WO2023130656A1 - 一种异构多节点互联拓扑生成方法和存储介质 - Google Patents

一种异构多节点互联拓扑生成方法和存储介质 Download PDF

Info

Publication number
WO2023130656A1
WO2023130656A1 PCT/CN2022/096236 CN2022096236W WO2023130656A1 WO 2023130656 A1 WO2023130656 A1 WO 2023130656A1 CN 2022096236 W CN2022096236 W CN 2022096236W WO 2023130656 A1 WO2023130656 A1 WO 2023130656A1
Authority
WO
WIPO (PCT)
Prior art keywords
heterogeneous multi
node
graph
interconnection topology
node interconnection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/096236
Other languages
English (en)
French (fr)
Inventor
杨宏斌
金良
胡克坤
赵雅倩
董刚
刘海威
蒋东东
晁银银
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Suzhou Wave Intelligent Technology Co Ltd
Original Assignee
Suzhou Wave Intelligent Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Suzhou Wave Intelligent Technology Co Ltd filed Critical Suzhou Wave Intelligent Technology Co Ltd
Publication of WO2023130656A1 publication Critical patent/WO2023130656A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/12—Discovery or management of network topologies
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Definitions

  • the present application relates to the technical field of computer networks, in particular to a heterogeneous multi-node interconnection topology generation method and storage medium.
  • the computing power of a single computing node is different, as large as a server, including multiple CPUs and multiple GPU computing cards, as small as a single dedicated computing chip, including hundreds or thousands of PEs; the interface is different, such as the QPI interface of Intel CPU, Nvidia GPU Nvlink of FPGA, SRIO of FPGA; multi-node interconnection application scenarios are different, such as supercomputer, distributed computing, ultra-heterogeneous platform, network-on-chip, many-core CPU multi-core architecture and multi-PE interconnection of heterogeneous acceleration chip, among which supercomputer, Distributed computing and ultra-heterogeneous platforms use device nodes, such as CPUs, GPUs, FPGAs, and dedicated ICs; network-on-chips, many-core CPU multi-core architectures, and heterogeneous acceleration chips use core nodes, such as CPU cores, CUDA cores,
  • each computing node is also various, such as: shared bus, Crossbar switching matrix (as shown in Figure 1 and Figure 2), Ring (as shown in Figure 3 is the topology diagram of the bus and Ring), Star connection, Mesh and Torus (as shown in Figure 4 is a schematic diagram of a 2D Mesh and 2D Torus distributed switching matrix, and Figure 5 is a schematic diagram of a 2D Torus structure), etc.
  • shared bus As shown in Figure 1 and Figure 2 and Figure 2D
  • Ring is the topology diagram of the bus and Ring
  • Star connection a Mesh and Torus (as shown in Figure 4 is a schematic diagram of a 2D Mesh and 2D Torus distributed switching matrix
  • Figure 5 is a schematic diagram of a 2D Torus structure
  • the network connecting multiple independent computers is called Network
  • the interconnection network inside a chip or between multiple chips is called Fabric
  • NoC noC
  • QoS Quality of Service
  • the receiving and sending ports of each node Crossbar can increase the queue buffer to realize QoS priority control; realize more advanced flow control strategy, make full use of queues; use more advanced routing algorithm and congestion judgment algorithm to calculate which way to go to the target node more smoothly.
  • FIG. 7 shows a schematic diagram of the internal architecture of an image processing chip.
  • Local Router in the figure refers to a local router
  • Global Router refers to a global router
  • Hierarchical Star refers to a hierarchical star
  • RISC Reduced Instruction-Set Computing (reduced instruction Set computer)
  • IIE refers to Integrated Information Environment (integrated information environment)
  • ME refers to Motion Estimation (motion estimation)
  • ST refers to store (storage)
  • SM refers to Shared Memory (shared memory)
  • PMC Power Management Controller (power management controller)
  • VAE refers to Variational autoEncoder (variational automatic encoder)
  • GTMU refers to GSM Transmission Timing Management Unit for BBU (GSM main control transmission unit)
  • OGW refers to Originating Gateway device (originating gateway device )
  • FMP refers to Functional Multiprocessor Architecture (functional multiprocessor structure)
  • VPE refers to Vector Processing Element (vector processing unit)
  • LTMU refers to Local Task Management Units (local task management unit)
  • SPE refers to
  • the chip internally uses 6 Crossbar Switches to form a star network. At the same time, 4 13x13 Crossbars are connected in series to form a Ring.
  • the overall hybrid topology is adopted.
  • the inventor realized that due to the different computing power provided by different computing devices or computing cores, the interface types, numbers and bandwidths are different, and the interconnection topologies between computing nodes are various. High computing power, multiple interfaces, and multiple interconnect lines correspond to high power consumption and cost. If the computing nodes and the interconnection do not match, the computing nodes or interconnect lines will be idle.
  • the heterogeneous multi-node topology selects computing units and connection relationships according to computing tasks. Since the performance of each computing node is different, the cost is also different. If there are many interconnecting lines between two nodes, the communication bandwidth will be high, but the cost will also be high. Larger, it will also increase the probability of channel congestion.
  • the present application provides a method for generating a heterogeneous multi-node interconnection topology, including the following steps:
  • Input feature integration information into a preset generation network to generate a heterogeneous multi-node interconnection topology
  • the eigenvalues of the heterogeneous multi-node interconnection topology are obtained to ensure that the heterogeneous multi-node interconnection topology meets the preset accuracy requirements.
  • the generating network specifically includes: an upsampling layer, a convolutional layer, a fully connected layer, a batch normalization layer, a modified linear unit, and a Sigmoid function.
  • the low-dimensional vector representation includes a node embedding vector and a connection embedding vector; based on the following formula, the node embedding vector of the graph convolutional network model is obtained:
  • v i represents the node Represents the neighbor nodes of node v i
  • e ij represents the connection between nodes v i and v j
  • mean represents the average function
  • f c0 and f c1 represent two feed-forward networks of different sizes
  • w ij e represents the learnable 1x1 weight of the corresponding adjacent edge
  • concat represents the concatenation function, and creates a node vector based on node features
  • v i and v j both represent node.
  • the upsampling process of the upsampling layer specifically includes: assuming that the feature integration information is a graph S(V, E) containing V vertices and E adjacent edges;
  • the first adjacency matrix is trained to obtain the optimal value A ⁇ of the first adjacency matrix.
  • the vertex features of the graph S'(V', E') are obtained based on the following formula:
  • f in is the vertex feature of graph S'(V',E')
  • k ij represents the geodesic distance between vertex j and vertex i of graph S'(V',E')
  • N*n The optimal value f j of the graph S'(V',E') obtained after the weight of any vertex in the vertices represents the vertex characteristics of the graph S(V,E).
  • the convolutional layer generates a global graph and an independent graph based on the upsampling result of the upsampling layer, and performs a convolution operation based on the global graph and the independent graph; the convolution operation specifically includes the following steps: initializing the independent graph For graph S(V,E), an independent graph is generated based on:
  • C k represents the independent graph
  • fin represents the vertex feature W ⁇ k of the graph S'(V',E') and are embedding functions ⁇ and Parameters
  • SoftMax is the normalization function
  • N represents the number of vertices in the graph S'(V',E'), ⁇ (v i ), Respectively represent two 1 ⁇ 1 convolutional layers with different initial values.
  • the eigenvalues of the heterogeneous multi-node interconnection topology are obtained based on the following formula:
  • B k represents the global graph
  • C k represents the independent graph
  • ⁇ represents the parameter to adjust the weight of the independent graph
  • fin represents the vertex features of the graph S'(V', E')
  • K v represents the kernel size of the spatial dimension
  • W k A vector of weights representing the 1x1 convolution operation.
  • ensuring that the heterogeneous multi-node interconnection topology meets the preset accuracy requirements specifically includes: combining the heterogeneous multi-node interconnection topology with the preset real heterogeneous multi-node interconnection topology based on the following formula Perform the cross entropy loss operation:
  • E represents the expected value of the distribution function
  • P data represents the distribution of actual topological samples
  • x is the real sample in P data
  • P z represents the distribution of input noise
  • D(x) represents the probability of judging the sample as correct
  • G(z ) represents the heterogeneous multi-node interconnection topology
  • z represents the input noise
  • P t represents the heterogeneous multi-node interconnection topology
  • P t ' represents the real heterogeneous multi-node interconnection topology
  • L topo represents the topological distance between the heterogeneous multi-node interconnection topology and the corresponding node of the real heterogeneous multi-node interconnection topology
  • ⁇ is the weight of the reconstruction item
  • L rec is the topology reconstruction loss
  • L cGAN is the cross-entropy loss
  • the performance parameters include noise signal, performance requirements, power consumption requirements and cost requirements.
  • ensuring that the heterogeneous multi-node interconnection topology meets preset accuracy requirements includes:
  • the cross-entropy loss results and topology reconstruction loss results of the heterogeneous multi-node interconnection topology and the preset real heterogeneous multi-node interconnection topology are obtained;
  • a non-volatile computer-readable storage medium storing computer-readable instructions, when the computer-readable instructions are executed by one or more processors, one or more processors are made to execute the method provided in any one of the above-mentioned embodiments.
  • the steps of the heterogeneous multi-node interconnection topology generation method are made to execute the method provided in any one of the above-mentioned embodiments.
  • Fig. 1 is a schematic diagram of Crossbar in the prior art
  • Fig. 2 is a schematic diagram of Crossbar cascading in the prior art
  • Fig. 3 is a topological structure diagram of a bus and a Ring in the prior art
  • Fig. 4 is the schematic diagram of 2D Mesh and 2D Torus distributed switch matrix in the prior art
  • Fig. 5 is the structural representation of 2D Torus in the prior art
  • Fig. 6 is a schematic diagram of the topological structure of banded rings and triple rings in the prior art
  • Fig. 7 is a schematic diagram of the internal architecture of an image processing dedicated chip in the prior art.
  • FIG. 8 is a flowchart of a method for generating a heterogeneous multi-node interconnection topology provided in one or more embodiments of the present application;
  • Fig. 9 is a schematic diagram of the application environment of the method provided in one or more embodiments of the present application.
  • FIG. 10 is a schematic diagram of the internal operation structure of the discriminant network provided in one or more embodiments of the present application.
  • FIG. 11 is a schematic diagram of the internal operation structure of the convolutional layer provided in one or more embodiments of the present application.
  • Fig. 12 is a schematic diagram of the internal operation structure of the generating network provided in one or more embodiments of the present application.
  • FIG. 8 is a flowchart of a method for generating a heterogeneous multi-node interconnection topology provided in Embodiment 1. Referring to FIG. 8 , FIG. 8 is a flowchart of a method for generating a heterogeneous multi-node interconnection topology provided in Embodiment 1. Referring to FIG. 8
  • the method provided in this embodiment is applied in the application environment shown in FIG. 9 .
  • the method of this embodiment includes the following steps:
  • Step S1 Based on the graph convolutional network model, feature extraction of node information and topological structure to obtain a low-dimensional vector representation of node information and topological structure.
  • the graph convolutional network model is based on node information library and topology library, feature node information and topology.
  • the node information library includes node models established based on characteristics such as node computing power, node core number, node interface number, and node interface broadband;
  • the topology model established by characteristics such as density and connection line length; performance parameters include performance, power consumption, cost, etc.
  • the node embedding vector of the graph convolutional network model is obtained based on the following formula:
  • v i represents the node Represents the neighbor nodes of node v i
  • e ij represents the connection between nodes v i and v j
  • mean represents the average function
  • Step S2 based on the node embedding vector, connection embedding vector and performance parameters input to the fully connected layer, the node embedding vector, connection embedding vector and performance parameters are fused in the fully connected layer to form feature integration information, and the feature integration information is input into the preset Generate networks to generate heterogeneous multi-node interconnection topologies.
  • the generation network includes, but is not limited to, an upsampling layer and a convolutional layer, as shown in FIG. 10 , which is a schematic diagram of the generation network of this application.
  • the spatial upsampling layer operates using an aggregation function defined by a graph A ⁇ that maps a graph S(V,E) with V vertices and E edges to a larger graph S'(V' , E'), by assigning different importances to new sets of vertices, the network can learn the optimal value of A ⁇ for good upsampling of the graph.
  • the upsampling process of the upsampling layer specifically includes: mapping the graph S(V, E) to a graph S'(V', E') containing N*n vertices and E*m adjacent edges; based on the graph S'(V',E') Generate the first adjacency matrix and obtain the initial value of the first adjacency matrix; based on the initial value of the first adjacency matrix, train the first adjacency matrix to obtain the optimal value of the first adjacency matrix A ⁇ .
  • the graph S'(V', E') vertex features are obtained based on the following formula:
  • f in is the vertex feature of the graph S'(V', E')
  • kij represents the geodesic distance A between the vertex j and the vertex i of the graph S'(V', E')
  • ⁇ is the calculation N* Any of the n vertices
  • f j represents the vertex feature of the graph S(V, E).
  • the convolution layer performs convolution operations based on the global graph and the independent graph.
  • the operation process of the convolutional layer is shown in Figure 11. Specifically, the independent graph is initialized as graph S(V,E); the independent graph is generated based on the following formula:
  • C k represents the independent graph
  • fin represents the vertex features W ⁇ k of the graph S'(V', E') and are embedding functions ⁇ and Parameters
  • SoftMax is a normalization function; where, the normalization function is; where, the normalization function is:
  • C k represents the independent graph
  • fin represents the vertex features W ⁇ k of the graph S'(V', E') and are embedding functions ⁇ and Parameters
  • SoftMax is a normalization function; where, the normalization function is; where, the normalization function is:
  • B k represents the global graph
  • C k represents the independent graph
  • ⁇ represents the parameter to adjust the weight of the independent graph
  • fin represents the vertex features of the graph S'(V', E')
  • K v represents the kernel size of the spatial dimension
  • W k A vector of weights representing the 1x1 convolution operation.
  • Step S3 Based on the pre-built discriminant network, ensure that the heterogeneous multi-node interconnection topology meets the preset accuracy requirements. Among them, there is a similarity between the operation structure of the discrimination network and the generation network, as shown in FIG. 12 , which is the internal structure of the discrimination network of the present application. Specifically, the discriminative network uses an aggregation matrix B ⁇ with trainable weights ⁇ that is different from the weights learned by the generator network, since the aggregation is mapped from a larger graph S'(V', E') to A smaller graph S 1 (V 1 , E 1 ), based on the following formula, obtains the vertex features of graph S 1 (V 1 , E 1 ):
  • f i is the vertex feature of graph S 1 (V 1 , E 1 );
  • f j ' is the vertex feature of graph S'(V', E');
  • k is the graph S'(V', E')
  • the geodesic distance B ⁇ between vertex j and vertex i is an aggregation matrix B ⁇ with trainable weights ⁇ .
  • ensuring that the heterogeneous multi-node interconnection topology meets the preset accuracy requirements specifically includes: intersecting the heterogeneous multi-node interconnection topology with the preset real heterogeneous multi-node interconnection topology based on the following formula entropy loss operation;
  • E represents the expected value of the distribution function
  • P data represents the distribution of actual topological samples
  • x is the real sample in P data
  • P z represents the distribution of input noise
  • D(x) represents the probability of judging the sample as correct
  • G(z ) represents the heterogeneous multi-node interconnection topology
  • z represents the input noise
  • P t represents the heterogeneous multi-node interconnection topology
  • P t ' represents the real heterogeneous multi-node interconnection topology
  • L topo represents the topological distance between the heterogeneous multi-node interconnection topology and the corresponding node of the real heterogeneous multi-node interconnection topology
  • ⁇ is the weight of the reconstruction item
  • L rec is the topology reconstruction loss
  • L cGAN is the cross-entropy loss
  • the method of this embodiment includes the following steps: based on the graph convolutional network model, feature extraction of node information and topology to obtain low-dimensional vector representations of node information and topology; input performance parameters and low-dimensional vector representations to the fully connected layer , to generate feature integration information; input feature integration information into the preset generation network to generate heterogeneous multi-node interconnection topology; based on the pre-built discriminant network, ensure that the heterogeneous multi-node interconnection topology meets the preset accuracy requirements .
  • the generating network specifically includes: an upsampling layer and a convolutional layer; the convolutional layer generates a heterogeneous multi-node interconnection topology based on the upsampling processing result of the upsampling layer.
  • the low-dimensional vector representation includes a node embedding vector and a connection embedding vector; based on the following formula, the node embedding vector of the graph convolutional network model is obtained:
  • v i represents the node Represents the neighbor nodes of node v i
  • e ij represents the connection between nodes v i and v j
  • mean represents the average function
  • the upsampling process of the upsampling layer specifically includes: assuming that the feature integration information is a graph S(V, E) containing V vertices and E adjacent edges; based on the graph S(V, E), sequentially Perform the following operations: map the graph S(V, E) to a graph S'(V', E') containing N*n vertices and E*m adjacent edges; based on the graph S'(V', E' ), generate the first adjacency matrix, and obtain the initial value of the first adjacency matrix; based on the initial value of the first adjacency matrix, train the first adjacency matrix to obtain the optimal value A ⁇ of the first adjacency matrix.
  • the graph S'(V', E') vertex features are obtained based on the following formula:
  • f in is the vertex feature of the graph S'(V', E')
  • kij represents the geodesic distance A between the vertex j and the vertex i of the graph S'(V', E')
  • ⁇ is the calculation N* Any of the n vertices
  • f j represents the vertex feature of the graph S(V, E).
  • the convolutional layer generates a global graph and an independent graph based on the upsampling result of the upsampling layer, and performs a convolution operation based on the global graph and the independent graph;
  • the convolution operation specifically includes the following steps: initializing the independent graph as a graph S(V,E), which generates an independent graph based on:
  • C k represents the independent graph
  • fin represents the vertex features W ⁇ k of the graph S'(V', E') and are embedding functions ⁇ and Parameters
  • SoftMax is a normalization function; where, the normalization function is; where, the normalization function is:
  • N represents the number of vertices ⁇ (v i ) of the graph S'(V', E') Respectively represent two 1x1 convolutional layers with different initial values.
  • the eigenvalues of the heterogeneous multi-node interconnection topology are obtained based on the following formula:
  • B k represents the global graph
  • C k represents the independent graph
  • ⁇ represents the parameter to adjust the weight of the independent graph
  • fin represents the vertex features of the graph S'(V', E')
  • K v represents the kernel size of the spatial dimension
  • W k A vector of weights representing the 1x1 convolution operation.
  • B k is a global map, which is unique to each layer.
  • Bk is an independent graph for learning a per-sample specific topology.
  • ⁇ and are two embedding functions, here a 1x1 convolutional layer.
  • Kv denotes the number of subgraphs, represents the residual operation
  • Represents the matrix multiplication operation is the gate that controls the importance weights of the two graphs. The importance of independent graphs in different layers is adjusted through a gating mechanism, using a different ⁇ value for each layer that is learned and updated through training.
  • ensuring that the heterogeneous multi-node interconnection topology meets preset accuracy requirements specifically includes:
  • heterogeneous multi-node interconnection topology and the preset real heterogeneous multi-node interconnection topology are calculated based on the following formula for cross-entropy loss;
  • E represents the expected value of the distribution function
  • P data represents the distribution of actual topological samples
  • x is the real sample in P data
  • P z represents the distribution of input noise
  • D(x) represents the probability of judging the sample as correct
  • G(z ) represents the heterogeneous multi-node interconnection topology
  • z represents the input noise
  • P t represents the heterogeneous multi-node interconnection topology
  • P t ' represents the real heterogeneous multi-node interconnection topology
  • L topo represents the topological distance between the heterogeneous multi-node interconnection topology and the corresponding node of the real heterogeneous multi-node interconnection topology
  • ⁇ is the weight of the reconstruction item
  • L rec is the topology reconstruction loss
  • L cGAN is the cross-entropy loss
  • the upsampling operation includes: the upsampling process of the upsampling layer specifically includes: assuming that the feature integration information is a graph S(V, E) containing V vertices and E adjacent edges; Perform the following operations: map the graph S(V, E) to a graph S'(V', E') containing N*n vertices and E*m adjacent edges; based on the graph S'(V', E' ), generate the first adjacency matrix, and obtain the initial value of the first adjacency matrix; based on the initial value of the first adjacency matrix, train the first adjacency matrix to obtain the best value A ⁇ of the first adjacency matrix; Get the best value A ⁇ , get the vertex features of the graph S'(V', E').
  • f in is the vertex feature of the graph S'(V', E')
  • kij represents the geodesic distance A between the vertex j and the vertex i of the graph S'(V', E')
  • ⁇ is the calculation N* The optimal value of the graph S'(V', E') obtained after the weight of any vertex in the n vertices
  • f j represents the vertex feature of the graph S(V, E).
  • the graph convolution operation includes the convolution layer generating a global graph and an independent graph based on the upsampling results of the upsampling layer, and performing convolution operations based on the global graph and the independent graph; the convolution operation specifically includes the following steps: initializing the independent graph as a graph S( V,E), generating an independent graph based on:
  • C k represents the independent graph
  • fin represents the vertex features W ⁇ k of the graph S'(V', E') and are embedding functions ⁇ and Parameters
  • SoftMax is a normalization function; where, the normalization function is; where, the normalization function is:
  • N represents the number of vertices ⁇ (v i ) of the graph S'(V', E'), Respectively represent two 1x1 convolutional layers with different initial values.
  • B k represents the global graph
  • C k represents the independent graph
  • ⁇ represents the parameter to adjust the weight of the independent graph
  • fin represents the vertex features of the graph S'(V', E')
  • K v represents the kernel size of the spatial dimension
  • W k represents 1x1 A vector of weights for the convolution operation.
  • the performance parameters include noise signal, performance requirements, power consumption requirements and cost requirements.
  • This embodiment provides a non-volatile computer-readable storage medium storing computer-readable instructions.
  • the program When the program is executed by one or more processors, one or more processors execute the program provided in any of the above-mentioned embodiments.
  • the steps of the heterogeneous multi-node interconnect topology generation method are described in detail below.
  • the embodiments in the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the embodiment of the present application may be in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present application may take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes therein. .
  • Embodiments of the present application are described with reference to flowcharts and/or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.
  • These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions
  • the device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Design And Manufacture Of Integrated Circuits (AREA)

Abstract

本申请涉及一种异构多节点互联拓扑生成方法和存储介质，方法包括以下步骤：基于图卷积网络模型，特征提取节点信息和拓扑结构，以获取节点信息和拓扑结构的低维向量表示；输入性能参数和所述低维向量表示至全连接层，以生成特征整合信息；将所述特征整合信息输入预设的生成网络，以生成异构多节点互联拓扑结构；获取所述异构多节点互联拓扑结构的特征值，确保所述异构多节点互联拓扑结构满足预设的准确性要求。

Description

一种异构多节点互联拓扑生成方法和存储介质
相关申请的交叉引用
本申请要求于2022年01月10日提交中国专利局，申请号为202210024578.5，申请名称为“一种异构多节点互联拓扑生成方法和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机网络技术领域，尤其是指一种异构多节点互联拓扑生成方法和存储介质。
背景技术
在目前的异构计算领域中，存在多种类型的计算设备，如CPU、GPU、FPGA和专用IC等。单个计算节点的算力不同，大到服务器，包含多个CPU和多张GPU计算卡，小到单个专用计算芯片，包含几百上千个PE；接口不同，如Intel CPU的QPI接口，Nvidia GPU的Nvlink，FPGA的SRIO；多节点的互联应用场景不同，比如超级计算机、分布式计算、超异构平台、片上网络、众核CPU多核架构和异构加速芯片多PE互联等，其中超级计算机、分布式计算和超异构平台这些使用设备节点，如CPU、GPU、FPGA和专用IC等；片上网络、众核CPU多核架构和异构加速芯片使用核心节点，如CPU核心、CUDA核心和PE阵列等。
各计算节点的互联方式也多种多样，典型的如：共享总线、Crossbar交换矩阵(如图1和图2所示)、Ring(如图3所示即为总线和Ring的拓扑结构图)、星形连接、Mesh和Torus(如图4所示即为2D Mesh和2D Torus分布式交换矩阵示意图，如图5所示即为2D Torus的结构示意图)等。通常将连接多台独立计算机的网络称为Network，将芯片内部或者多个芯片之间的互联网络称为Fabric，将嵌入到单个芯片内部的用于连接片内多个不同模块的大规模互联网络称为NoC。不同于计算机网络，在NoC中可以实现更多特色技术，如QoS(Quality of Service)，可以决定先转发哪个请求，后转发哪个，每个节点Crossbar的接收和发送端口可以增加队列缓冲，以实现QoS优先级控制；实现更先进的流量控制策略，更加充分地利用队列；利用更先进的路由算法和拥塞判断算法，算出从哪条路走到目标节点更加顺畅。常用的Fabric拓扑如HyperCube超立方也是Intel QPI所使用的拓扑，胖树(Fat Tree)拓扑也是天河II超级计算机连接大量计算机节点使用的拓扑，金字塔(Pyramid)拓扑，蝴蝶(Butterfly)拓扑，Intel在其12核心的Ivy Bridge CPU微架构中串接其12核心使用的带弦环和三重环(Triple Ring)，带弦环(如图6所示即为带弦环和三重环拓扑结构示意图)，带环立方体和ClosNetwork拓扑等。如图7所示的图像处理专用芯片内部架构示意图，图中的Local Router是指局部路由器，Global Router是指全局路由器，Hierarchical Star是指分级星形，RISC是指Reduced Instruction-Set Computing(精简指令集计算机)，IIE是指Integrated Information Environment(综合信息环境)，ME是指Motion Estimation(运动估测)，ST是指store(存储)，SM是指Shared Memory(共享内存)，PMC是指Power Management Controller(电源管理控制器)，VAE是指Variational autoEncoder(变分自动编码器)，GTMU是指GSM Transmission Timing Management Unit for BBU(GSM主控传输单元)，OGW是指Originating Gateway device(始发网关设备)，FMP是指Functional  Multiprocessor Architecture(功能型多处理器结构)，VPE是指Vector Processing Element(向量处理单元)，LTMU是指Local Task Management Units(本地任务管理单元)，SPE是指Scalar Processing Element(标量处理单元)，FEC是指Forwarding Equivalence Class(转发等价类)。该芯片内部使用了6个Crossbar Switch形成了一个星形网络，同时又将4个13x13的Crossbar串接起来形成了一个Ring，总体采用混合拓扑方式。其中，发明人意识到，由于不同的计算设备或计算核心能提供的算力不同，其接口类型、数量和带宽不同，各计算节点之间的互联拓扑又多种多样。算力高、接口多、互联线多对应的功耗和成本高，如果计算节点和互联不匹配会造成计算节点空闲或互联线空闲，因此在部署一个计算任务或设计异构多节点硬件电路时，节点类型、数量如何选择，采用哪种拓扑，各节点如何互联，使其在满足计算性能的情况下尽量低功耗低成本，是一个优化问题。
也就是说，不同于芯片的布局布线任务其网表中确定的单元和连接关系，异构多节点拓扑是根据计算任务来选取计算单元和连接关系。由于每个计算节点的性能不同，成本也不同，两节点间互联线多，其通信带宽就高，但成本也高；相反，如果采用级联方式，其级数越多，对应的延迟可能越大，也会增大通道拥塞的概率。
发明内容
本申请提供了一种异构多节点互联拓扑生成方法，包括以下步骤：
基于图卷积网络模型，特征提取节点信息和拓扑结构，以获取节点信息和拓扑结构的低维向量表示；
输入性能参数和低维向量表示至全连接层，以生成特征整合信息；
将特征整合信息输入预设的生成网络，以生成异构多节点互联拓扑结构；及
获取异构多节点互联拓扑结构的特征值，确保异构多节点互联拓扑结构满足预设的准确性要求。
在本申请的一个实施例中，生成网络具体包括：上采样层、卷积层、全连接层、批归一化层、修正线性单元和S型函数。
在本申请的一个实施例中，低维向量表示包括节点嵌入向量和连接嵌入向量；基于下式，获取得到所述图卷积网络模型的节点嵌入向量：
Figure PCTCN2022096236-appb-000001
其中，v i表示节点
Figure PCTCN2022096236-appb-000002
表示节点v i的邻居节点，e ij表示节点v i和v j的连接，mean表示平均值函数；基于下式，获取得到图卷积网络模型的连接嵌入向量：
Figure PCTCN2022096236-appb-000003
其中，f c0和f c1表示两个不同大小的前馈网络，w ij e表示对应邻边的可学习的1x1权重，concat表示拼接函数，基于节点特征创建节点向量，v i和v j均表示节点。
在本申请的一个实施例中，上采样层的上采样过程具体包括：假设特征整合信息为一个包含V个顶点和E个邻边的图S(V,E)；
基于图S(V,E)依次进行如下操作：将图S(V,E)映射到一个包含N*n个顶点和E*m个邻边的图 S'(V',E')；
基于图S'(V',E')，生成第一邻接矩阵，并获得第一邻接矩阵的初始值；及
基于第一邻接矩阵的初始值，对第一邻接矩阵进行训练，以获得第一邻接矩阵的最佳值A ω。
在本申请的一个实施例中，基于下式获得图S'(V',E')的顶点特征：
Figure PCTCN2022096236-appb-000004
其中，f in为图S'(V',E')的顶点特征，k ij表示图S'(V',E')的顶点j和顶点i之间的测地距离，为计算N*n个顶点中任一顶点的权重后得到的图S'(V',E')的最佳值f j表示图S(V,E)的顶点特征。
在本申请的一个实施例中，卷积层基于上采样层的上采样结果生成全局图和独立图，并基于全局图和独立图进行卷积运算；卷积运算具体包括以下步骤：初始化独立图为图S(V,E)，基于下式生成独立图：
Figure PCTCN2022096236-appb-000005
其中，C k表示独立图，f in表示图S'(V',E')的顶点特征W θk和
Figure PCTCN2022096236-appb-000006
分别为嵌入函数θ和
Figure PCTCN2022096236-appb-000007
的参数，SoftMax为归一化函数；其中，归一化函数为：
Figure PCTCN2022096236-appb-000008
其中，N表示图S'(V',E')的顶点数量,θ(v i)、
Figure PCTCN2022096236-appb-000009
分别表示两个初始值不等的1×1卷积层。
在本申请的一个实施例中，基于下式获取异构多节点互联拓扑结构的特征值：
Figure PCTCN2022096236-appb-000010
其中，B k表示全局图，C k表示独立图α表示调整独立图权重的参数，f in表示图S'(V'，E')的顶点特征，K v表示空间维的核大小，W k表示1x1卷积操作的权重向量。
在本申请的一个实施例中，确保异构多节点互联拓扑结构满足预设的准确性要求具体包括：将异构多节点互联拓扑结构与预设的真实异构多节点互联拓扑结构基于下式进行交叉熵损失运算：
Figure PCTCN2022096236-appb-000011
其中，E表示分布函数的期望值，P data代表实际拓扑样本的分布，x是P data中的真实样本，P z表示输入噪音的分布，D(x)表示判别样本为正确的概率，G(z)表示异构多节点互联拓扑图；z表示输入噪音；基于下式，获取拓扑重建损失结果：
Figure PCTCN2022096236-appb-000012
其中，P t表示异构多节点互联拓扑结构P t'表示真实异构多节点互联拓扑结构，L topo表示异构多节点互联拓扑结构与真实异构多节点互联拓扑结构对应节点的拓扑距离；根据拓扑重建损失与交叉熵损失运算结果，基于下式获取异构多节点互联拓扑结构的最终损失：
L＝L cGAN+λL rec
其中，λ为重建项加权L rec为拓扑重建损失L cGAN为交叉熵损失；将异构多节点互联拓扑结构的最终损失与预设的异构多节点互联拓扑结构的损失进行比较，若异构多节点互联拓扑结构的最终损失大于预设的异构多节点互联拓扑结构的损失，则重复执行生成异构多节点互联拓扑结构直至异构多节点互联拓扑结构的最终损失不大于预设的异构多节点互联拓扑结构的损失，确保异构多节点互联拓扑结构满足预设的准确性要求。
在本申请的一个实施例中，性能参数包括噪音信号、性能要求、功耗要求和成本要求。
在本申请的一个实施例中，确保异构多节点互联拓扑结构满足预设的准确性要求，包括：
基于异构多节点互联拓扑结构的特征值，获得异构多节点互联拓扑结构与预设的真实异构多节点互联拓扑结构的交叉熵损失结果及拓扑重建损失结果；
基于交叉熵损失结果及拓扑重建损失结果，获得异构多节点互联拓扑结构的最终损失；及
将异构多节点互联拓扑结构的最终损失与预设的异构多节点互联拓扑结构的损失进行比较,若异构多节点互联拓扑结构的最终损失大于预设的异构多节点互联拓扑结构的损失,则重复执行上采样层的上采样过程和卷积层的卷积运算直至异构多节点互联拓扑结构的最终损失不大于预设的异构多节点互联拓扑结构的损失。
为实现上述目的，本申请还提出第二技术方案：
一种存储有计算机可读指令的非易失性计算机可读存储介质，当计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行上述任一实施例中提供的异构多节互联拓扑生成方法的步骤。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1是现有技术中Crossbar示意图；
图2是现有技术中Crossbar级联示意图；
图3是现有技术中总线和Ring的拓扑结构图；
图4是现有技术中2D Mesh和2D Torus分布式交换矩阵示意图；图5是现有技术中2D Torus的结构 示意图；
图6是现有技术中带弦环和三重环拓扑结构示意图；
图7是现有技术中图像处理专用芯片内部架构示意图；
图8是本申请一个或多个实施例中提供的异构多节互联拓扑生成方法的流程图；
图9是本申请一个或多个实施例中提供的方法的应用环境示意图；
图10是的本申请一个或多个实施例中提供判别网络的内部运算结构示意图；
图11是的本申请一个或多个实施例中提供卷积层的内部运算结构示意图；
图12是的本申请一个或多个实施例中提供生成网络的内部运算结构示意图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合本申请的实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例，都属于本申请保护的范围。
实施例一：
参照图8所示，图8为实施例一提供给的异构多节互联拓扑生成方法流程图。
本实施例提供的该方法应用于如图9所示的应用环境中。本实施例的方法，包括以下步骤：
步骤S1、基于图卷积网络模型，特征提取节点信息和拓扑结构，以获取节点信息和拓扑结构的低维向量表示。
在其中一个实施方式中，图卷积网络模型基于节点信息库和拓扑结构库，特征节点信息和拓扑结构。其中，节点信息库包括分别基于节点算力、节点核心数量、节点接口数量和节点接口宽带等特征建立的节点模型；所述拓扑结构库包括基于连接方式、连接点数量、连接级数、连接线密度、连接线长度等特征建立的拓扑结构模型；性能参数包括性能、功耗、成本等。
在其中一个实施方式中，基于性能参数、节点信息和拓扑结构的低维向量表示，基于下式，获取得到图卷积网络模型的节点嵌入向量：
Figure PCTCN2022096236-appb-000013
其中，v i表示节点
Figure PCTCN2022096236-appb-000014
表示节点v i的邻居节点，e ij表示节点v i和v j的连接，mean表示平均值函数；基于下式，获取得到图卷积网络模型的连接嵌入向量：
Figure PCTCN2022096236-appb-000015
其中，
Figure PCTCN2022096236-appb-000016
和
Figure PCTCN2022096236-appb-000017
表示两个不同大小的前馈网络，
Figure PCTCN2022096236-appb-000018
是对应于邻边的可学习的1x1权重，concat表示拼接函数，基于节点特征创建节点向量，v i和v j均表示节点。
步骤S2、基于节点嵌入向量、连接嵌入向量和性能参数输入至全连接层，将节点嵌入向量、连接嵌入向量和性能参数在全连接层融合之后形成特征整合信息，将特征整合信息输入预设的生成网络，以生成 异构多节点互联拓扑结构。
在其中一个实施方式中，生成网络包括但是不限于上采样层和卷积层，如图10所示即为本申请的生成网络的示意图。其中，空间上采样层使用由图A ω定义的聚合函数进行操作，该图由一个包含V个顶点和E个边的图S(V,E)映射到一个更大的图S'(V'，E')，通过给新顶点集分配不同的重要度，网络可以学习到A ω的最佳值，从而对图进行良好的上采样。上采样层的上采样过程具体包括：将图S(V,E)映射到一个包含N*n个顶点和E*m个邻边的图S'(V'，E')；基于图S'(V'，E')生成第一邻接矩阵，并获得第一邻接矩阵的初始值；基于第一邻接矩阵的初始值，对第一邻接矩阵进行训练，以获得第一邻接矩阵的最佳值A ω。基于下式获得图S'(V'，E')顶点特征：
Figure PCTCN2022096236-appb-000019
其中，f in为图S'(V'，E')的顶点特征，k ij表示图S'(V'，E')的顶点j和顶点i之间的测地距离A ω为计算N*n个顶点中任一；顶点的权重后得到的图S'(V'，E')的最佳值；f j表示图S(V,E)的顶点特征。
在其中一个实施例中，在上采样层进行上采样处理之后，卷积层基于全局图和独立图进行卷积运算。卷积层的运算过程如图11所示，具体地，初始化独立图为图S(V,E)；基于下式生成独立图：
Figure PCTCN2022096236-appb-000020
其中，C k表示独立图，f in示图S'(V'，E')的顶点特征W θk和
Figure PCTCN2022096236-appb-000021
分别为嵌入函数θ和
Figure PCTCN2022096236-appb-000022
的参数，SoftMax为归一化函数；其中，归一化函数为；其中，归一化函数为：
Figure PCTCN2022096236-appb-000023
其中，C k表示独立图，f in示图S'(V'，E')的顶点特征W θk和
Figure PCTCN2022096236-appb-000024
分别为嵌入函数θ和
Figure PCTCN2022096236-appb-000025
的参数，SoftMax为归一化函数；其中，归一化函数为；其中，归一化函数为：
Figure PCTCN2022096236-appb-000026
其中，B k表示全局图，C k表示独立图α表示调整独立图权重的参数，f in表示图S'(V'，E')的顶点特征，K v表示空间维的核大小，W k表示1x1卷积操作的权重向量。
步骤S3、基于预构建的判别网络，确保异构多节点互联拓扑结构满足预设的准确性要求。其中，判别网络与生成网络的运算结构存在相似性，如图12所示，即为本申请的判别网络的内部结构。具体地，判别网络使用的是与生成网络学习的权值不同、具有可训练权值φ的聚合矩阵B φ，由于该聚合是从一个较大 的图S'(V'，E')映射到一个较小的图S 1(V 1，E 1)，基于下式，获得图S 1(V 1，E 1)的顶点特征：
Figure PCTCN2022096236-appb-000027
其中，f i为图S 1(V 1，E 1)的顶点特征；f j'表示图S’(V‘，E’)的顶点特征；k表示图S’(V‘，E’)的顶点j和顶点i之间的测地距离B φ为具有可训练权值φ的聚合矩阵B φ。
在其中一个实施例中，确保异构多节点互联拓扑结构满足预设的准确性要求具体包括：将异构多节点互联拓扑结构与预设的真实异构多节点互联拓扑结构基于下式进行交叉熵损失运算；
Figure PCTCN2022096236-appb-000028
其中，E表示分布函数的期望值，P data代表实际拓扑样本的分布，x是P data中的真实样本，P z表示输入噪音的分布，D(x)表示判别样本为正确的概率，G(z)表示异构多节点互联拓扑图；z表示输入噪音；基于下式，获取拓扑重建损失结果：
Figure PCTCN2022096236-appb-000029
其中，P t表示异构多节点互联拓扑结构P t'表示真实异构多节点互联拓扑结构，L topo表示异构多节点互联拓扑结构与真实异构多节点互联拓扑结构对应节点的拓扑距离；根据拓扑重建损失与交叉熵损失运算结果，基于下式获取异构多节点互联拓扑结构的最终损失：
L＝L cGAN+λL rec
其中，λ为重建项加权L rec为拓扑重建损失L cGAN为交叉熵损失；将异构多节点互联拓扑结构的最终损失与预设的异构多节点互联拓扑结构的损失进行比较，若异构多节点互联拓扑结构的最终损失大于预设的异构多节点互联拓扑结构的损失，则重复执行步骤S2直至异构多节点互联拓扑结构的最终损失不大于预设的异构多节点互联拓扑结构的损失，确保异构多节点互联拓扑结构满足预设的准确性要求。
实施例二：
本实施例的方法，包括以下步骤：基于图卷积网络模型，特征提取节点信息和拓扑结构，以获取节点信息和拓扑结构的低维向量表示；输入性能参数和低维向量表示至全连接层，以生成特征整合信息；将特征整合信息输入预设的生成网络，以生成异构多节点互联拓扑结构；基于预构建的判别网络，确保异构多节点互联拓扑结构满足预设的准确性要求。
在其中一个实施方式中，生成网络具体包括：上采样层和卷积层；卷积层基于上采样层的上采样处理结果，以生成异构多节点互联拓扑结构。
在其中一个实施方式中，低维向量表示包括节点嵌入向量和连接嵌入向量；基于下式，获取得到图卷积网络模型的节点嵌入向量：
Figure PCTCN2022096236-appb-000030
其中，v i表示节点
Figure PCTCN2022096236-appb-000031
表示节点v i的邻居节点，e ij表示节点v i和v j的连接，mean表示平均值函数；基于下式，获取得到图卷积网络模型的连接嵌入向量：
Figure PCTCN2022096236-appb-000032
其中，
Figure PCTCN2022096236-appb-000033
和
Figure PCTCN2022096236-appb-000034
表示两个不同大小的前馈网络，
Figure PCTCN2022096236-appb-000035
是对应于邻边的可学习的1x1权重，concat表示拼接函数，基于节点特征创建节点向量，v i和v j均表示节点。
在其中一个实施方式中，上采样层的上采样过程具体包括：假设特征整合信息为一个包含V个顶点和E个邻边的图S(V,E)；基于图S(V,E)依次进行如下操作：将图S(V,E)映射到一个包含N*n个顶点和E*m个邻边的图S'(V'，E')；基于图S'(V'，E')，生成第一邻接矩阵，并获得第一邻接矩阵的初始值；基于第一邻接矩阵的初始值，对第一邻接矩阵进行训练，以获得第一邻接矩阵的最佳值A ω。
在其中一个实施方式中，基于下式获得图S'(V'，E')顶点特征：
Figure PCTCN2022096236-appb-000036
其中，f in为图S'(V'，E')的顶点特征，k ij表示图S'(V'，E')的顶点j和顶点i之间的测地距离A ω为计算N*n个顶点中任一；顶点的权重后得到的图S'(V'，E')的最佳值；f j表示图S(V,E)的顶点特征。
在其中一个实施方式中，卷积层基于上采样层的上采样结果生成全局图和独立图，并基于全局图和独立图进行卷积运算；卷积运算具体包括以下步骤：初始化独立图为图S(V,E)，基于下式生成独立图：
Figure PCTCN2022096236-appb-000037
其中，C k表示独立图，f in示图S'(V'，E')的顶点特征W θk和
Figure PCTCN2022096236-appb-000038
分别为嵌入函数θ和
Figure PCTCN2022096236-appb-000039
的参数，SoftMax为归一化函数；其中，归一化函数为；其中，归一化函数为：
Figure PCTCN2022096236-appb-000040
其中，N表示图S'(V'，E')的顶点数量θ(v i)
Figure PCTCN2022096236-appb-000041
分别表示两个初始值不等的1x1卷积层。
在其中一个实施方式中，基于下式获取异构多节点互联拓扑结构的特征值：
Figure PCTCN2022096236-appb-000042
其中，B k表示全局图，C k表示独立图α表示调整独立图权重的参数，f in表示图S'(V'，E')的顶点特征，K v表示空间维的核大小，W k表示1x1卷积操作的权重向量。卷积层的内部的具体地运算过程如图11所示，图中，B k是全局图，对每一层是唯一的。B k是独立图，用于学习每个样本特定的拓扑结构。
θ和
Figure PCTCN2022096236-appb-000043
是两个嵌入函数，在这里为1x1的卷积层。K v表示子图的数量，
Figure PCTCN2022096236-appb-000044
表示残差操作
Figure PCTCN2022096236-appb-000045
表示矩阵乘操作
Figure PCTCN2022096236-appb-000046
是控制两种图重要性权重的门。通过门控机制来调整不同层中独立图的重要性，对每层使用一个不同的α该值通过训练来学习和更新。
在其中一个实施方式中，确保异构多节点互联拓扑结构满足预设的准确性要求具体包括：
将异构多节点互联拓扑结构与预设的真实异构多节点互联拓扑结构基于下式进行交叉熵损失运算；
Figure PCTCN2022096236-appb-000047
其中，E表示分布函数的期望值，P data代表实际拓扑样本的分布，x是P data中的真实样本，P z表示输入噪音的分布，D(x)表示判别样本为正确的概率，G(z)表示异构多节点互联拓扑图；z表示输入噪音；基于下式，获取拓扑重建损失结果：
Figure PCTCN2022096236-appb-000048
其中，P t表示异构多节点互联拓扑结构P t'表示真实异构多节点互联拓扑结构，L topo表示异构多节点互联拓扑结构与真实异构多节点互联拓扑结构对应节点的拓扑距离；根据拓扑重建损失与交叉熵损失运算结果，基于下式获取异构多节点互联拓扑结构的最终损失：
L＝L cGAN+λL rec
其中，λ为重建项加权L rec为拓扑重建损失L cGAN为交叉熵损失；将异构多节点互联拓扑结构的最终损失与预设的异构多节点互联拓扑结构的损失进行比较，若异构多节点互联拓扑结构的最终损失大于预设的异构多节点互联拓扑结构的损失，则重复执行上采样操作和图卷积操作直至异构多节点互联拓扑结构的最终损失不大于预设的异构多节点互联拓扑结构的损失，确保异构多节点互联拓扑结构满足预设的准确性要求。其中，上采样操作包括：上采样层的上采样过程具体包括：假设特征整合信息为一个包含V个顶点和E个邻边的图S(V,E)；基于图S(V,E)依次进行如下操作：将图S(V,E)映射到一个包含N*n个顶点和E*m个邻边的图S'(V'，E')；基于图S'(V'，E')，生成第一邻接矩阵，并获得第一邻接矩阵的初始值；基于第一邻接矩阵的初始值，对第一邻接矩阵进行训练，以获得第一邻接矩阵的最佳值A ω；基 于最佳值A ω，获得图S'(V'，E')的顶点特征。基于下式获得图S'(V'，E')顶点特征：
Figure PCTCN2022096236-appb-000049
其中，f in为图S'(V'，E')的顶点特征，k ij表示图S'(V'，E')的顶点j和顶点i之间的测地距离A ω为计算N*n个顶点中任一顶点的权重后得到的图S'(V'，E')的最佳值；f j表示图S(V,E)的顶点特征。
图卷积操作包括卷积层基于上采样层的上采样结果生成全局图和独立图，并基于全局图和独立图进行卷积运算；卷积运算具体包括以下步骤：初始化独立图为图S(V,E)，基于下式生成独立图：
Figure PCTCN2022096236-appb-000050
其中，C k表示独立图，f in示图S'(V'，E')的顶点特征W θk和
Figure PCTCN2022096236-appb-000051
分别为嵌入函数θ和
Figure PCTCN2022096236-appb-000052
的参数，SoftMax为归一化函数；其中，归一化函数为；其中，归一化函数为：
Figure PCTCN2022096236-appb-000053
N表示图S'(V'，E')的顶点数量θ(v i)、
Figure PCTCN2022096236-appb-000054
分别表示两个初始值不等的1x1卷积层。
基于下式获取异构多节点互联拓扑结构的特征值：基于下式获取异构多节点互联拓扑结构的特征值：
Figure PCTCN2022096236-appb-000055
其中，B k表示全局图，C k表示独立图α表示调整独立图权重的参数，f in表示图S'(V'，E')的顶点特征K v表示空间维的核大小W k表示1x1卷积操作的权重向量。
在其中一个实施方式中，性能参数包括噪音信号、性能要求、功耗要求和成本要求。
实施例三：
本实施例提供一种存储有计算机可读指令的非易失性计算机可读存储介质，当程序被一个或多个处理器执行时，使得一个或多个处理器执行上述任一实施例中提供的异构多节互联拓扑生成方法的步骤。
本领域内的技术人员应明白，本申请实施例中的实施例可提供为方法、系统、或计算机程序产品。因此，本申请实施例中可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请实施例中可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请实施例中是参照根据本申请实施例中实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌 入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
注意，上述仅为本申请的较佳实施例及所运用技术原理。本领域技术人员会理解，本申请不限于这里的特定实施例，对本领域技术人员来说能够进行各种明显的变化、重新调整和替代而不会脱离本申请的保护范围。因此，虽然通过以上实施例对本申请进行了较为详细的说明，但是本申请不仅仅限于以上实施例，在不脱离本申请构思的情况下，还可以包括更多其它等效实施例，而本申请的范围由所附的权利要求范围决定。

Claims (11)

  1. 一种异构多节点互联拓扑生成方法，其特征在于：所述方法包括以下步骤：
    基于图卷积网络模型，特征提取节点信息和拓扑结构，以获取节点信息和拓扑结构的低维向量表示；
    输入性能参数和所述低维向量表示至全连接层，以生成特征整合信息；
    将所述特征整合信息输入预设的生成网络，以生成异构多节点互联拓扑结构；及
    获取所述异构多节点互联拓扑结构的特征值，确保所述异构多节点互联拓扑结构满足预设的准确性要求。
  2. 根据权利要求1所述的异构多节点互联拓扑生成方法，其特征在于：所述生成网络具体包括：上采样层、卷积层、全连接层、批归一化层、修正线性单元和S型函数。
  3. 根据权利要求1所述的异构多节点互联拓扑生成方法，其特征在于：所述低维向量表示包括节点嵌入向量和连接嵌入向量；
    基于下式，获取得到所述图卷积网络模型的节点嵌入向量：
    Figure PCTCN2022096236-appb-100001
    其中ν i表示节点，
    Figure PCTCN2022096236-appb-100002
    表示节点ν i的邻居节点，e ij表示节点ν i和ν j的连接，mean表示平均值函数；及
    基于下式，获取得到所述图卷积网络模型的连接嵌入向量：
    Figure PCTCN2022096236-appb-100003
    其中，f c0和f c1表示两个不同大小的前馈网络，w ij e表示对应邻边的可学习的1x1权重，concat表示拼接函数，基于节点特征创建节点向量，v i和v j均表示节点。
  4. 根据权利要求2所述的异构多节点互联拓扑生成方法，其特征在于：所述上采样层的上采样过程具体包括：假设特征整合信息为一个包含V个顶点和E个邻边的图S(V,E)；基于所述图S(V,E)依次进行如下操作：
    将所述图S(V,E)映射到一个包含N*n个顶点和E*m个邻边的图S'(V'，E')；
    基于所述图S'(V'，E')，生成第一邻接矩阵，并获得所述第一邻接矩阵的初始值；及
    基于所述第一邻接矩阵的初始值，对所述第一邻接矩阵进行训练，以获得所述第一邻接矩阵的最佳值A ω。
  5. 根据权利要求4所述的异构多节点互联拓扑生成方法，其特征在于：基下式获得所述图S'(V'，E')的顶点特征：
    Figure PCTCN2022096236-appb-100004
    其中，f in为图S'(V'，E')的顶点特征k ij表示图S'(V'，E')的顶点j和顶点i之间的测地距离,A ω为 计算所述N*n个顶点中任一顶点的权重后得到的图S'(V'，E')的最佳值，f j表示图S(V,E)的顶点特征。
  6. 根据权利要求2所述的异构多节点互联拓扑生成方法，其特征在于：所述卷积层基于所述上采样层的上采样结果生成全局图和独立图，并基于所述全局图和所述独立图进行卷积运算；所述卷积运算具体包括以下步骤：
    初始化所述独立图为图S(V,E),并基于下式生成所述独立图：：
    Figure PCTCN2022096236-appb-100005
    其中，C k表示独立图，f in表示图S'(V'，E')的顶点特征W θk和
    Figure PCTCN2022096236-appb-100006
    分别为嵌入函数θ和
    Figure PCTCN2022096236-appb-100007
    的参数，SoftMax为归一化函数；及
    其中，所述归一化函数为：
    Figure PCTCN2022096236-appb-100008
    其中，N表示图S'(V'，E')的顶点数量，θ(v i)、
    Figure PCTCN2022096236-appb-100009
    分别表示两个初始值不等的1x1卷积层。
  7. 根据权利要求6所述的异构多节点互联拓扑生成方法，其特征在于：基于下式获取所述异构多节点互联拓扑结构的特征值：
    Figure PCTCN2022096236-appb-100010
    其中B k表示全局图，C k表示独立图α表示调整独立图权重的参数。f in表示图S'(V'，E')的顶点特征K v表示空间维的核大小W K表示1x1卷积操作的权重向量。
  8. 根据权利要求7所述的异构多节点互联拓扑生成方法，其特征在于：确保所述异构多节点互联拓扑结构满足预设的准确性要求具体包括：
    将所述异构多节点互联拓扑结构与预设的真实异构多节点互联拓扑结构，基于下式进行交叉熵损失运算：
    Figure PCTCN2022096236-appb-100011
    其中，E表示分布函数的期望值，P data代表实际拓扑样本的分布，x是P data中的真实样本，P z表示输入噪音的分布，D(x)表示判别样本为正确的概率，G(z)表示异构多节点互联拓扑图；z表示输入噪音；
    基于下式，获取拓扑重建损失结果：
    Figure PCTCN2022096236-appb-100012
    其中，P t表示异构多节点互联拓扑结构P t’表示真实异构多节点互联拓扑结构，L topo表示异构多节点 互联拓扑结构与真实异构多节点互联拓扑结构对应节点的拓扑距离；
    根据所述拓扑重建损失结果与所述交叉熵损失运算结果，基于下式获取所述异构多节点互联拓扑结构的最终损失：
    L＝L cGAN+λL rec
    其中，λ为重建项加权，L rec为拓扑重建损失，L cGAN为交叉熵损失；及
    将所述异构多节点互联拓扑结构的最终损失与预设的异构多节点互联拓扑结构的损失进行比较，若所述异构多节点互联拓扑结构的最终损失大于所述预设的异构多节点互联拓扑结构的损失，则重复执行权利要求4～7直至所述异构多节点互联拓扑结构的最终损失不大于所述预设的异构多节点互联拓扑结构的损失，确保所述异构多节点互联拓扑结构满足预设的准确性要求。
  9. 根据权利要求1所述的异构多节点互联拓扑生成方法，其特征在于：所述性能参数包括噪音信号、性能要求、功耗要求和成本要求。
  10. 根据权利要求1所述的异构多节点互联拓扑生成方法，其特征在于，所述确保所述异构多节点互联拓扑结构满足预设的准确性要求，包括：
    基于所述异构多节点互联拓扑结构的特征值，获得所述异构多节点互联拓扑结构与预设的真实异构多节点互联拓扑结构的交叉熵损失结果及拓扑重建损失结果；
    基于所述交叉熵损失结果及所述拓扑重建损失结果，获得所述异构多节点互联拓扑结构的最终损失；及
    将所述异构多节点互联拓扑结构的最终损失与预设的异构多节点互联拓扑结构的损失进行比较,若所述异构多节点互联拓扑结构的最终损失大于所述预设的异构多节点互联拓扑结构的损失,则重复执行所述上采样层的上采样过程和所述卷积层的卷积运算直至所述异构多节点互联拓扑结构的最终损失不大于所述预设的异构多节点互联拓扑结构的损失。
  11. 一个或多个存储有计算机可读指令的非易失性计算机可读存储介质，其特征在于，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如权利要求1至10中任一项所述的方法的步骤。
PCT/CN2022/096236 2022-01-10 2022-05-31 一种异构多节点互联拓扑生成方法和存储介质 Ceased WO2023130656A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210024578.5 2022-01-10
CN202210024578.5A CN114050975B (zh) 2022-01-10 2022-01-10 一种异构多节点互联拓扑生成方法和存储介质

Publications (1)

Publication Number Publication Date
WO2023130656A1 true WO2023130656A1 (zh) 2023-07-13

Family

ID=80196189

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/096236 Ceased WO2023130656A1 (zh) 2022-01-10 2022-05-31 一种异构多节点互联拓扑生成方法和存储介质

Country Status (2)

Country Link
CN (1) CN114050975B (zh)
WO (1) WO2023130656A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117560312A (zh) * 2023-10-25 2024-02-13 中勍科技股份有限公司 一种基于交换机的自适应网络动态拓扑方法及装置
CN117807939A (zh) * 2023-12-29 2024-04-02 苏州异格技术有限公司 电路超图确定方法、装置、计算机设备及存储介质
CN118657193A (zh) * 2024-05-31 2024-09-17 浪潮电子信息产业股份有限公司 异构计算平台及其模型数据处理方法、装置、设备、介质
CN121580834A (zh) * 2025-11-28 2026-02-27 合肥益丕基信息技术有限公司 一种基于图神经网络的管道拓扑关系构建方法

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114050975B (zh) * 2022-01-10 2022-04-19 苏州浪潮智能科技有限公司 一种异构多节点互联拓扑生成方法和存储介质
CN114726739B (zh) * 2022-04-18 2024-04-09 深圳市智象科技有限公司 拓扑数据处理方法、装置、设备及存储介质
CN114884908B (zh) 2022-04-29 2024-02-13 浪潮电子信息产业股份有限公司 一种数据同步方法、装置、设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210034737A1 (en) * 2019-07-30 2021-02-04 Sakif Hossain Khan Detection of adverserial attacks on graphs and graph subsets
CN112417219A (zh) * 2020-11-16 2021-02-26 吉林大学 基于超图卷积的超边链接预测方法
CN112925989A (zh) * 2021-01-29 2021-06-08 中国计量大学 一种属性网络的群体发现方法及系统
CN114050975A (zh) * 2022-01-10 2022-02-15 苏州浪潮智能科技有限公司 一种异构多节点互联拓扑生成方法和存储介质

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111861178B (zh) * 2020-07-13 2024-06-07 北京嘀嘀无限科技发展有限公司 服务匹配模型的训练方法、服务匹配方法、设备及介质
CN111914484B (zh) * 2020-08-07 2024-03-26 中国南方电网有限责任公司 一种用于电网暂态稳定评估的递归图卷积网络系统
CN112163219A (zh) * 2020-08-27 2021-01-01 北京航空航天大学 一种基于词嵌入和gcn的恶意程序识别和分类方法
CN112651492B (zh) * 2020-12-30 2023-10-13 广州大学华软软件学院 一种自连接宽度图卷积神经网络模型系统及训练方法
CN112800903B (zh) * 2021-01-19 2022-08-26 南京邮电大学 一种基于时空图卷积神经网络的动态表情识别方法及系统
CN113222328B (zh) * 2021-03-25 2022-02-25 中国科学技术大学先进技术研究院 基于路段污染相似性的空气质量监测设备布点选址方法
CN113240187B (zh) * 2021-05-26 2022-10-11 合肥工业大学 预测模型生成方法、系统、设备、存储介质及预测方法
CN113904786B (zh) * 2021-06-29 2023-05-30 重庆大学 一种基于线路拓扑分析和潮流特性的虚假数据注入攻击辨识方法
CN113780470B (zh) * 2021-09-28 2024-03-08 西安聚全网络科技有限公司 一种基于自适应多通道交叉图卷积网络的图分类方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210034737A1 (en) * 2019-07-30 2021-02-04 Sakif Hossain Khan Detection of adverserial attacks on graphs and graph subsets
CN112417219A (zh) * 2020-11-16 2021-02-26 吉林大学 基于超图卷积的超边链接预测方法
CN112925989A (zh) * 2021-01-29 2021-06-08 中国计量大学 一种属性网络的群体发现方法及系统
CN114050975A (zh) * 2022-01-10 2022-02-15 苏州浪潮智能科技有限公司 一种异构多节点互联拓扑生成方法和存储介质

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117560312A (zh) * 2023-10-25 2024-02-13 中勍科技股份有限公司 一种基于交换机的自适应网络动态拓扑方法及装置
CN117807939A (zh) * 2023-12-29 2024-04-02 苏州异格技术有限公司 电路超图确定方法、装置、计算机设备及存储介质
CN118657193A (zh) * 2024-05-31 2024-09-17 浪潮电子信息产业股份有限公司 异构计算平台及其模型数据处理方法、装置、设备、介质
CN121580834A (zh) * 2025-11-28 2026-02-27 合肥益丕基信息技术有限公司 一种基于图神经网络的管道拓扑关系构建方法

Also Published As

Publication number Publication date
CN114050975A (zh) 2022-02-15
CN114050975B (zh) 2022-04-19

Similar Documents

Publication Publication Date Title
WO2023130656A1 (zh) 一种异构多节点互联拓扑生成方法和存储介质
US12147829B2 (en) Data processing system and method for heterogeneous architecture
Chen et al. NoC-based DNN accelerator: A future design paradigm
US20220374288A1 (en) Distributed placement of linear operators for accelerated deep learning
US20220029845A1 (en) Scalable in-network computation for massively-parallel shared-memory processors
US20220350598A1 (en) Instruction processing apparatus, acceleration unit, and server
CN111191784A (zh) 转置的稀疏矩阵乘以稠密矩阵用于神经网络训练
CN113449859A (zh) 一种数据处理方法及其装置
WO2015196833A1 (zh) 一种集成网络路由器的众核处理器系统及其集成方法和实现方法
US20240311668A1 (en) Optimizing quantum computing circuit state partitions for simulation
US20240311667A1 (en) Simulating quantum computing circuits using sparse state partitioning
CN119692408B (zh) 一种视觉Transformer加速器实现方法和系统
US20230229916A1 (en) Scalable tensor network contraction using reinforcement learning
US20260105340A1 (en) Simulating quantum computing circuits using kronecker factorization
CN114707648A (zh) 一种基于存内计算的图卷积网络软硬件协同加速方法
CN117278552A (zh) 基于端边云自适应协同的视频或图像分析方法及装置
CN115222014A (zh) 用于神经网络模型执行的加速单元及服务器
CN114691457A (zh) 一种确定硬件性能的方法、装置、存储介质以及电子设备
Truong et al. Hybrid electrical/optical switch architectures for training distributed deep learning in large-scale
CN115016943B (zh) 一种并行计算方法、系统、设备及存储介质
CN120123103B (zh) 面向异构集群环境的模型流水线划分及部署方法和系统
Ortega-Cisneros Design and implementation of an noc-based convolution architecture with gemm and systolic arrays
Ouyang et al. PRLM: A parallel loading mechanism for a deep neural network accelerator based on NoC
CN117389921A (zh) 一种存储扩展卡和系统
JP2023109670A (ja) 加速システム及びその動的な構成方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22918114

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22918114

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC ( EPO FORM 1205A DATED 02/01/2025 ).

122 Ep: pct application non-entry in european phase

Ref document number: 22918114

Country of ref document: EP

Kind code of ref document: A1