Space-based computing power constellation-oriented multi-hop routing and load distribution method and system, and storage medium

Abstract

The invention provides a space-based computing power constellation-oriented multi-hop routing and load distribution method and system and a storage medium, and the method comprises the steps: 1, building a model for the space-based computing power constellation to execute a matrix vector multiplication calculation task, giving a source node to encode and split the task, and giving a relay node to transmit the task hop by hop, the relay node and the destination node carry out distributed calculation on the task in the end-to-end routing process; 2, establishing an optimization problem of minimizing task execution time delay, and proposing a computing power perception and load distribution algorithm based on a block coordinate descent method to split an original problem into routing and load distribution problems; and step 3, selecting an optimal relay node based on a routing algorithm of dynamic programming design computing power perception, and designing a multi-hop coding load distribution algorithm based on an interior point method to realize communication and computing load balancing. The method has the advantages that a large amount of original data is prevented from being transmitted to a ground center, tasks are processed in a distributed mode in the end-to-end routing process, and task execution time delay is reduced.

Description

Technical Field

[0001] The present invention relates to the field of satellite communication technology, and in particular to a multi-hop routing and load distribution method, system and storage medium for a space-based computing constellation. Background Art

[0002] Space-based computing constellations include a large number of low-orbit constellations. They have attracted widespread attention from the industry and academia because of their high bandwidth, ubiquitous coverage, ubiquitous connectivity, and ubiquitous computing. For example, my country's GW constellation plans to deploy 12,992 satellites at orbital altitudes of 590km, 600km, and 1,145km. With the development of integrated circuits, satellites have a high on-orbit processing speed. For example, the Xiphos Q8S processor launched by Xilinx can achieve a computing speed of 1,800 GFLOPS. However, with the application of deep learning and convolutional neural networks in image processing, the computing-intensive business of space-based computing networks is becoming increasingly large. In traditional satellite networks, satellites only serve as transmission relays to forward a large amount of raw data to ground computing centers for processing. This generates a satellite-to-ground communication delay of up to 16s when the mission arrival rate is high, making it difficult to support delay-sensitive tasks. At the same time, space-based computing constellations have problems such as insufficient computing power of a single satellite, the straggler problem caused by large-scale satellite computing power heterogeneity, and extended multi-hop communication time. These problems restrict the execution efficiency of space-based computing constellations in processing computationally intensive and delay-sensitive tasks, resulting in the inability to directly apply existing routing strategies and load distribution strategies to space-based computing constellations.

[0003] At present, many routing strategies for satellite constellations ignore the impact of computing delay on task execution delay, and the load distribution strategy lacks methods to resist stragglers in distributed computing of heterogeneous space-based computing constellations.

[0004] At present, there are some inadequate considerations in the routing and load distribution methods for space-based computing constellations to perform computationally intensive and delay-sensitive tasks. In terms of routing strategies, existing methods only focus on reducing communication delays. They can only find the shortest path with communication delay from the source node to the node used for computing. This not only ignores the impact of computing delays on task execution efficiency, but also fails to fully utilize the idle computing resources of relay satellites, which easily leads to the bottleneck problem of limited computing resources of a single satellite. In terms of load distribution strategies, distributed computing can break through the bottleneck of computing resources of a single satellite through the cooperation of multiple satellites, but it is easy to have a straggler node whose computing time is significantly slower than that of most nodes. The introduction of computing redundancy through coding can reduce the impact of stragglers, but it is still necessary to design the balance of coding load according to the heterogeneous computing and communication capabilities of satellites, so as to reduce the task execution delay by introducing redundancy. Summary of the invention

[0005] In order to solve the problems in the prior art, the present invention provides a multi-hop routing and load distribution method for a space-based computing constellation, comprising the following steps:

[0006] Step 1: Establish a model for the space-based computing constellation to perform matrix-vector multiplication calculation tasks, provide source nodes to encode and split the tasks, relay nodes to transmit the tasks hop by hop, and relay nodes and destination nodes to perform distributed calculations on the tasks during the end-to-end routing process;

[0007] Step 2: Establish an optimization problem of minimizing task execution delay, and propose a computing power perception and load distribution algorithm based on block coordinate descent method to split the original problem into routing and load distribution problems;

[0008] Step 3: Design a computing power-aware routing algorithm based on dynamic programming to select the best relay node, and design a multi-hop coding load distribution algorithm based on the interior point method to achieve communication and computing load balancing.

[0009] As a further improvement of the present invention, the step 1 comprises:

[0010] Step S1: The source node S first selects the optimal path p i ,use The MDS code encodes the task matrix A as Then it is divided into H sub-matrices, where the allocation path p i The jth u The submatrix of the jump work node u is a matrix composed of the encoding combinations of each row in the task matrix A. is a positive integer, Random linear combinations of the L rows from the task matrix A;

[0011] Step S2: The master node encodes the matrix It is sent to the destination node D hop by hop through H-1 relay nodes. Each relay node receives the corresponding sub-matrix assigned to itself. For the sub-matrix not assigned to itself, the relay node does not store it but directly sends it to the next hop node, i.e., a transparent transmission mechanism is adopted.

[0012] Step S3: The destination node and each relay node complete the calculation of the sub-matrix assigned to them, until the destination node and each relay node complete at least L results in total.

[0013] As a further improvement of the present invention, in step S3, the communication delay of each node receiving the allocated sub-matrix is ​​composed of the transmission delay and the propagation delay of the multi-hop ISL;

[0014] is the sum of the computation delay, propagation delay and transmission delay of node u, and the formula is as follows:

[0015]

[0016] in, Indicates the selection path p i Time u The total delay of the satellite node u to complete the task, Indicates the selection path p i Time u The computation time of the satellite hopping node u, Indicates the selection path p i When the task is transferred to the jth u The propagation delay of the satellite hopping node u, Indicates the selection path p i When the task is transferred to the jth u Transmission time of hopping satellite node u;

[0017] The cumulative probability distribution function of the total delay of satellite node u to complete the task is as follows:

[0018]

[0019] Among them, the positive number μ u is the stragling parameter determined by the computing power of satellite u, a positive number u is the shift time parameter required for satellite u to complete a single inner product operation, They represent the number of intra-orbital inter-satellite link hops and the number of inter-orbital inter-satellite link hops required for transmission from source node S to node u, respectively. iv represents the assignment to path p i The number of rows of the encoding matrix of the v-th hop node on Represents the path p i The jth u Hop transmission rate, considering the mesh satellite network topology, d v represents the distance between satellites in the same orbit, d h represents the distance between the intersatellite links of adjacent satellites in different orbits, and b represents the coding matrix The number of bits in a row, c is the speed of light;

[0020] T exe The CDF of The product of is expressed as follows:

[0021]

[0022] T exe Represents the total delay of the system to complete the task, Indicates that as of T exe The node set that completes the submatrix calculation when

[0023] Due to Texe is a positive number, and its expectation is obtained by integrating the CDF, as follows:

[0024]

[0025] As a further improvement of the present invention, in the step 2, it also includes:

[0026] Step y1: By finding the optimal path p i and optimal load distribution Minimize the average total delay The problem is set up as follows:

[0027]

[0028] in is the set of natural numbers;

[0029] Step y2: Replace the objective function with a desired constraint, as follows:

[0030]

[0031] Among them, t exe* Indicates the expected deadline.

[0032] Step y3: Based on the block coordinate descent method, the total problem of step y2 is The problem is broken down into two sub-problems. For the optimal path problem, given the total delay of the initial feasible task execution and load distribution strategy i(0) , find the path p that can complete the calculation with the largest number of inner products i ,question The load distribution problem is a constrained optimization problem. Given a path p i Finally, find the minimum task execution delay t that can ensure the completion of the task exe* and specific load distribution strategies

[0033]

[0034] By continuously executing the above steps, the algorithm will gradually converge and obtain the routing and load distribution solution that minimizes the task execution delay.

[0035] As a further improvement of the present invention, in step 3, the computing power-aware routing algorithm based on dynamic programming design runs the following steps:

[0036] Step 1: Input source node S, destination node D, Total inner product of tasks L, dv ,d h ,R v ,R h , the total delay of the result of the (r-1)th iteration Load distribution vector l i(r-1) ;

[0037] Step 2: Initial State

[0038] Step 3: Calculate the state according to equation (19) And according to formula (20), record the optimal previous hop node for node u

[0039] Step 4: Output the optimal path

[0040]

[0041] As a further improvement of the present invention, in step 3, the multi-hop coding load distribution algorithm designed based on the interior point method runs the following steps:

[0042] Step a1: read the route obtained by the routing algorithm as a path, and read the load distribution and task execution delay of the previous iteration as an initial feasible solution;

[0043] Step a2: Initialize the number of iterations and enter the loop;

[0044] Step a3: Construct barrier function

[0045]

[0046] Step a4: X (k-1) ={t exe*(k-1) ,l i (k-1)} is used as the initial point to solve the unconstrained minimum problem of minimizing the barrier function. The formula is as follows:

[0047]

[0048] where r k is the barrier factor;

[0049] Step a5: Determine whether the solution converges. If so, exit the loop. Otherwise, reduce the barrier factor r. k The value of , the number of iterations increases by 1, and the process returns to step a3;

[0050] Step a6: Output optimal delay and optimal load distribution.

[0051] As a further improvement of the present invention, in step 1, a distributed matrix multiplication operation is performed on the space-based computing power constellation. Assuming that H hops are required for transmission from the source node S to the destination node D, the computing task y=Ax generated by the source node is distributedly calculated through the destination node D and H-1 relay nodes, where

[0052] The present invention also discloses a multi-hop routing and load distribution system for space-based computing constellations, including:

[0053] Module 1: It is used to establish a model for the space-based computing constellation to perform matrix-vector multiplication calculation tasks, provide source nodes to encode and split tasks, relay nodes to transmit tasks hop by hop, and relay nodes and destination nodes to perform distributed calculations on tasks during end-to-end routing;

[0054] Module 2: Used to establish the optimization problem of minimizing the task execution delay, a computing power perception and load distribution algorithm based on block coordinate descent method is proposed to split the original problem into routing and load distribution problems;

[0055] Module 3: Used to design a computing power-aware routing algorithm based on dynamic programming to select the best relay node, and design a multi-hop coding load distribution algorithm based on the interior point method to achieve communication and computing load balancing.

[0056] As a further improvement of the present invention, the multi-hop routing and load distribution system runs the following steps:

[0057] Step 1: Input the constellation parameters of space-based computing power, including inter-satellite distance, inter-satellite transmission rate, displacement time parameters related to satellite computing power, straggler parameters, task source node, destination node, and number of task rows;

[0058] Step 2: Initialize any feasible routing and load distribution;

[0059] Step 3: Call the computing power-aware routing algorithm based on dynamic programming, fix the load distribution scheme, and search for computing power-aware routing by maximizing the number of coded inner product calculations expected to be completed on the transmission path;

[0060] Step 4: Call the coding load distribution algorithm based on the interior point method to solve the computation and communication load of each hop on the path according to the route obtained in step 3;

[0061] Step 5: Calculate the task execution delay under the routing in step 3 and the load distribution in step 4;

[0062] Step 6: Determine whether the task execution delay is reduced compared with the previous iteration result. If it is reduced, use the current iteration result as the starting point for the next iteration and return to step 3. Otherwise, jump to step 7.

[0063] Step 7: Output computing power routing and computing communication load distribution, end.

[0064] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the multi-hop routing and load distribution method of the present invention when called by a processor.

[0065] The beneficial effects of the present invention are: 1. It is proposed to embed coded distributed computing into the space-based computing power constellation, which can avoid transmitting a large amount of raw data to the ground center, and at the same time, distribute the processing of tasks in the end-to-end routing process, rather than processing them on a single satellite, thereby significantly reducing the task execution delay; 2. In order to select satellites with higher computing power to participate in task calculations, the computing power-aware multi-hop routing scheme simultaneously evaluates the communication capability and computing capability of the relay satellite, integrating the communication and computing processes; 3. The designed load distribution scheme can resist the influence of straggler by introducing coding technology, thereby improving the robustness of the space-based computing power constellation, and achieving a balance between computing and communication loads through on-orbit processing of tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a schematic diagram of the topology of the space-based computing power constellation of the present invention;

[0067] Figure 2 This is a schematic diagram of the topology of the space-based computing power constellation of the present invention;

[0068] Figure 3 This is a workflow diagram of the space-based computing power constellation coding distributed computing system of the present invention;

[0069] Figure 4 It is a time sequence diagram of task execution delay of the present invention;

[0070] Figure 5 Schematic diagram of average task execution delay comparison of the present invention; (a) H = 14, scenario 1, (b) L = 2000, scenario 1, (c) H = 14, scenario 2, (d) L = 2000, scenario 2; DETAILED DESCRIPTION

[0071] According to the characteristics of space-based computing power constellations, the present invention analyzes the time delay of space-based computing power constellations in executing computing tasks, designs a distributed computing method to solve the problem of insufficient computing power of a single satellite, designs a coding load distribution method to solve the straggler problem, and designs computing power-aware routing to solve the problem of extended multi-hop communication time, laying a technical foundation for the planning of space-based computing power constellations.

[0072] The core goal of the method designed by the present invention is to embed coded distributed computing into the space-based computing constellation to avoid transmitting a large amount of raw data to the ground center, and at the same time encode and split the computing tasks, and distribute them in the end-to-end routing process instead of processing them on a single satellite, thereby improving the execution efficiency of the space-based computing constellation in processing computing-intensive and delay-sensitive services. This method can find the most suitable satellite node for relay transmission and relay computing from the source node to the destination node in a given satellite network, and design specific communication and computing loads for each node to reduce the task execution delay of the space-based computing constellation.

[0073] The present invention discloses a multi-hop routing and load distribution method for a space-based computing constellation, comprising the following steps:

[0074] Step 1: Establish a model for the space-based computing constellation to perform matrix-vector multiplication calculation tasks, provide source nodes to encode and split the tasks, relay nodes to transmit the tasks hop by hop, and relay nodes and destination nodes to perform distributed calculations on the tasks during the end-to-end routing process;

[0075] Step 2: Establish an optimization problem of minimizing task execution delay, and propose a computing power perception and load distribution algorithm based on block coordinate descent method to split the original problem into routing and load distribution problems;

[0076] Step 3: Design a computing power-aware routing algorithm based on dynamic programming to select the best relay node, and design a multi-hop coding load distribution algorithm based on the interior point method to achieve communication and computing load balancing.

[0077] The scheme of the present invention is specifically as follows:

[0078] (1) System workflow diagram Figure 1 As shown, the workflow is as follows:

[0079] Step 1: Input the constellation parameters of space-based computing power, including inter-satellite distance, inter-satellite transmission rate, displacement time parameters related to satellite computing power, straggler parameters, task source node, destination node, and number of task rows;

[0080] Step 2: Initialize any feasible routing and load distribution;

[0081] Step 3: Call the computing power-aware routing algorithm based on dynamic programming, fix the load distribution scheme, and search for computing power-aware routing by maximizing the number of coded inner product calculations expected to be completed on the transmission path;

[0082] Step 4: Call the coding load distribution algorithm based on the interior point method to solve the computation and communication load of each hop on the path according to the route obtained in step 3;

[0083] Step 5: Calculate the task execution delay under the routing in step 3 and the load distribution in step 4;

[0084] Step 6: Determine whether the task execution delay is reduced compared with the previous iteration result. If it is reduced, use the current iteration result as the starting point for the next iteration and return to step 3. Otherwise, jump to step 7.

[0085] Step 7: Output computing power routing and computing communication load distribution.

[0086] (2) Task execution delay analysis

[0087] Consider distributed matrix multiplication on a space-based computing constellation. Assume that it takes H hops to transmit from source node S to destination node D. The computing task y = Ax generated by the source node is distributed through the destination node D and H-1 relay nodes, where

[0088] like Figure 2 As shown in the figure, the multi-hop network topology of the space-based computing constellation is defined as a graph G = {U, E, D, R}, where U represents the endpoint of the graph, i.e., the satellite node. For each u∈U, define They represent the number of ISL hops and ISL hops required for the transmission from source node S to node u. u ,but E represents the edge of the graph, i.e. ISL. vu ∈E represents the ISL connecting satellite nodes v and u. Each satellite is equipped with four intersatellite links, two ISLs connecting adjacent co-orbital satellites and two ISLs connecting adjacent off-orbital satellites, forming Figure 1 The topology shown. For each e vu ∈E has two attributes, namely the distance d of ISL vu ∈D and the emission rate R vu ∈R, if e vu is the same-orbit intersatellite link, d vu =d v , R vu =R v , otherwise d vu =d h , R vu =R h .

[0089] There are a total of Q reachable paths from the source node S to the destination node D. The path set is defined as

[0090] P={p1,p2,...,p i ,,p Q} (1)

[0091] Obviously, each path has the same hop count H, so each path has the same number of nodes participating in the calculation. The node set contained in the path set P is denoted as

[0092] The system workflow is as follows Figure 3 As shown, the specific steps are as follows:

[0093] Step S1: The source node S first selects the optimal path p i ,use The MDS code encodes the task matrix A as Then it is divided into H sub-matrices, where the allocation path p i The jth u The submatrix of the jump work node u is a matrix consisting of the encoding combinations of each row in A, is a positive integer. In order to assign computing tasks to each working node, Random linear combinations of L rows from matrix A, so that all nodes can recover the result Ax after completing at least L rows of inner products.

[0094] Step S2: During the task delivery phase, the master node will encode the matrix It is sent to the destination node D hop by hop through H-1 relay nodes. Each relay node receives the corresponding sub-matrix assigned to itself. For the sub-matrix not assigned to itself, the relay node does not store it but directly sends it to the next hop node, that is, a transparent transmission mechanism is adopted.

[0095] Step S3: During the calculation phase, each destination node and each relay node completes the calculation of the sub-matrix assigned to it, until the destination node and each relay node complete at least L results in total.

[0096] The communication delay of each node receiving the assigned submatrix consists of the transmission delay and the propagation delay of the multi-hop ISL. Indicates the selection path p i When the task is transferred to the jth u Transmission time of hopping satellite node u:

[0097]

[0098] Among them, R i,j Represents the path p i The jth u Hop transfer rate, b represents the number of bits in a row of tasks.

[0099] Defining variables The package indicates the selected path p i When the task is transferred to the jth uPropagation delay of satellite hopping node u

[0100]

[0101] Where c is the speed of light.

[0102] Defining random variables Indicates the selection path p i Time u The computation time of the satellite hopping node u, and H random variables Independent of each other. Consider that the computation time of each node follows a two-parameter shift exponential distribution, that is, Worker node u of row vector inner product, calculation time The CDF of is as follows:

[0103]

[0104] in Positive number a u The shift time parameter required for satellite u to complete a single inner product operation, a positive number μ u is a stragling parameter determined by the computing power of satellite u.

[0105] Defining variables Indicates the selection path p i Time u The total delay of the satellite hopping node u to complete the task can be obtained from the above model: is the sum of the computation delay, transmission delay and propagation delay of node u

[0106]

[0107] Then the cumulative probability distribution function (CDF) of the total delay of satellite node u to complete the task is as follows:

[0108]

[0109] Define variable X i,j (t) represents the selected path p i Time u The number of calculations completed by the satellite-hopping node u by time t. For a given time t, if node u does not complete the assigned calculation task, the calculation result of node u is abandoned, so X i,j The expectation of (t) can be expressed as:

[0110]

[0111] Define the random variable T exe Indicates the total delay of the system to complete the task. To ensure that the master node can decode the result, T exeThe total number of inner product calculations completed by the nodes that have completed the calculation before exceeds L. Definition A set of nodes that meet the following conditions

[0112]

[0113] in Represents the path p i The set of nodes participating in the calculation. Then T exe It can be expressed as

[0114]

[0115] exist Figure 4 In the example shown, after relay node 3 completes its own calculation task, the accumulated L-row inner product results have been completed, so

[0116] T exe The CDF of The product of is:

[0117]

[0118] Due to T exe is a positive number, and its expectation can be obtained by integrating the CDF:

[0119]

[0120] (3) Establishing the optimization problem

[0121] The main problem is to find the optimal path p i and optimal load distribution Minimize the average total delay The problem is set up as follows:

[0122]

[0123] in is a set of natural numbers.

[0124] because It is not independent and identically distributed, so it is difficult to obtain elements in , which makes it difficult to obtain according to (10) and (11) therefore, It is impossible to directly solve for . We consider The approximation of is to replace the objective function with a desired constraint, as follows:

[0125]

[0126] In order to ensure the recovery of the result Ax, the expected deadline texe* At least L rows of inner products are calculated, so there is constraint (15). At the same time, for large-scale matrix multiplication, A usually has a large number of rows, so the load distribution is also very large, and rounding has little effect on the result. Therefore, formula (16) further converts The constraint of is relaxed from a positive integer to a positive integer.

[0127] The number of reachable paths Q increases with the increase of the number of hops H. If we further consider the load distribution l i , which will produce more feasible solutions and a larger search space. Therefore, based on the block coordinate descent method, the total problem The problem is broken down into two sub-problems. For the optimal path problem, given the total delay of the initial feasible task execution and load distribution strategy i(0) , find the path p that can complete the calculation with the largest number of inner products i .question The load distribution problem is a constrained optimization problem. Given a path p i Then, find the minimum task execution delay t that can ensure the completion of the task (i.e., the master node receives at least L inner products). exe* and specific load distribution strategies

[0128]

[0129] By continuously executing the above steps, the algorithm will gradually converge and obtain the routing and load distribution solution that minimizes the task execution delay.

[0130] (4) Problem Solving

[0131] is a multi-stage decision problem. If an exhaustive algorithm is used to solve it, the complexity is This paper proposes an algorithm based on dynamic programming, which decomposes the original problem into a series of sub-problems and gradually obtains the solution of the original problem by solving the sub-problems. The algorithm complexity is

[0132] For each relay node, there are only two candidate transmission directions, namely, transmission through the same track ISL or transmission through the different track ISL. Therefore, it is defined that all nodes on the path from the source node S to the node u have a given load distribution l. i Next deadline t exe* The total number of inner products that can be completed is Where node u is the jth node on its path u Jump node. The recursive formula is

[0133]

[0134] Where nodes v and u are path p i The adjacent nodes on the same orbit are connected by intersatellite links, and nodes w and u are on path p o The adjacent nodes connected by intersatellite links on different orbits are defined as It indicates the optimal previous hop node for node u, which is:

[0135]

[0136] Algorithm 1 traverses all Repeat equation (19) and equation (20) until we get So we get the problem The solution process is shown in Table 1.

[0137] Table 1 Computing power-aware routing algorithm based on dynamic programming

[0138]

[0139] The load distribution method under a given route is shown in Table 2, and the complexity is The main steps of Algorithm 2 are as follows:

[0140] Step a1: Read the route obtained by Algorithm 1 as the path, and read the load distribution and task execution delay of the previous iteration as the initial feasible solution (line 1).

[0141] Step a2: Initialize the number of iterations and enter the loop (line 2-3).

[0142] Step a3: Construct barrier function (line 4)

[0143]

[0144] Step a4: The result X of the previous iteration (k-1) ={t exe*(k-1) ,l i (k-1)} as the initial point to solve the unconstrained minimization problem of minimizing the barrier function (line 5)

[0145]

[0146] where r k It is an obstacle factor.

[0147] Step a5: Determine whether the solution converges. If so, exit the loop. Otherwise, reduce the barrier factor r. k The value of , the number of iterations increases by 1, and returns to step a3 (line 6-9).

[0148] Step a6: Output optimal delay and optimal load distribution.

[0149] Table 2 Coding load distribution algorithm based on interior point method

[0150]

[0151] (5) Simulation experiment verification

[0152] To verify the performance of the algorithm, we consider using the satellites of the first phase of Starlink to complete the matrix multiplication operation with a task row number of L. A total of 1584 satellites are deployed in 22 orbits, with 72 satellites in each orbit. The inter-satellite distance d of the satellites in the same orbit is v = 600km, the inter-satellite distance d of different orbit satellites h =1577km. The intersatellite link transmission rate is set to R v =2Mb / s, R h =0.8Mb / s, coding matrix Each row can be represented by 512 bits.

[0153] Considering two scenarios with different heterogeneous performance, the parameter settings are shown in Table 3

[0154] Table 3 Simulation parameter setting table

[0155]

[0156] Among them, the parameters of scenario one are extracted with equal probability from a given finite set, and the parameters of scenario two are generated from a uniformly distributed random source in a given range.

[0157] The schemes used for comparison are as follows:

[0158] 1) Random path - evenly divided without encoding

[0159] A reachable path from a given master node to a target node is randomly selected; the task matrix A is not encoded, and the size of the matrix assigned to each satellite is the same, that is,

[0160] 2) Random Path-HCMM

[0161] A reachable path from a given master node to a target node is randomly selected; the load distribution strategy is calculated by the HCMM algorithm for coding distributed computing in a heterogeneous environment without considering communication.

[0162] 3) Shortest path - evenly divided without encoding

[0163] The shortest path is calculated by the Dijkstra algorithm to ensure the lowest transmission and propagation delay; the mission matrix A is not encoded, and the size of the matrix assigned to each satellite is the same, that is,

[0164] 4) Shortest Path - HCMM

[0165] The shortest path is calculated by the Dijkstra algorithm to ensure the minimum transmission and propagation delay; the load distribution strategy is calculated by the HCMM algorithm for coded distributed computing in a heterogeneous environment without considering communication.

[0166] 5) CRLA

[0167] The optimal routing strategy and load distribution strategy obtained by executing the algorithm proposed in the present invention are optimization results obtained by comprehensively considering the computing power, communication capacity and inter-satellite distance of the satellite.

[0168] from Figure 5 (a) It can be seen that the algorithm proposed in this invention has the lowest average task execution delay compared with the other four algorithms. Figure 5 As shown in (b), with a fixed number of task rows, increasing the number of satellite hops allows more satellites to participate in task execution, thereby reducing the average task execution delay, which shows that coded distributed computing can improve the on-orbit computing speed of space-based computing constellations. As the number of hops increases, the effect of CRLA in reducing the average task execution compared to the shortest path-HCMM increases from 12.44% to 16.88%, and then decreases to 5.57%. The reason for this is that when the communication or computing delay accounts for a large proportion, the shortest path-HCMM can maintain good performance. However, since the shortest path-HCMM does not integrate routing and load distribution, its performance is not as good as CRLA.

[0169] In order to compare the performance of the algorithm in different scenarios, we repeated the experiment in the more heterogeneous scenario 2. When H is fixed to 14, we can see that Figure 5 (c).,The average task execution delay of CRLA is reduced by 24.16% to 26.06% compared with the shortest path-HCMM; Figure 5 In (d), when L = 2000, it is reduced by 7.58% to 25.51%, and reaches a peak when the number of hops is 14. Compared with scenario 1, the shortest path-HCMM algorithm is less effective in reducing task execution delay, indicating that CRLA has more obvious advantages in more heterogeneous scenarios. CRLA can avoid selecting relay nodes with weaker computing power by jointly optimizing routing and load distribution. At the same time, the average number of hops for low-orbit computing constellations is around 14, which shows that our algorithm can significantly reduce the average task execution delay in most cases.

[0170] The computing power routing and load distribution method for space-based computing power constellations proposed in the present invention has the following characteristics:

[0171] 1. Joint optimization: First, find the appropriate satellite to participate in the mission through routing, and then perform specific load distribution on each satellite after determining the route. In order to solve the problem of too large feasible domain caused by the coupling of routing and load distribution, a computing power perception and load distribution (CRLA) algorithm based on block coordinate descent is proposed to ensure low algorithm complexity.

[0172] 2. Computing power awareness: The computing power-aware routing strategy is designed by maximizing the number of coded inner product calculations expected to be completed on the transmission path, fully considering the impact of computing delay on task execution delay. It can evaluate and select satellites with better computing and communication conditions to participate in the task execution while maintaining a low algorithm complexity.

[0173] 3. Load balancing: The load distribution strategy takes into account the randomness of computing delay and the multi-hop characteristics of communication delay. Redundancy is introduced through coding to resist the impact of stragglers slowing down distributed computing. It achieves balance of computing and communication loads for constellations with heterogeneous computing power, and assigns more tasks to nodes with strong capabilities, thereby reducing task execution delay.

[0174] Aiming at the demand of large-scale computing on orbit, this paper proposes a model to embed coded distributed computing into the space-based computing constellation, theoretically analyzes the task execution delay of distributed computing on orbit, establishes the delay minimization problem and proposes a load allocation (CRLA) algorithm to solve the problem, and experimentally verifies that the proposed algorithm is more effective than the existing algorithm in reducing the average task execution delay. Specifically, this algorithm provides a multi-hop routing and load allocation solution for the distributed processing of computing tasks on orbit for the space-based computing constellation, with the following innovative work:

[0175] 1. It is proposed to embed coded distributed computing into the space-based computing constellation, which can not only avoid transmitting a large amount of raw data to the ground center, but also distribute the processing of tasks in the end-to-end routing process instead of processing them on a single satellite, thereby significantly reducing the task execution delay.

[0176] 2. In order to select satellites with higher computing power to participate in mission calculations, the computing power-aware multi-hop routing solution simultaneously evaluates the communication and computing capabilities of the relay satellites, integrating the communication and computing processes.

[0177] 3. The designed load distribution scheme can resist the influence of straggler by introducing coding technology, thereby improving the robustness of the space-based computing constellation and achieving a balance between computing and communication loads through on-orbit processing of tasks.

[0178] The present invention also discloses a multi-hop routing and load distribution system for space-based computing constellations, including:

[0179] Module 1: It is used to establish a model for the space-based computing constellation to perform matrix-vector multiplication calculation tasks, provide source nodes to encode and split tasks, relay nodes to transmit tasks hop by hop, and relay nodes and destination nodes to perform distributed calculations on tasks during end-to-end routing;

[0180] Module 2: Used to establish the optimization problem of minimizing the task execution delay, a computing power awareness and load allocation (CRLA) algorithm based on block coordinate descent method is proposed to split the original problem into routing and load allocation problems;

[0181] Module 3: Used to design a computing power-aware routing algorithm based on dynamic programming to select the best relay node, and design a multi-hop coding load distribution algorithm based on the interior point method to achieve communication and computing load balancing.

[0182] The multi-hop routing and load distribution system operates in the following steps:

[0183] Step 1: Input the constellation parameters of space-based computing power, including inter-satellite distance, inter-satellite transmission rate, displacement time parameters related to satellite computing power, straggler parameters, task source node, destination node, and number of task rows;

[0184] Step 2: Initialize any feasible routing and load distribution;

[0185] Step 3: Call the computing power-aware routing algorithm based on dynamic programming, fix the load distribution scheme, and search for computing power-aware routing by maximizing the number of coded inner product calculations expected to be completed on the transmission path;

[0186] Step 4: Call the coding load distribution algorithm based on the interior point method to solve the computation and communication load of each hop on the path according to the route obtained in step 3;

[0187] Step 5: Calculate the task execution delay under the routing in step 3 and the load distribution in step 4;

[0188] Step 6: Determine whether the task execution delay is reduced compared with the previous iteration result. If it is reduced, use the current iteration result as the starting point for the next iteration and return to step 3. Otherwise, jump to step 7.

[0189] Step 7: Output computing power routing and computing communication load distribution.

[0190] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the multi-hop routing and load distribution method of the present invention when called by a processor.

[0191] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A multi-hop routing and load distribution method for space-based computing constellations, characterized in that: The following steps are involved: Step 1: Establish a model for the space-based computing constellation to perform matrix-vector multiplication tasks, encode and split the tasks at the source node, and transmit the tasks hop by hop at the relay node. In the end-to-end routing process, the relay node and the destination node perform distributed computing on the tasks. Step 2: Establish an optimization problem of minimizing task execution delay, and propose a computing power perception and load distribution algorithm based on block coordinate descent method to split the original problem into routing and load distribution problems; Step 3: Design a computing power-aware routing algorithm based on dynamic programming to select the best relay node, and design a multi-hop coding load distribution algorithm based on the interior point method to achieve communication and computing load balancing.

2. The multi-hop routing and load distribution method according to claim 1, characterized in that: The step one comprises: Step S1: The source node S first selects the optimal path p i ,use The MDS code encodes the task matrix A as Then it is divided into H sub-matrices, where the allocation path p i The jth u The submatrix of the jump work node u is a matrix composed of the encoding combinations of each row in the task matrix A. is a positive integer, Random linear combinations of the L rows from the task matrix A; Step S2: The master node encodes the matrix It is sent to the destination node D hop by hop through H-1 relay nodes; each relay node receives the corresponding sub-matrix assigned to itself. For the sub-matrix not assigned to itself, the relay node does not store it but directly sends it to the next hop node, that is, a transparent transmission mechanism is adopted; Step S3: The destination node and each relay node complete the calculation of the sub-matrix assigned to them, until the destination node and each relay node complete at least L results in total.

3. The multi-hop routing and load distribution method according to claim 2, characterized in that: In step S3, the communication delay of each node receiving the assigned sub-matrix is ​​composed of the transmission delay and the propagation delay of the multi-hop ISL; is the sum of the computation delay, propagation delay and transmission delay of node u, and the formula is as follows: in, Indicates the selection path p i Time u The total delay of the satellite node u to complete the task, Indicates the selection path p i Time u The computation time of the satellite hopping node u, Indicates the selection path p i When the task is transferred to the jth u The propagation delay of the satellite hopping node u, Indicates the selection path p i When the task is transferred to the jth u Transmission time of hopping satellite node u; The CDF of the total delay of satellite node u to complete the task is as follows: Among them, the positive number μ u is the stragling parameter determined by the computing power of satellite u, a positive number u is the shift time parameter required for satellite u to complete a single inner product operation, They represent the number of intra-orbital inter-satellite link hops and the number of inter-orbital inter-satellite link hops required for transmission from source node S to node u, respectively. i,v represents the assignment to path p i The number of rows of the encoding matrix of the v-th hop node on Represents the path p i The jth u Hop transmission rate, considering the mesh satellite network topology, d v represents the distance between satellites in the same orbit, d h represents the distance between the intersatellite links of adjacent satellites in different orbits, and b represents the coding matrix The number of bits in a row, c is the speed of light; T exe The CDF of The product of is expressed as follows: T exe Represents the total delay of the system to complete the task, Indicates that as of T exe The node set that completes the submatrix calculation when Due to T exe is a positive number, and its expectation is obtained by integrating the CDF, as follows:

4. The multi-hop routing and load distribution method according to claim 3, characterized in that: In the step 2, it also includes: Step y1: By finding the optimal path p i and optimal load distribution Minimize the average total delay The problem is set up as follows: in is the set of natural numbers; Step y2: Replace the objective function with a desired constraint, as follows: Among them, t exe* Indicates the expected deadline. Step y3: Based on the block coordinate descent method, the total problem of step y2 is The problem is broken down into two sub-problems. For the optimal path problem, given the total delay of the initial feasible task execution and load distribution strategy i(0) , find the path p that can complete the calculation with the largest number of inner products i ,question The load distribution problem is a constrained optimization problem. Given a path p i Finally, find the minimum task execution delay t that can ensure the completion of the task exe* and specific load distribution strategies By continuously executing the above steps, the algorithm will gradually converge and obtain the routing and load distribution solution that minimizes the task execution delay.

5. The multi-hop routing and load distribution method according to claim 4, characterized in that: In step 3, the computing power-aware routing algorithm based on dynamic programming design runs the following steps: Step 1: Input source node S, destination node D, Total inner product of tasks L, d v ,d h ,R v ,R h , the total delay of the result of the (r-1)th iteration Load distribution vector l i(r-1) ; Step 2: Initial State Step 3: Calculate the state according to equation (19) And according to formula (20), record the optimal previous hop node for node u Step 4: Output the optimal path 6. The multi-hop routing and load distribution method according to claim 5, characterized in that: In step 3, the multi-hop coding load distribution algorithm designed based on the interior point method runs the following steps: Step a1: read the route obtained by the routing algorithm as the path, and read the load distribution and task execution delay of the previous iteration as the initial feasible solution; Step a2: Initialize the number of iterations and enter the loop; Step a3: Construct barrier function Step a4: X (k-1) ={t exe*(k-1) ,l i (k-1) } as the initial point to solve the unconstrained minimum problem of minimizing the barrier function. The formula is as follows: where r k is the barrier factor; Step a5: Determine whether the solution converges. If so, exit the loop. Otherwise, reduce the barrier factor r. k The value of , the number of iterations increases by 1, and the process returns to step a3; Step a6: Output optimal delay and optimal load distribution.

7. The multi-hop routing and load distribution method according to claim 1, characterized in that: In step 1, a distributed matrix multiplication operation is performed on the space-based computing constellation. Assuming that H hops are required for transmission from the source node S to the destination node D, the computing task y=Ax generated by the source node is distributedly calculated through the destination node D and H-1 relay nodes, where 8. A multi-hop routing and load distribution system for space-based computing constellations, characterized in that: It includes: Module 1: used to establish a model for the space-based computing constellation to perform matrix-vector multiplication calculation tasks, provide source nodes to encode and split the tasks, relay nodes to transmit the tasks hop by hop, and relay nodes and destination nodes to perform distributed calculations on the tasks during end-to-end routing; Module 2: Used to establish the optimization problem of minimizing the task execution delay, a computing power perception and load distribution algorithm based on block coordinate descent method is proposed to split the original problem into routing and load distribution problems; Module 3: Used to design a computing power-aware routing algorithm based on dynamic programming to select the best relay node, and design a multi-hop coding load distribution algorithm based on the interior point method to achieve communication and computing load balancing.

9. The multi-hop routing and load distribution system according to claim 8, characterized in that: The multi-hop routing and load distribution system operates in the following steps: Step 1: Input the constellation parameters of space-based computing power, including inter-satellite distance, inter-satellite transmission rate, displacement time parameters related to satellite computing power, straggler parameters, task source node, destination node, and number of task rows; Step 2: Initialize any feasible routing and load distribution; Step 3: Call the computing power-aware routing algorithm based on dynamic programming, fix the load distribution scheme, and search for computing power-aware routing by maximizing the number of coded inner product calculations expected to be completed on the transmission path; Step 4: Call the coding load distribution algorithm based on the interior point method to solve the computation and communication load of each hop on the path according to the route obtained in step 3; Step 5: Calculate the task execution delay under the routing in step 3 and the load distribution in step 4; Step 6: Determine whether the task execution delay is reduced compared with the previous iteration result. If it is reduced, use the current iteration result as the starting point for the next iteration and return to step 3. Otherwise, jump to step 7. Step 7: Output computing power routing and computing communication load distribution, end.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the multi-hop routing and load distribution method according to any one of claims 1 to 7 when called by a processor.

Images