the ant colony algorithm; 2. Initialize the pheromone level and ant colony of each path in the candidate path set; 3. The ants in the ant colony calculate the path cost of each path, and calculate the probability of being selected based on the pheromone level and cost of each path, and sum them up to get the total selection probability P sum , where the path cost C i The link cost is obtained by summing the costs of all links in the path. The calculation formula is shown in Formula 2-1, where u and v represent the source node and destination node of the link respectively. is the bandwidth utilization of the link, c uv is the link bandwidth, is the maximum queue depth allowed by the port corresponding to switch v, l v is the actual queue depth of the port corresponding to switch v at this time, d uv is the link delay, The maximum delay allowed for the task; Path Q i The formula for calculating the probability of being selected is shown in Formula 2-2, where N i For path Q i pheromone concentration; 4. The ants traverse all candidate paths. When the probability of the path being selected is greater than or equal to the randomly generated path in (0, P sum ), add it to your own path combination until the total probability of the selected path combination is greater than 0.65*P sum , the traversal ends (lines 6-12), and the total cost of the path set M selected by the ant is calculated (line 13), where the total cost is calculated as shown in formula 2-3. in, maxT f_M is the maximum delay of the entire path after replacing the congested link with the selected path combination, maxT RAR is the maximum transmission delay between hosts of the current task, maxT M is the maximum path delay in the selected path combination, minT M is the minimum path delay in the selected path combination; 5. Update the pheromone level of each path in the selected path combination according to the total cost of the selected path combination. The pheromone volatility is 0.
2. The specific calculation formula is shown in Formula 2-4 (line 14); N i =N i *0.8+1 / C sum Formula 2-4 6. Repeat steps 3-5 until the maximum number of iterations is reached; 7. Select the path combination with the minimum total cost selected by the ants in the ant colony after the last iteration as the optimal path combination.
4. The method according to claim 3, It is characterized in that The controller replaces the corresponding congested link in the original transmission path with the optimal path combination, generates a multi-path transmission path set between task hosts, and adds new forwarding rules to all switch nodes involved in the multi-path transmission path set, including: The controller replaces the corresponding congested link in the original path with the optimal path combination, generates a multi-path transmission path set between task hosts, traverses all switch nodes involved in the multi-path transmission path set, clears the original forwarding rules for the destination host, and then adds new forwarding rules to all switch nodes: if a switch node has only a single next hop, the IPv4 forwarding flow table is sent down normally; if there are multiple next hops, a multi-path forwarding flow table is sent down, so that the data packets sent to the task destination host through the switch are forwarded between multiple next hops in a polling manner; When the multi-path start node receives a data packet whose destination IP address is the task destination host, a hash operation is performed based on the data packet's source IP address, destination IP address, source port, and destination port to obtain the flow identifier to which it belongs, and then the data packet sequence number value stored in the register corresponding to the flow identifier is updated, and then the total number of next hops is modulo the data packet sequence number value to obtain the next hop path identifier, and then the flow table is matched according to the path identifier to forward the data packet.
