Runtime multi-task management method and system for extended security aggregation acceleration service

By precompiling the aggregator table and dynamic allocation aggregator on a programmable switch, combining the three-stage verification algorithm to process multi-GAIN task messages, the problem of long training duration and lack of multi-task management in federated learning is solved, and efficient Multitasking management and privacy protection.

CN118972350APending Publication Date: 2024-11-15NAT UNIV OF DEFENSE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410960504.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

During the training process, existing federated learning technology has a long training duration due to the huge communication overhead of model updates and lacks runtime multi-task management capabilities, which affects business performance.

Method used

By precompiling the aggregator table on a programmable switch and dynamically allocating the aggregator, a three-stage verification heuristic algorithm is used for routing calculations, ensuring that the routing that meets the constraints process multi-GAIN task messages, realizing runtime multi-task management.

Benefits of technology

Effectively support runtime multi-task management, avoiding the central server from collecting data from a single client, protecting user privacy, reducing bandwidth and resource overhead, and improving business performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118972350A_ABST
    Figure CN118972350A_ABST
Patent Text Reader

Abstract

The present application relates to a runtime multitask management method and system for extended secure aggregation acceleration services by extending a GAIN to support runtime multitask management that pre-compiles an aggregator table on each programmable switch and dynamically allocates aggregators on the aggregator table between different federated learning tasks, a GAIN task is expanded to a multi-task application, different tasks can use an aggregator on the same switch and do not interfere with each other, and a route of a federated learning task is calculated through an efficient task routing algorithm in a multi-task management process. The situation that the central server collects data from a single client side is avoided, it is guaranteed that the data of each client side is aggregated at least once, the data privacy of a user is protected, and the effect that the GAIN has the multi-task management capacity during operation is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of data task processing, and relates to a runtime multi-task management method and system for extending a secure aggregation acceleration service. Background Art

[0002] Federated learning (FL) is a privacy-preserving training technique in which multiple devices collaborate to train a global model. In each round of training, a central server updates the global model weights (i.e., model parameters) by aggregating local model updates from a large number of clients using federated averaging or its derivative algorithms, and then distributes the updated global model back to these clients for the next round of training. The size of model updates typically ranges from megabytes to gigabytes, and there may be hundreds of clients, so the training process usually incurs huge communication overhead and greatly affects the duration of training.

[0003] The emergence of programmable switches enables traditional network devices to provide a higher level of flexibility and customization. Many studies advocate the deployment of programmable switches at the edge of the network to achieve network autonomy and adaptability, ultimately improving the business performance of the edge network. When the traffic of multiple clients converges to the same switch, the switch can actively aggregate and forward the aggregated traffic to the downstream, thereby reducing traffic and communication delays. This method is called in-network aggregation. In addition, the switch can aggregate the received messages immediately and only temporarily store part of the aggregated results to achieve stateless aggregation. Based on this, existing studies have designed a secure aggregation acceleration service (GAIN) designed for federated learning to meet practical application needs. However, the above technologies do not yet have the ability to manage multiple tasks at runtime. Summary of the invention

[0004] In view of the problems existing in the above-mentioned traditional methods, the present invention proposes a runtime multi-task management method for extended security aggregation acceleration service and a runtime multi-task management system for extended security aggregation acceleration service, which can efficiently support runtime multi-task management.

[0005] In order to achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0006] On the one hand, a runtime multi-task management method for extending a security aggregation acceleration service is provided, comprising the steps of:

[0007] Receive messages from multiple GAIN tasks;

[0008] According to the message content of the message, the matching rule table inside the aggregation interaction machine is searched, the task identity of the message is returned, and the offset of the matched aggregator table is retrieved; the aggregator table is precompiled on the aggregation interaction machine;

[0009] Match the message to the aggregator in the aggregator table according to the offset;

[0010] The three-stage verification heuristic algorithm is used to calculate the routing of the aggregation switch to determine the routing that meets the constraints. The three-stage verification heuristic algorithm includes verifying the initial routing, verifying the aggregation node, and verifying the number of aggregations. The constraints include that the remaining aggregation capacity when using the aggregation switch for aggregation is greater than or equal to the rate required by multiple GAIN tasks, that the routing meets the bandwidth constraints of each link, and that the data of each client is aggregated at least once.

[0011] The messages of multiple GAIN tasks are aggregated according to the routes that meet the constraints.

[0012] On the other hand, a runtime multi-task management system for extending the security aggregation acceleration service is also provided, including:

[0013] Message receiving module, used to receive messages of multiple GAIN tasks;

[0014] A rule matching module is used to search the matching rule table inside the aggregation interaction machine according to the message content of the message, return the task identity of the message and retrieve the offset of the matched aggregator table; the aggregation interaction machine is precompiled with the aggregator table;

[0015] An aggregation matching module, used to match the message to the aggregator in the aggregator table according to the offset;

[0016] The routing calculation module is used to perform routing calculation of the aggregation switch using a three-stage verification heuristic algorithm to determine the routing that meets the constraints. The three-stage verification heuristic algorithm includes verifying the initial routing, verifying the aggregation node, and verifying the number of aggregations. The constraints include that the remaining aggregation capacity when using the aggregation switch for aggregation is greater than or equal to the rate required by multiple GAIN tasks, that the routing meets the bandwidth constraints of each link, and that the data of each client is aggregated at least once.

[0017] The aggregation processing module is used to complete the aggregation processing of the messages of multiple GAIN tasks according to the routes that meet the constraint conditions.

[0018] One of the above technical solutions has the following advantages and beneficial effects:

[0019] The runtime multi-task management method and system of the extended security aggregation acceleration service, by extending GAIN to support runtime multi-task management, precompiles an aggregator table on each programmable switch and dynamically allocates aggregators on the aggregator table between different federated learning tasks, so as to extend the GAIN task to multi-task applications. Different tasks can use the aggregator on the same switch without interfering with each other. In the multi-task management process, the routing of the federated learning task is calculated through an efficient task routing algorithm, which avoids the central server collecting data from a single client and ensures that the data of each client is aggregated at least once, thereby protecting the user's data privacy and achieving the effect of enabling GAIN to have the ability of runtime multi-task management. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the conventional technology, the drawings required for use in the embodiments or the conventional technology descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A schematic diagram of a flow chart of a runtime multi-task management method for extending a security aggregation acceleration service in one embodiment;

[0022] Figure 2 A schematic diagram of a routing example in an embodiment;

[0023] Figure 3 A schematic diagram of adjusting a running task to a new area in one embodiment, wherein (a) is an ideal situation and (b) is an aggregate interruption;

[0024] Figure 4 Schematic diagram of bandwidth overhead under different topologies and different proportions of programmable switches in a network in one embodiment, wherein (a) is the normalized bandwidth overhead in different topologies, and (b) is the normalized bandwidth overhead when programmable switches replace traditional switches in different proportions;

[0025] Figure 5 Schematic diagram of the impact of different numbers of clients on bandwidth overhead and request completion rate in one embodiment, wherein (a) is the impact of the number of clients on bandwidth overhead, and (b) is the request completion rate;

[0026] Figure 6 A schematic diagram of the module structure of a runtime multi-task management system for extending the security aggregation acceleration service in one embodiment. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0029] It should be noted that the reference to "embodiment" in this document means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The presentation of the phrase at various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments.

[0030] Those skilled in the art will appreciate that the embodiments described herein may be combined with other embodiments. The term "and / or" used in the present specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0031] The following will describe the implementation of the present invention in detail with reference to the accompanying drawings in the embodiment diagram of the present invention.

[0032] In one embodiment, Figure 1 As shown, a runtime multi-task management method for an extended security aggregation acceleration service is provided, and the method may include the following processing steps S12 to S20:

[0033] S12, receiving the message of multiple GAIN tasks;

[0034] S14, searching the matching rule table inside the aggregation interaction machine according to the message content of the message, returning the task identity of the message and retrieving the offset of the matched aggregator table; the aggregation interaction machine is precompiled with the aggregator table;

[0035] S16, matching the message to an aggregator in the aggregator table according to the offset;

[0036] S18, using a three-stage verification heuristic algorithm to perform routing calculations on the aggregation switch to determine a route that satisfies the constraints; the three-stage verification heuristic algorithm includes verifying the initial route, verifying the aggregation node, and verifying the number of aggregations; the constraints include that the remaining aggregation capacity when using the aggregation switch for aggregation is greater than or equal to the rate required by multiple GAIN tasks, that the route satisfies the bandwidth constraints of each link, and that the data of each client is guaranteed to be aggregated at least once;

[0037] S20, completing aggregation processing of the messages of multiple GAIN tasks according to the routes that meet the constraint conditions.

[0038] It can be understood that this embodiment supports runtime multi-task management by extending GAIN. Configuring a new task on an aggregation switch requires recompiling the switch register to provide a new aggregator for the task, which requires the switch to be restarted. Therefore, in the process of coping with the configuration of multiple tasks, the edge network may be frequently interrupted. In order to achieve runtime multi-task management, this embodiment precompiles a set of aggregators on each programmable switch, called an aggregator table, and dynamically allocates the aggregators on the aggregator table between different federated learning tasks.

[0039] Specifically, the TaskID (task identity) in the data packet header is used to specify data packets from different GAIN tasks. Users can obtain TaskID by applying for GAIN service from the network provider. Inside each aggregation switch, there is a matching rule table, which records the offset (denoted as Offset) of each task that needs to be aggregated on the aggregation switch on the aggregator table. After the message arrives, the aggregation switch searches the matching rule table according to the message content to return the TaskID of the message and retrieve the corresponding Offset. If any matching rule in the matching rule table is not matched, the aggregation switch directly forwards the message; otherwise, the aggregation switch matches the message to the aggregator in the aggregator table, and the index idx of the aggregator table can be calculated as:

[0040] idx=Offset+pkt.SN%2L (1)

[0041] Among them, pkt.SN represents the sequence number of the currently arriving message, and L represents the number of aggregators on the aggregation switch. If the aggregator completes the aggregation, then the aggregation switch should initialize another corresponding aggregator. The formula that allows the aggregation switch to calculate an aggregator index and initialize a corresponding aggregator is written as:

[0042] idx=Offset+((pkt.SN%2L)+L)%2L (2)

[0043] In this way, different GAIN tasks can use the aggregators on the same aggregation switch without interfering with each other. Once a task is completed, the network provider can reclaim its corresponding resources by deleting the corresponding matching rules and initializing the task's aggregators according to the offset Offset and the number of aggregators L configured for the task. These aggregators can be reallocated to new tasks at runtime.

[0044] In the process of multi-task management, an efficient task routing algorithm is proposed. Specifically, the source node of a request is represented by S = {s1, s2, ...}, and the destination node is represented by dst. The required transmission rate (i.e., throughput) is represented by r, which can be specified by the user according to his own needs. The network is modeled as a directed graph G = (V, E), where the node set V corresponds to the switch and the edge set E represents the network link. From the node v i To node v j The edge is denoted by e ij , and its available bandwidth is b ij Indicates. ij The value is not necessarily equal to the reverse bandwidth (b ji ). In addition, the source node and the target node of the request are also added to the graph G to expand the graph G. These nodes are connected to their corresponding switches through virtual links. The switches in the graph G include aggregation switches that undertake data aggregation functions and ordinary switches that do not undertake data aggregation functions.

[0045] A switch attribute PAT (Peak Aggregation Throughput) is defined to describe the maximum aggregation capacity that the switch can support. According to the memory overhead required by the aggregation switch to achieve the expected transmission throughput, the value of PAT is linearly related to the number of aggregators configured on the aggregation switch. Considering that GAIN requires additional memory for error handling, the PAT of an aggregation switch configured with N aggregators can be expressed by the following formula:

[0046]

[0047] Where d0 represents the packet size, and R represents the round-trip time (RTT) of the packet in the network. When multiple tasks need the same aggregation switch to aggregate their data flows, if the PAT of the aggregation switch is greater than or equal to their total rate, the aggregation switch can aggregate all flows, and each task will have only one flow forwarded through its respective egress link. However, if the PAT of the aggregation switch is less than the sum of the rates of these tasks, it indicates that the aggregation switch cannot handle all the data flows from these tasks, and some tasks need to consider other switches for data aggregation.

[0048] In addition to ensuring the connectivity between the source node and the destination node, GAIN routing also needs to consider three constraints. First, if the aggregation switch v i The remaining polymerization capacity p i (equal to its PAT minus the total rate of other tasks that require the switch to aggregate) should be greater than or equal to the rate required by the task, that is, p i≥r. Second, the routing should satisfy the bandwidth constraints of each link; specifically, when the switch cannot perform data aggregation as a branch node, its egress link can carry multiple data flows for the task, that is, for the link e carrying the m flows of the task, ij , its bandwidth should satisfy condition b ij ≥m×r. Third, to avoid the central server collecting data from a single client, GAIN should ensure that the data of each client is aggregated at least once. This ensures that the central server can only obtain the aggregated results but cannot obtain the user's private data.

[0049] The goal of the above-mentioned task routing algorithm for multiple GAIN tasks is to minimize the bandwidth overhead on the link and the resource overhead on the switch (i.e., the total number of occupied aggregators). However, even in the simplest case (i.e., not considering the above three constraints and only considering minimizing the bandwidth overhead), it belongs to a class of Steiner tree problems and has been proven to be NP-hard (non-deterministic polynomial hard) problems. To meet this challenge, a heuristic algorithm, also known as the Three-Phase Validation (TPV) algorithm (i.e., the heuristic algorithm for three-phase validation), is further designed to obtain a feasible solution while reducing the computational overhead of online route generation. The TPV algorithm consists of three stages: verifying the initial route, verifying the aggregation node, and verifying the number of aggregations. It continuously checks and adjusts the current path so that it eventually meets the GAIN constraints.

[0050] Verify the initial route: In this phase, each source node s in the TPV computation graph G i ∈S to the target node dst. These paths together form a tree subgraph, called the shortest path tree (denoted as graph T). In this tree subgraph, if a node is the endpoint of multiple edges, it is called a branch node, and the branch node can be used as an aggregation node.

[0051] TPV first assumes that these branch nodes have sufficient residual aggregation capacity. This means that the egress link of each branch node transmits only one aggregate flow, whose size is equal to the size of the flow carried by each ingress link. To this end, TPV verifies whether all links in the graph T can meet the requested rate requirements. If the available bandwidth of any link is less than the required rate r, it means that even if all branch nodes are used for data aggregation, the transmission rate cannot meet the requirements. Therefore, TPV believes that the request cannot be met. It should be noted that although recalculating a new route in the graph G to meet the rate requirement may produce a feasible solution, it may cause "detours" and increase the length of the route. When processing a large number of requests, "detours" for each request to meet its rate requirements may bring huge bandwidth overhead to the network, which needs to be avoided, so this practice is ignored here.

[0052] Verify aggregation nodes: For routes that meet the constraints of the first stage, TPV further verifies whether all branch nodes have sufficient remaining aggregation capacity. To this end, TPV first finds all branch nodes in the graph T and forms a set A, then selects nodes whose remaining aggregation capacity is less than r, recorded as set A′.

[0053] For each node in set A', the flows of its child nodes cannot be aggregated due to insufficient aggregation capabilities and can only be forwarded to the downstream branch nodes for aggregation. This makes the number of flows on the egress link equal to the number of flows entering the node. Therefore, the routing in graph T should be adjusted to set this branch node as a forwarding node. In addition, TPV will verify whether set A and set A' are the same. If A = A', it means that all nodes on the path cannot aggregate data. Therefore, TPV believes that the request cannot be satisfied.

[0054] use Figure 2 The example in is used to illustrate this adjustment process, where s1 to s5 are source nodes, dst is the destination node, v1 to v5 are switching nodes in the network, and the set A = {v1, v3, v4}. If node v1 has sufficient aggregation capacity, it will aggregate the flows from source nodes s1, s2, and s3, and then output only a single aggregated flow. Therefore, links {v1, v2} and {v4, v5} only need to transmit one flow, which requires that their available bandwidth is greater than or equal to r. However, if node v1 lacks sufficient aggregation capacity, it can only forward flows from source nodes s1, s2, and s3. These flows are expected to be aggregated at the downstream branch node v4. In this scenario, TPV specifies node v1 as a forwarding node. Therefore, links {v1, v2} and {v4, v5} must carry three unaggregated flows. This route is valid only when the available bandwidth of the two links is greater than or equal to 3r. It should be noted that if such an adjustment cannot meet the bandwidth constraints of these links, TPV considers that the request cannot be satisfied.

[0055] Verify the number of aggregations: The routes generated after the above two stages are already able to meet the resource constraints of the switch and the bandwidth constraints of the link. However, in order to ensure the privacy of the data, it is necessary to ensure that the data of each source node is aggregated at least once. To this end, TPV checks whether there is a branch node on the path from each source node to the target node, which performs aggregation instead of forwarding. If there is no aggregation, the data of the source node can be forwarded directly to the target node without aggregation, allowing the target node to obtain its original data.

[0056] It should be noted that TPV only verifies the remaining part of the available bandwidth at each step. If a link is used in the previous step, its updated available bandwidth is verified in the subsequent step to meet the bandwidth requirements of the transmission.

[0057] The runtime multi-task management method of the extended security aggregation acceleration service extends GAIN to support runtime multi-task management. It precompiles an aggregator table on each programmable switch and dynamically allocates aggregators on the aggregator table between different federated learning tasks to expand the GAIN task to multi-task applications. Different tasks can use the aggregator on the same switch without interfering with each other. During the multi-task management process, the routing of the federated learning task is calculated through an efficient task routing algorithm, which avoids the central server collecting data from a single client and ensures that the data of each client is aggregated at least once, thereby protecting the user's data privacy and achieving the effect of enabling GAIN to have the ability of runtime multi-task management.

[0058] In one embodiment, when the three-stage verification heuristic algorithm is used to perform routing calculation of the aggregation switch and a route that satisfies the constraint condition cannot be determined, the runtime multi-task management method for the extended security aggregation acceleration service may further include the following processing steps:

[0059] Use the three-stage verification heuristic algorithm to update all the branch point sets in the shortest path tree;

[0060] For each source node that needs to recalculate the route, a three-stage verification heuristic algorithm is used to calculate the shortest path from the source node to each node in the bifurcation point set, and the shortest path is selected;

[0061] A three-stage verification heuristic algorithm is used to verify whether all links on the shortest path have available bandwidth greater than or equal to the rate required by the multi-GAIN task request;

[0062] If not, it is determined that the request is unsuccessful, otherwise the shortest path is determined to be the recalculated route.

[0063] Specifically, for source nodes that do not meet the above constraints, TPV needs to recalculate their routes. Specifically, TPV updates the set A to represent the branch nodes responsible for performing the aggregation operation. Then, for each source node s that needs to recalculate the route i , TPV is calculated from s i The shortest path to each node in set A is selected, and the shortest of these paths is selected. Then TPV verifies whether all links on this path have available bandwidth greater than or equal to the requested rate r. If there is a path that does not meet this requirement, the request is considered unsuccessful. Otherwise, since the endpoints of each path are aggregation nodes, this ensures that the data of each source node can be aggregated at least once after recalculating the route, thereby achieving privacy protection.

[0064] In one embodiment, the runtime multi-task management method of the extended security aggregation acceleration service may further include the following processing steps:

[0065] If it is determined that switch memory fragmentation occurs, synchronize the windows of all source nodes;

[0066] Clear the regions of all multi-GAIN tasks running on the aggregation switch, and re-assign the multi-GAIN tasks to the new aggregator region of the tasks in the aggregator table in the original order.

[0067] After the current task is completed, memory fragmentation is eliminated by dynamically adjusting the task area on the aggregator table, and the aggregator area is reallocated to the new task to avoid the failure of new task aggregation due to lack of sufficient continuous areas, so as to effectively avoid switch memory fragmentation. Finally, GAIN can achieve the effect of having the ability of runtime multi-task management.

[0068] It is understandable that when a task is completed, the occupied area in the aggregator table will be recycled and the related aggregators will be initialized. Subsequently, the new task can reuse the aggregators in the area to aggregate the data packets. However, due to the different rate requirements of different tasks, not all aggregators in the recycled area are used by new tasks, which leads to a large number of fragmented areas in the aggregator table after tasks dynamically join and exit. Therefore, even if the TPV algorithm believes that the switch has enough remaining aggregation capacity, the task may still fail due to the lack of sufficiently continuous areas. This problem is called switch memory fragmentation.

[0069] Eliminate fragmentation by dynamically adjusting the task regions on the aggregator table. It should be noted that TPV estimates the remaining aggregation capacity of the aggregation switch based on the number of remaining aggregators. If an aggregation switch has sufficient aggregation capacity but lacks continuous aggregators, this indicates that there is a serious memory fragmentation problem on the aggregation switch. Therefore, it is necessary to clear the regions of all GAIN tasks running on the aggregation switch and reassign these tasks to their new aggregator regions on the aggregator table in their original order. This method effectively eliminates memory fragmentation while ensuring that there is a continuous region at the end of the aggregator table.

[0070] However, the adjustment process may introduce potential errors in the aggregation process. Figure 3 In the example shown, all aggregators on the aggregation switch are initialized, and then the regions of the running tasks are adjusted from their previous locations to their new locations. Figure 3(a)), adjusting the region does not affect correctness. When the region changes, packets sent before the adjustment are considered lost and trigger retransmissions. These retransmitted packets are then aggregated in the new region. After this stage, there is no change in the aggregation process. However, if window desynchronization occurs (such as Figure 3 (b) shows the aggregation interruption), the aggregation process will be interrupted. Window resynchronization relies on the aggregation results previously cached on the switch, but the adjustment will cause these results to be lost. Therefore, the message with sequence number SN=1 from client 2 will match an empty aggregator, which does not contain the previous aggregation results. Therefore, the switch must wait for the packet with sequence number SN=1 from client 1 to complete the aggregation, but this will never happen. Among them, Figure 3 Part (1) of (a) is when the packet is retransmitted due to timeout. Figure 3 Part (2) of (a) is to clear the original aggregator and aggregate data packets in the new area. Figure 3 Part (1) of (b) is the window asynchrony before adjustment. Figure 3 Part (2) of (b) is for the packet with SN=1 from client 2 to wait for the packet with SN=1 from client 1, but this does not happen.

[0071] To avoid interruption, the windows of all source nodes must be synchronized before adjustment. When a regional adjustment is required, the network provider should notify the target node (i.e., the central server) of each task involved. Each target node records the sequence number SN of the currently received data packet and notifies all source nodes to synchronize the starting position of its window to SN+1. Figure 3 For example, if the central server receives the data packet with sequence number SN=1 but has not received the data packet with sequence number SN=2 before the adjustment, the window start positions of the two clients will be synchronized to the position of the data packet with sequence number SN=2. Subsequently, the client sends data packets according to this synchronized window configuration. These packets are then aggregated in the new area.

[0072] In one embodiment, in order to more intuitively illustrate the performance and effect of the runtime multi-task management method of the above-mentioned extended security aggregation acceleration service, one of the examples is given here to evaluate the performance of GAIN supporting runtime multi-task management through FPGA-based experiments and large-scale simulations. Those skilled in the art will understand that this experimental example is only illustrative and is not the only limitation implemented on the runtime multi-task management method of the above-mentioned extended security aggregation acceleration service.

[0073] The existing DEFO dataset is used to simulate and evaluate the performance of GAIN for a large number of clients in a real ISP (Internet Service Provider) network. The default experimental network is selected as the largest topology in the dataset (rf1239, 315 nodes, 1944 edges), and the default number of clients is set to 1000. In each network, half of the nodes are randomly assigned as ingress switches and their PAT is assumed to be equal to the link bandwidth. By default, the remaining nodes are considered to be regular switches. The link bandwidth is configured to 10Gbps, and the rate required for each request is randomly specified between 10Mbps and 100Mbps.

[0074] like Figure 4 As shown in the figure, it is a schematic diagram of bandwidth overhead under different topologies and different proportions of programmable switches in the network. Figure 4 (a) shows the normalized bandwidth overhead in different topologies, where GAIN-Ingress means that only the ingress switch is used for data aggregation, while GAIN-All assumes that all switches are able to aggregate data. The data transmitted by all methods is of the same size, and the bandwidth overhead refers to the total amount of data transmitted on all links. Conv is a traditional model update method, and Conv-Quant is a traditional method that quantizes the update data to float16 type to reduce the amount of communication data. Compared with Conv, GAIN-Ingress significantly reduces the bandwidth overhead, with an average reduction of more than 96.3% on all topologies, and the maximum reduction reaches 99.0% (rf3967). GAIN-All achieves even lower bandwidth overhead, reaching 99.2% (rf1221, rf1755, and rf3967). Figure 4 (b) shows the normalized bandwidth overhead when replacing different proportions of traditional switches with programmable switches. Increasing the proportion of programmable switches allows more traffic to be aggregated, thereby reducing the bandwidth overhead accordingly. For 1000 clients (yield 1000 clients), increasing the proportion of programmable switches from 0% to 100% can reduce the bandwidth overhead by an additional 31.5%.

[0075] like Figure 5 The figure shows the effect of different numbers of clients on bandwidth overhead and request completion rate. The effect of the number of clients on bandwidth overhead is shown in Figure 5(a). Since GAIN aggregates client data using ingress switches, it effectively reduces the traffic entering the edge network. As the number of clients increases, the bandwidth overhead of both GAIN-Ingress and GAIN-All increases only slightly. In contrast, Conv and Conv-Quant fail to avoid increased congestion. When the number of clients is 200, the bandwidth overhead of Conv is 3.67 times that of GAIN-Ingress; when the number of clients reaches 1200, the gap further widens to 4.27 times.

[0076] The request completion rate is Figure 5 As shown in (b), 100 requests are generated sequentially and then they are tried to be configured into the network. If the rate requirement of the request cannot be met, the request is abandoned and marked as failed. As the number of clients increases, the completion rate of Conv decreases as the number of source nodes for each task increases. When the number of clients reaches 1200, Conv only achieves a completion rate of about 2.8%. In contrast, GAIN-Ingress and GAIN-All are least affected, with completion rates of 78.3% and 98.4%, respectively, which are 21.7 times and 32.4 times higher than before.

[0077] It should be understood that although Figure 1 The steps in the method are shown in sequence as indicated by the arrows, but the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of the steps, and the steps can be executed in other orders. Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequentially, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0078] In one embodiment, Figure 6As shown, a runtime multi-task management system 100 for extending a secure aggregation acceleration service is also provided, which may include a message receiving module 11, a rule matching module 13, an aggregation matching module 15, a routing calculation module 17, and an aggregation processing module 19. Among them, the message receiving module 11 is used to receive messages of multiple GAIN tasks. The rule matching module 13 is used to search the matching rule table inside the aggregation interactive machine according to the message content of the message, return the task identity of the message and retrieve the offset of the matched aggregator table; the aggregation interactive machine is precompiled with an aggregator table. The aggregation matching module 15 is used to match the message to the aggregator in the aggregator table according to the offset. The routing calculation module 17 is used to perform routing calculation of the aggregation switch using a three-stage verification heuristic algorithm to determine the route that meets the constraint conditions; the three-stage verification heuristic algorithm includes verifying the initial route, verifying the aggregation node, and verifying the number of aggregations. The constraint conditions include that the remaining aggregation capacity is greater than or equal to the rate required by the multiple GAIN tasks when the aggregation switch is used for aggregation, the route meets the bandwidth constraint of each link, and ensures that the data of each client is aggregated at least once. The aggregation processing module 19 is used to complete the aggregation processing of the messages of multiple GAIN tasks according to the routes that meet the constraint conditions.

[0079] The runtime multi-task management system 100 of the extended security aggregation acceleration service extends GAIN to support runtime multi-task management. It precompiles an aggregator table on each programmable switch and dynamically allocates aggregators on the aggregator table between different federated learning tasks to expand the GAIN task to multi-task applications. Different tasks can use the aggregator on the same switch without interfering with each other. During the multi-task management process, the routing of the federated learning task is calculated through an efficient task routing algorithm, which avoids the central server collecting data from a single client and ensures that the data of each client is aggregated at least once, thereby protecting the user's data privacy and achieving the effect of enabling GAIN to have the ability of runtime multi-task management.

[0080] In one embodiment, the runtime multi-task management system 100 of the extended security aggregation acceleration service may further include a routing recalculation module for updating all bifurcation point sets in the shortest path tree using a three-stage verified heuristic algorithm; for each source node that needs to recalculate the route, the three-stage verified heuristic algorithm is used to calculate the shortest path from the source node to each node in the bifurcation point set, and select the shortest path; the three-stage verified heuristic algorithm is used to verify whether all links on the shortest path have an available bandwidth greater than or equal to the rate required by the multi-GAIN task request; if not, it is determined that the request is unsuccessful, otherwise the shortest path is determined to be the recalculated route.

[0081] In one embodiment, the runtime multi-task management system 100 of the extended security aggregation acceleration service may also include a memory fragmentation elimination module, which is used to synchronize the windows of all source nodes when it is determined that switch memory fragmentation occurs, clear the areas of all multi-GAIN tasks running on the aggregation switch, and re-assign the multi-GAIN tasks to a new aggregator area of ​​the tasks in the aggregator table in the original order.

[0082] For the specific limitations of the runtime multi-task management system 100 for the extended security aggregation acceleration service, please refer to the corresponding limitations of the runtime multi-task management method for the extended security aggregation acceleration service above, which will not be repeated here. Each module in the runtime multi-task management system 100 for the extended security aggregation acceleration service can be implemented in whole or in part through software, hardware and a combination thereof. The above modules can be embedded in or independent of a device with data processing functions in the form of hardware, or can be stored in the memory of the aforementioned device in the form of software, so that the processor can call and execute the operations corresponding to the above modules. The aforementioned device can be, but is not limited to, various types of edge communication devices already available in the art.

[0083] In one embodiment, a computer device is also provided, including a memory and a processor, the memory stores a computer program, and the processor implements the following processing steps when executing the computer program: receiving a message of a multi-GAIN task; searching a matching rule table inside an aggregation interactive machine according to the message content of the message, returning the task identity of the message and retrieving the offset of the matched aggregator table; the aggregation interactive machine is precompiled with an aggregator table; matching the message to an aggregator in the aggregator table according to the offset; performing routing calculation of the aggregation switch using a three-stage verification heuristic algorithm to determine a route that meets constraint conditions; the three-stage verification heuristic algorithm includes verifying an initial route, verifying an aggregation node, and verifying the number of aggregations, and the constraint conditions include that the remaining aggregation capacity is greater than or equal to the rate required by the multi-GAIN task when the aggregation switch is used for aggregation, the route meets the bandwidth constraint of each link, and ensures that the data of each client is aggregated at least once; and completing the aggregation processing of the message of the multi-GAIN task according to the route that meets the constraint conditions.

[0084] It can be understood that in addition to the memory and processor mentioned above, the above-mentioned computer device may also include other software and hardware components not listed in this specification. The specific details can be determined according to the model of the specific edge communication device in different application scenarios, and this specification will no longer list them one by one in detail.

[0085] In one embodiment, when the processor executes the computer program, it can also implement the steps or sub-steps added in each embodiment of the runtime multi-task management method for extending the security aggregation acceleration service.

[0086] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, the following processing steps are implemented: receiving a message of a multi-GAIN task; searching a matching rule table inside an aggregation interactive machine according to the message content of the message, returning the task identity of the message and retrieving the offset of the matched aggregator table; the aggregation interactive machine is precompiled with an aggregator table; matching the message to an aggregator in the aggregator table according to the offset; performing routing calculation of the aggregation switch using a three-stage verification heuristic algorithm to determine a route that meets the constraint conditions; the three-stage verification heuristic algorithm includes verifying the initial route, verifying the aggregation node, and verifying the number of aggregations, and the constraint conditions include that the remaining aggregation capacity is greater than or equal to the rate required by the multi-GAIN task when the aggregation switch is used for aggregation, the route meets the bandwidth constraint of each link, and ensures that the data of each client is aggregated at least once; completing the aggregation processing of the message of the multi-GAIN task according to the route that meets the constraint conditions.

[0087] In one embodiment, when the computer program is executed by a processor, it can also implement the steps or sub-steps added in each embodiment of the runtime multi-task management method for extending the security aggregation acceleration service described above.

[0088] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus dynamic random access memory (RambusDRAM, referred to as RDRAM) and interface dynamic random access memory (DRDRAM).

[0089] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of protection of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and all of them belong to the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the attached claims.

Claims

1. A runtime multi-task management method for extending security aggregation acceleration services, It is characterized in that Includes steps: Receive messages from multiple GAIN tasks; Searching the matching rule table inside the aggregation interaction machine according to the message content of the message, returning the task identity of the message and retrieving the offset of the matched aggregator table; The aggregation interaction machine is precompiled with an aggregator table; Matching the message to an aggregator in the aggregator table according to the offset; The heuristic algorithm of the three-stage verification is used to calculate the routing of the aggregation switch to determine the routing that meets the constraint conditions; the heuristic algorithm of the three-stage verification includes verifying the initial routing, verifying the aggregation node and verifying the number of aggregations, and the constraint conditions include that the remaining aggregation capacity when the aggregation switch is used for aggregation is greater than or equal to the rate required by multiple GAIN tasks, the routing meets the bandwidth constraint of each link, and ensures that the data of each client is aggregated at least once; The messages of multiple GAIN tasks are aggregated according to the routes that meet the constraints.

2. The runtime multi-task management method for extended security aggregation acceleration service according to claim 1, It is characterized in that When the route calculation of the aggregation switch is performed using the three-stage verification heuristic algorithm and a route that satisfies the constraint condition cannot be determined, the runtime multi-task management method further includes the steps of: Use the three-stage verification heuristic algorithm to update all the branch point sets in the shortest path tree; For each source node that needs to recalculate the route, a three-stage verification heuristic algorithm is used to calculate the shortest path from the source node to each node in the bifurcation point set, and the shortest path is selected; A three-stage verification heuristic algorithm is used to verify whether all links on the shortest path have available bandwidth greater than or equal to the rate required by the multi-GAIN task request; If not, it is determined that the request is unsuccessful, otherwise the shortest path is determined to be the recalculated route.

3. The runtime multi-task management method for extended security aggregation acceleration service according to claim 1 or 2, It is characterized in that Also includes the steps: If it is determined that switch memory fragmentation occurs, synchronize the windows of all source nodes; Clear the regions of all multi-GAIN tasks running on the aggregation switch, and re-assign the multi-GAIN tasks to the new aggregator region of the tasks in the aggregator table in the original order.

4. A runtime multi-task management system for extended security aggregation acceleration services, It is characterized in that include: Message receiving module, used to receive messages of multiple GAIN tasks; A rule matching module, used for searching a matching rule table inside the aggregation interaction machine according to the message content of the message, returning the task identity of the message and retrieving the offset of the matched aggregator table; the aggregation interaction machine is precompiled with the aggregator table; An aggregation matching module, used for matching the message to an aggregator in the aggregator table according to the offset; A routing calculation module, used to perform routing calculation of the aggregation switch using a three-stage verification heuristic algorithm to determine a route that meets the constraint conditions; the three-stage verification heuristic algorithm includes verifying the initial route, verifying the aggregation node, and verifying the number of aggregations, and the constraint conditions include that the remaining aggregation capacity when the aggregation switch is used for aggregation is greater than or equal to the rate required by multiple GAIN tasks, the route meets the bandwidth constraint of each link, and ensures that the data of each client is aggregated at least once; The aggregation processing module is used to complete the aggregation processing of the messages of multiple GAIN tasks according to the routes that meet the constraint conditions.

5. The runtime multi-task management system for the extended security aggregation acceleration service according to claim 4, It is characterized in that The method also includes a route recalculation module, which is used to update all the bifurcation point sets in the shortest path tree using a three-stage verified heuristic algorithm; for each source node that needs to recalculate the route, the three-stage verified heuristic algorithm is used to respectively calculate the shortest path from the source node to each node in the bifurcation point set, and select the shortest path; the three-stage verified heuristic algorithm is used to verify whether all links on the shortest path have an available bandwidth greater than or equal to the rate required by the multi-GAIN task request; if not, it is determined that the request is unsuccessful, otherwise the shortest path is determined to be the recalculated route.

6. The runtime multi-task management system of the extended security aggregation acceleration service according to claim 4 or 5, It is characterized in that It also includes a memory fragmentation elimination module, which is used to synchronize the windows of all source nodes when it is determined that switch memory fragmentation occurs, clear the areas of all multi-GAIN tasks running on the aggregation switch, and re-assign the multi-GAIN tasks to a new aggregator area of ​​the tasks on the aggregator table in the original order.