CN118394486B - Calculation network task scheduling method and device - Google Patents
Calculation network task scheduling method and device Download PDFInfo
- Publication number
- CN118394486B CN118394486B CN202410832822.XA CN202410832822A CN118394486B CN 118394486 B CN118394486 B CN 118394486B CN 202410832822 A CN202410832822 A CN 202410832822A CN 118394486 B CN118394486 B CN 118394486B
- Authority
- CN
- China
- Prior art keywords
- network
- power
- supply
- calculation
- matrix
- 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.)
- Active
Links
- 238000004364 calculation method Methods 0.000 title claims abstract description 277
- 238000000034 method Methods 0.000 title claims abstract description 34
- 239000011159 matrix material Substances 0.000 claims abstract description 236
- 238000011156 evaluation Methods 0.000 claims abstract description 37
- 238000004891 communication Methods 0.000 claims description 29
- 238000004422 calculation algorithm Methods 0.000 claims description 13
- 230000008859 change Effects 0.000 claims description 10
- 230000009466 transformation Effects 0.000 claims description 6
- 239000013598 vector Substances 0.000 description 21
- 230000008569 process Effects 0.000 description 8
- 238000003860 storage Methods 0.000 description 5
- 230000005540 biological transmission Effects 0.000 description 4
- 238000007667 floating Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 239000000523 sample Substances 0.000 description 3
- HPTJABJPZMULFH-UHFFFAOYSA-N 12-[(Cyclohexylcarbamoyl)amino]dodecanoic acid Chemical group OC(=O)CCCCCCCCCCCNC(=O)NC1CCCCC1 HPTJABJPZMULFH-UHFFFAOYSA-N 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000008278 dynamic mechanism Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000000977 initiatory effect Effects 0.000 description 1
- 238000005304 joining Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/48—Program initiating; Program switching, e.g. by interrupt
- G06F9/4806—Task transfer initiation or dispatching
- G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
- G06F9/4881—Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5066—Algorithms for mapping a plurality of inter-dependent sub-tasks onto a plurality of physical CPUs
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The embodiment of the specification discloses a method and a device for dispatching a task of a computing power network. The method of the embodiment of the specification comprises the following steps: forming an evaluation matrix based on the demand matrix and the supply matrix; and (3) carrying out matching degree scoring on the supply and demand computing capacity of the evaluation matrix to form a computing power supply set comprising a plurality of computing power supply nodes arranged from high to low according to the matching degree scoring. And taking one calculation force supply node in the calculation force supply set as a target node, taking a calculation force demand node as a source node, and calculating the shortest delay path between the two nodes. When the shortest delay path meets the network delay constraint, the current calculation force supply node is used as a calculation force supply node with high matching degree, and the shortest delay path between the current calculation force supply node and the calculation force demand node is output. And when the shortest delay path does not meet the network delay constraint, traversing the next computing power supply node in the computing power supply set to perform the calculation and judgment until a scheduling strategy meeting the network delay constraint is obtained.
Description
Technical Field
One or more embodiments of the present disclosure relate to the field of computer network technologies, and in particular, to a method and an apparatus for scheduling tasks in a computing power network.
Background
As a distributed system, a computing power network faces many challenges and bottlenecks in computing power resource collection and computing power scheduling. These challenges are related not only to the technical level, but also to aspects of system design, reliability, etc.
The collection of computational resources involves obtaining information from the various nodes in the distributed network to learn about their availability, performance, and resource conditions. However, the heterogeneous nature of distributed networks is a major challenge. Nodes may have different types of hardware and software configurations, resulting in variability in resource information, which increases the complexity of resource collection. Furthermore, the dynamics of the nodes is also a problem. Nodes may join or leave the network at any time and their resource information may change, thus requiring a dynamic mechanism to manage and update the resource information.
Computational power scheduling involves assigning tasks to individual nodes and efficiently utilizing their computational power resources to complete the tasks. One major challenge is how to efficiently allocate and utilize the computational power resources of the various nodes to maximize the performance and efficiency of the system. This involves a complex optimization problem that requires consideration of several factors such as task priority, availability of nodes, performance and load conditions. The task scheduling policy should be able to dynamically adjust the allocation and scheduling of tasks based on the nature of the tasks and the real-time state of the system. Due to the dynamics of nodes in the computing network (e.g., joining, exiting, or failing nodes) and the diversity of tasks, the scheduling system needs to have a certain degree of dynamics and flexibility to adapt to changing environments.
Disclosure of Invention
One or more embodiments of the present disclosure describe a method and an apparatus for scheduling a task of a computing power network, where by performing overall planning on a demand parameter of a computing power demand end, a network transmission layer capability, and a capability index of a computing power supply end, a computing power supply node that is most matched is obtained by computing, and a shortest delay path between the computing power supply node and the computing power demand node that is most matched is further output, so as to output an optimal scheduling scheme.
In a first aspect, an embodiment of the present disclosure provides a computing power network task scheduling method, which is executed in a computing power scheduling device, including:
Forming a demand matrix of a demand side of the power calculation network based on indexes of a CPU, a GPU, a memory and a hard disk required by power calculation demand nodes in the power calculation network, and forming a supply matrix of a supply side of the power calculation network based on indexes of the CPU, the GPU, the memory and the hard disk which can be supplied by a plurality of power calculation supply nodes in the power calculation network;
performing dot product operation on the demand matrix and the supply matrix to obtain an evaluation matrix of the calculation network; the evaluation matrix stores comprehensive CPU, GPU, memory and hard disk indexes of a plurality of calculation power supply nodes in the calculation power network;
based on comprehensive CPU, GPU, memory and hard disk indexes, carrying out supply-demand matching degree scoring on each calculation power supply node in the evaluation matrix, and arranging each calculation power supply node from high to low according to the supply-demand matching degree scoring to form a calculation power supply set;
based on bandwidth and delay indexes among all computing nodes in the computing network, respectively forming a bandwidth adjacency matrix and a delay adjacency matrix of the computing network; the power calculation nodes in the power calculation network comprise power calculation supply nodes, power calculation demand nodes and network communication nodes which are positioned between the power calculation supply nodes and the power calculation demand nodes and are used for realizing communication between the power calculation supply nodes and the power calculation demand nodes;
Taking the force calculation demand side as a source node, taking the first force calculation supply node of the force calculation supply set as a destination node, and calculating the shortest delay path between the demand side and the supply side based on a delay adjacency matrix of a force calculation network by using Dijkstra algorithm;
and judging whether the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side, if so, outputting the shortest delay path calculated based on the current calculation force supply node as a destination node, otherwise, traversing the calculation force supply set in sequence, taking the next calculation force supply node as the destination node, and repeating the calculation flow of the shortest delay path between the demand side and the supply side until the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side.
In some embodiments, forming a demand matrix of a demand side of the power computing network based on CPU, GPU, memory, and hard disk metrics required by power computing demand nodes in the power computing network includes:
According to the service type of the power network demand side, determining the CPU, GPU, memory and hard disk indexes required by the service;
and forming a demand matrix of the demand side of the power calculation network based on the CPU, GPU, memory and hard disk indexes required by the service.
In some embodiments, forming a supply matrix of a power network supply side based on indexes of a CPU, a GPU, a memory, and a hard disk that can be supplied by a plurality of power supply nodes in the power network includes:
Acquiring related data of indexes of a plurality of computing power supply nodes representing a CPU, a GPU, a memory and a hard disk in a computing power network, and respectively forming a CPU supply matrix, a GPU supply matrix, a memory supply matrix and a hard disk supply matrix;
The CPU supply matrix, the GPU supply matrix, the memory supply matrix, and the hard disk supply matrix constitute a supply matrix on the power network supply side.
In some embodiments, the scoring the supply-demand matching degree of each computing power supply node in the evaluation matrix based on the comprehensive CPU, GPU, memory, and hard disk index includes:
Determining index weight coefficients of a CPU, a GPU, a memory and a hard disk related to a service according to the service type of a power calculation network demand side;
And carrying out weighted average calculation on the comprehensive CPU, GPU, memory and hard disk indexes of each computing power supply node in the evaluation matrix by using the obtained CPU, GPU, memory and hard disk index weight coefficients to obtain the supply and demand matching degree score of each computing power supply node.
In some embodiments, forming a bandwidth adjacency matrix and a delay adjacency matrix of the power network based on bandwidth and delay indexes among power nodes in the power network, respectively, includes:
acquiring a network topology structure of the power calculation network and bandwidth and delay indexes among nodes in the power calculation network;
Forming a bandwidth adjacency matrix of the computing power network based on a network topology structure of the computing power network and bandwidth indexes among nodes in the computing power network; forming a delay adjacency matrix of the power calculation network based on a network topology structure of the power calculation network and delay indexes among nodes in the power calculation network; the bandwidth index between the nodes without direct connection is 0, and the delay index between the nodes without direct connection is infinity.
In some embodiments, based on the bandwidth and the delay index between each power node in the power network, a bandwidth adjacency matrix and a delay adjacency matrix of the power network are respectively formed, and the method further comprises:
Setting the inter-node bandwidth index smaller than the bandwidth constraint value at the demand side in the bandwidth adjacent matrix of the computing power network as 0, setting the inter-node bandwidth index not smaller than the bandwidth constraint value at the demand side in the bandwidth adjacent matrix of the computing power network as 1, and further forming the bandwidth adjacent matrix of the computing power network after change;
when the bandwidth index between nodes in the bandwidth adjacent matrix of the changed computing power network is 0, setting the delay index between related nodes in the delay adjacent matrix of the computing power network to infinity, and further forming the delay adjacent matrix of the changed computing power network;
and using the time delay adjacency matrix of the changed computing power network for the subsequent shortest time delay path calculation.
In some embodiments, the bandwidth constraint value and the network delay constraint value on the demand side are determined according to the traffic type on the demand side of the power network.
In a second aspect, an embodiment of the present disclosure provides a computing power network task scheduling device, provided in a computing power scheduling apparatus, including:
The first module is used for forming a demand matrix of a demand side of the power calculation network based on indexes of a CPU, a GPU, a memory and a hard disk required by the power calculation demand nodes in the power calculation network, and forming a supply matrix of a supply side of the power calculation network based on indexes of the CPU, the GPU, the memory and the hard disk which can be supplied by a plurality of power calculation supply nodes in the power calculation network;
the second module is used for carrying out dot multiplication operation on the demand matrix and the supply matrix to obtain an evaluation matrix of the calculation network; the evaluation matrix stores comprehensive CPU, GPU, memory and hard disk indexes of a plurality of calculation power supply nodes in the calculation power network;
The matching module is used for scoring the supply and demand matching degree of each computing power supply node in the evaluation matrix based on the comprehensive CPU, the GPU, the memory and the hard disk indexes, and arranging each computing power supply node from high to low according to the supply and demand matching degree score to form a computing power supply set;
the third module is used for respectively forming a bandwidth adjacency matrix and a delay adjacency matrix of the power calculation network based on bandwidth and delay indexes among all power calculation nodes in the power calculation network; the power calculation nodes in the power calculation network comprise power calculation supply nodes, power calculation demand nodes and network communication nodes which are positioned between the power calculation supply nodes and the power calculation demand nodes and are used for realizing communication between the power calculation supply nodes and the power calculation demand nodes;
the computing module is used for taking the computing force demand side as a source node, taking the first computing force supply node of the computing force supply set as a destination node, and computing the shortest delay path between the demand side and the supply side based on a delay adjacency matrix of a computing force network by using a Dijkstra algorithm;
And the judging module is used for judging whether the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side, if so, outputting the shortest delay path calculated based on the current calculation force supply node as a destination node, otherwise, traversing the calculation force supply set in sequence, taking the next calculation force supply node as the destination node, triggering the calculating module to repeat the shortest delay path calculating flow between the demand side and the supply side until the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side.
In some embodiments, the matching module comprises:
The weight coefficient acquisition unit is used for determining the index weight coefficients of the CPU, the GPU, the memory and the hard disk related to the service according to the service type of the power calculation network demand side;
The computing unit is used for carrying out weighted average computation on the comprehensive CPU, GPU, memory and hard disk indexes of each computing power supply node in the evaluation matrix by utilizing the obtained CPU, GPU, memory and hard disk index weight coefficients to obtain supply and demand matching degree scores of each computing power supply node;
and the forming unit is used for arranging each calculation force supply node from high to low according to the supply-demand matching degree score to form a calculation force supply set.
In some embodiments, the third module comprises:
The acquisition unit is used for acquiring the network topology structure of the power calculation network and the bandwidth and delay indexes among all nodes in the power calculation network;
The matrix forming unit is used for forming a bandwidth adjacency matrix of the power calculation network based on the network topology structure of the power calculation network and bandwidth indexes among all nodes in the power calculation network; forming a delay adjacency matrix of the power calculation network based on a network topology structure of the power calculation network and delay indexes among nodes in the power calculation network; wherein, the bandwidth index between nodes without direct connection is 0, and the delay index between nodes without direct connection is infinity;
The first transformation unit is used for setting the inter-node bandwidth index of the bandwidth constraint value smaller than the bandwidth constraint value at the demand side in the bandwidth adjacent matrix of the computational power network to be 0, setting the inter-node bandwidth index of the bandwidth adjacent matrix of the computational power network not smaller than the bandwidth constraint value at the demand side to be 1, and further forming a bandwidth adjacent matrix of the varied computational power network;
the second transformation unit is used for setting the delay index between related nodes in the delay adjacent matrix of the computing power network to be infinite when the bandwidth index between the nodes in the bandwidth adjacent matrix of the computing power network after the change is 0, so as to form the delay adjacent matrix of the computing power network after the change; the time delay adjacency matrix of the changed computing power network is used for the subsequent shortest time delay path calculation.
The technical scheme provided by some embodiments of the present specification has the following beneficial effects:
According to the method and the device for dispatching the task of the power computing network, evaluation and quantification of power computing resources of all power computing supply nodes and network resources among the nodes in the power computing network are achieved, and then a demand matrix, a supply matrix, a delay adjacency matrix and a bandwidth adjacency matrix of the power computing network are built at a power computing dispatching equipment side; forming an evaluation matrix based on the demand matrix and the supply matrix, and quantitatively evaluating a matching relation between a computing force supply end and a computing force demand end in a force network based on the evaluation matrix to obtain a computing force supply node set which is arranged from high to low according to a matching degree score; and finally, an optimal network running path is planned according to the current situation of the computational power network between the computational power supply node and the computational power demand node with high matching, so that intensive and efficient utilization of computational power network resources is realized.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present description, the drawings that are required in the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present description, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a system architecture diagram of a power computing scheduling device for collecting and distributing resources between nodes and nodes in a power computing network;
FIG. 2 is a flowchart of a method for scheduling tasks in a computing power network according to an embodiment of the present disclosure;
FIG. 3 is an example block diagram of a method of task scheduling for a computing network in accordance with an embodiment of the present disclosure;
fig. 4 is a block diagram of a computing power network task scheduling device according to an embodiment of the present disclosure.
Detailed Description
The technical solutions in the embodiments of the present specification will be clearly and completely described below with reference to the drawings in the embodiments of the present specification.
The terms first, second, third and the like in the description and in the claims and in the above drawings are used for distinguishing between different objects and not necessarily for describing a particular sequential or chronological order. Furthermore, the terms "comprise" and "have," as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those listed steps or elements but may include other steps or elements not listed or inherent to such process, method, article, or apparatus.
As in fig. 1, the computing network is formed by a plurality of computing nodes forming a topology. The power calculation network comprises a plurality of power calculation supply nodes on a supply side, a plurality of power calculation demand nodes on a demand side, and a plurality of network communication nodes for realizing communication between the supply measurement and the demand side nodes. The network communication node is located in a communication network between a provisioning side and a demand side.
The computing power supply node is a computing device with a certain computing power resource and capable of processing the service on the demand side, such as a server cluster and a cloud server cluster. The power demand node is an initiating terminal of service request processing, is generally arranged at a service system end, and can be computing equipment at the service system end, such as intelligent mobile equipment, a terminal computer, a small-scale server node and terminal equipment (such as a monitoring camera, a network-connected automobile terminal and the like) in a service scene. The service types are various, and according to different service systems, the services such as scientific calculation and simulation, financial modeling analysis, image video processing, data mining machine learning, multimedia storage, file server or backup, content distribution and the like can be available. The network communication node is a network device with communication capability, such as a switch, router, modem, etc.
In order to implement the task scheduling of the power computing network in the embodiment of the present specification, a resource collection engine is required to be set at the power computing supply side, multiple types of power computing resources at the power computing supply side are collected, evaluation tests of performance, capacity and other parameter indexes are performed on the power computing resources, and relevant results are fed back to the power computing scheduling device; network probes are arranged at each network communication node of the communication network and are used for collecting network link information, network bandwidth information and network delay information between the network communication node and the power calculation supply node, between the network communication node and between the network communication node and the power calculation demand node, and synchronizing related information to the power calculation scheduling equipment.
The power computing scheduling device utilizes the method of the embodiment of the specification to realize task scheduling according to the service request of the power computing demand side, the reported power computing resource information and the reported network capability information.
Each of the power demand nodes on the power demand side transmits the resource demand and the network demand required by the service according to the type of the service being processed. And the power computing scheduling equipment sequentially calculates the optimal scheduling strategy according to the demands of power computing demand nodes according to the priority of service processing and the currently reported power computing resource information and network capability information. And after the task scheduling of a certain calculation force demand node is processed, performing task scheduling on the service demand node initiated by the task under the next priority. At this time, it is necessary to collect information again by using the resource collection engine and the network probe, and perform task scheduling based on the updated computing resource information and network capability information in the computing network.
The computing power scheduling device may be a computing device, such as a server, having a certain computing power. The power calculation scheduling device is provided with a power calculation scheduling platform, the power calculation scheduling platform collects the resource measurement information and the network capacity information of a power calculation network transmission layer, and when the power calculation demands reported by a demand side are received, an optimal scheduling strategy can be calculated.
As shown in fig. 2, an embodiment of the present disclosure provides a method for scheduling tasks of a computing power network, including:
102, forming a demand matrix of a demand side of the power computing network based on indexes of a CPU, a GPU, a memory and a hard disk required by power computing demand nodes in the power computing network, and forming a supply matrix of a supply side of the power computing network based on indexes of the CPU, the GPU, the memory and the hard disk which can be supplied by a plurality of power computing supply nodes in the power computing network;
104, performing dot product operation on the demand matrix and the supply matrix to obtain an evaluation matrix of the calculation network; the evaluation matrix stores comprehensive CPU, GPU, memory and hard disk indexes of a plurality of calculation power supply nodes in the calculation power network;
Step 106, based on the comprehensive CPU, GPU, memory and hard disk indexes, carrying out supply and demand matching degree scoring on each calculation force supply node in the evaluation matrix, and arranging each calculation force supply node from high to low according to the supply and demand matching degree scoring to form a calculation force supply set;
step 108, respectively forming a bandwidth adjacency matrix and a delay adjacency matrix of the power calculation network based on bandwidth and delay indexes among all power calculation nodes in the power calculation network; the power calculation nodes in the power calculation network comprise power calculation supply nodes, power calculation demand nodes and network communication nodes which are positioned between the power calculation supply nodes and the power calculation demand nodes and are used for realizing communication between the power calculation supply nodes and the power calculation demand nodes;
Step 110, using the force calculation demand side as a source node, using the first force calculation supply node of the force calculation supply set as a destination node, and calculating the shortest delay path between the demand side and the supply side based on a delay adjacency matrix of the force calculation network by using Dijkstra algorithm;
And 112, judging whether the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side, if so, outputting the shortest delay path calculated by taking the current calculation force supply node as the destination node, otherwise, traversing the calculation force supply set in sequence, taking the next calculation force supply node as the destination node, and repeating the calculation flow of the shortest delay path between the demand side and the supply side until the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side.
The method of the embodiment of the specification is executed by the power dispatching equipment, and can be used for carrying out supply and demand matching on the business power demand of the power network demand side based on the power resource information summarized by the power network supply side and the network capacity information summarized by the power network transmission layer, so as to calculate and obtain a power supply node with high matching degree and the shortest delay path between the power supply node and the power demand node. The method of the embodiment of the specification can flexibly output the optimal task scheduling strategy according to the dynamic property of the nodes in the power network and the diversity of the tasks, and realizes intensive and efficient utilization of the power network resources.
Next, the steps 102 to 112 will be described in detail.
It should be noted that, the method of the embodiment of the present disclosure does not limit the execution sequence according to the number of the steps, wherein the forming process of the computing power supply set (i.e. steps 102 to 106) and the forming process of the bandwidth adjacency matrix and the delay adjacency matrix (i.e. step 108) may be performed according to the sequence of steps 102 to 108, or may be performed in parallel. After forming the computational power supply set, the bandwidth adjacency matrix, and the delay adjacency matrix, steps 110, 112 are performed.
From the viewpoint of execution efficiency, steps 102 to 106, and step 108 are generally executed in parallel regardless of the order (see fig. 3).
In step 102, a demand matrix of a demand side of the power computing network is formed based on indexes of a CPU, a GPU, a memory, and a hard disk required by a power computing demand node in the power computing network, including:
Step A11, determining CPU, GPU, memory and hard disk indexes required by the service according to the service type of the power network demand side;
and step A13, forming a demand matrix of the demand side of the computing power network based on the CPU, GPU, memory and hard disk indexes required by the service.
In step a11, when the power computing scheduling device receives the service power computing demand of the power computing demand node, the service type of the service power computing demand and the CPU, GPU, memory and hard disk index required by the service may be obtained. Under an example, a service side sends the service type and CPU, GPU, memory and hard disk indexes required by the service to a computing power scheduling device along with the service computing power requirement; in another example, different CPU, GPU, memory, and hard disk indexes required by multiple service types are stored in the power computing scheduling device in advance, and when the service side sends the service type to the power computing scheduling device, the power computing scheduling device determines the CPU, GPU, memory, and hard disk indexes required by the service under the service type according to the service type.
The CPU, GPU, memory and hard disk indexes required by the services corresponding to different service types are determined according to historical data and manual experience.
For example, according to historical experience, the amount of the required computational resources of 100 persons is A, and when the service is carried out on 10000 persons, the amount of the required computational resources is 100A according to parameter evaluation; or for a storage type service may be provided directly according to the service requirement, such as the storage space provided by the service requires the hard disk capacity C.
Step a13 forms a demand matrix r=Wherein, the method comprises the steps of, wherein,For the CPU demand vector,For the GPU demand vector,As the memory demand vector,Is a hard disk demand vector.
The CPU index may contain relevant data characterizing the CPU index, such as CPU instruction set, cache capacity, IPC, floating point computing performance, power consumption, CPU available core number, etc.
The GPU metrics may include relevant data characterizing the GPU metrics, such as GPU architecture, memory bandwidth, CUDA cores, cache capacity, floating point computing performance, power consumption, the number of cores available to the GPU, and so on.
The memory metrics may include related data characterizing the memory metrics, such as bandwidth, latency, capacity, frequency, ECC, timing, number of channels, available memory capacity, etc. of the memory.
The hard disk metrics may contain relevant data characterizing the hard disk metrics, such as type of hard disk, IOPS, latency, bandwidth, throughput, equalization, queue depth, durability, storage available capacity, etc.
In step 102, a supply matrix of a power network supply side is formed based on indexes of a CPU, a GPU, a memory, and a hard disk that can be supplied by a plurality of power supply nodes in the power network, including:
step B11, acquiring related data of a plurality of calculation power supply nodes representing CPU, GPU, memory and hard disk indexes in a calculation power network, and respectively forming a CPU supply matrix, a GPU supply matrix, a memory supply matrix and a hard disk supply matrix;
in step B13, the CPU supply matrix, the GPU supply matrix, the memory supply matrix, and the hard disk supply matrix form a supply matrix on the power network supply side.
In step B11, the resource collection engine obtains the computing power resources of the respective computing power supply nodes. Comprising
For example, relevant data of the CPU index is acquired, and a CPU supply matrix C CPU is formed based on the relevant data.
The supply matrix C CPU includes a plurality of row vectors, each of which contains data about CPU metrics of a certain power supply node.
The parameters in each row vector are in turn the CPU instruction set, cache capacity, IPC, floating point computing performance, power consumption, CPU performance, CPU capacity of a certain computing power supply node.
For example, relevant data of GPU metrics is acquired, and a GPU supply matrix C GPU is formed based on the relevant data.
Supply matrix C GPU includes a plurality of row vectors, each of which contains data about GPU metrics for a particular computational power supply node.
The parameters in each row vector are sequentially the framework, the video memory bandwidth, the CUDA core, the Cache capacity, the floating point computing performance, the power consumption, the GPU capacity and the GPU capacity of a certain computation power supply node.
For example, related data of the memory index is acquired, and a memory supply matrix C MEM is formed based on the related data.
The supply matrix C MEM includes a plurality of row vectors, each of which contains data related to the memory index of a certain power supply node.
The parameters in each row vector are in turn the bandwidth, delay, capacity, frequency, ECC, timing, number of channels, memory available capacity of the memory of a given power supply node.
For example, related data of the hard disk index is acquired, and a hard disk supply matrix C STO is formed based on the related data.
The supply matrix C STO includes a plurality of row vectors, each of which contains data related to the hard disk index of a certain computing power supply node.
The parameters in each row vector are in turn the hard disk type, IOPS, latency, bandwidth, throughput, equalization, queue depth, durability, storage availability capacity of a certain power supply node.
The related data belong to basic attributes of a CPU, a GPU, a memory and a hard disk. The resource collection engine can be utilized to directly acquire related hardware parameter information at the operating system level, or acquire the related hardware parameter information by searching parameter specification specifications of specific types of hardware corresponding to hardware manufacturers.
The CPU supply matrix, the GPU supply matrix, the memory supply matrix and the hard disk supply matrix respectively contain resource data of all computing power supply nodes in the computing power network.
In step 104, each column vector of the CPU supply matrix, the GPU supply matrix, the memory supply matrix, the hard disk supply matrix, and the demand matrix is subjected to dot multiplication, and 4 sets of column vector combinations are obtained as a result of dot multiplication, so as to obtain a comprehensive evaluation matrix N of each computing power supply node in the business scene.
N=。
The row rank of N is the maximum column rank of the four node capacity matrix, and the part of the column vector of N, which is not full of rank, is complemented by a zero vector. The row sequence number of the element with the largest row vector value is the node sequence number with the highest corresponding efficiency under the calculation power index.
Disassembling N in a row vector mode:
N mid-row vectors And the four-tuple vector is composed of indexes of the CPU, the GPU, the memory and the hard disk of the ith computing power supply node.
In step 106, the scoring the supply-demand matching degree of each computing power supply node in the evaluation matrix based on the comprehensive CPU, GPU, memory, and hard disk index includes:
Step 1061, determining index weight coefficients of a CPU, a GPU, a memory and a hard disk related to a service according to the service type of the power network demand side;
Step 1063, performing weighted average calculation on the comprehensive CPU, GPU, memory and hard disk indexes of each power supply node in the evaluation matrix by using the obtained CPU, GPU, memory and hard disk index weight coefficients, so as to obtain the supply-demand matching degree score of each power supply node.
In step 1061, when the power computing scheduling device receives the service power demand sent by the power computing demand node, the power computing scheduling device may acquire service-related CPU, GPU, memory, and hard disk index weight coefficients included in the service power demand; the pre-stored index weight coefficient can be queried according to the service types contained in the service calculation power demand, and the calculation power scheduling equipment stores various service types in advance and the index weight coefficient of the CPU, the GPU, the memory and the hard disk corresponding to the service types.
In either way, the index weight coefficient is obtained, and the index weight coefficient is determined based on historical data and artificial experience. When the traffic is a type of resource intensive traffic, the type of resource indicator weight may be greater. For example, when the service is a CPU-intensive service, the CPU index weight coefficient may be larger. For another example, when the service is a memory intensive service, the memory index weight coefficient may be larger.
In step 1063, the supply-demand matching score for each computing force supply node is calculated using a weighted average.
Si=wCPU*γCPU+wGPU*γGPU+wMem*γMem+wSto*γSto，i∈[1,k]
Si is the supply-demand matching score for the ith computational power supply node. w CPU、wGPU、wMem、wSto is the index weight coefficient of CPU, GPU, memory and hard disk respectively. And each index weight coefficient is respectively subjected to weighted average calculation with indexes of the CPU, the GPU, the memory and the hard disk of the row vector in the corresponding evaluation matrix. The sum of the index weight coefficients of the CPU, the GPU, the memory and the hard disk is 1.
And after carrying out supply-demand matching degree scoring calculation on each calculation force supply node in the evaluation matrix, arranging each calculation force supply node from high to low according to the supply-demand matching degree scoring to form a calculation force supply set. The 1 st calculation force supply node in the calculation force supply set is the calculation force supply node with the highest score, and the scores of the follow-up calculation force supply nodes are gradually reduced. The higher the score the more power supply nodes whose power characteristics match the power business needs.
In step 108, based on the bandwidth and delay indexes between the computing nodes in the computing network, a bandwidth adjacency matrix and a delay adjacency matrix of the computing network are respectively formed, including:
Step 1081, obtaining a network topology structure of the power calculation network and bandwidth and delay indexes among nodes in the power calculation network;
step 1083, forming a bandwidth adjacency matrix of the power calculation network based on the network topology structure of the power calculation network and the bandwidth index among the nodes in the power calculation network; forming a delay adjacency matrix of the power calculation network based on a network topology structure of the power calculation network and delay indexes among nodes in the power calculation network; the bandwidth index between the nodes without direct connection is 0, and the delay index between the nodes without direct connection is infinity.
In step 1081, a network probe located at a network communication node collects network transport layer data between each of the power computing nodes, including network link information, network bandwidth information, network latency information between the network communication node and the power computing supply node, between the network communication node and the network communication node, and between the network communication node and the power computing demand node, and sends the network link information, the network bandwidth information, the network latency information to the power computing scheduling device.
The computational power scheduling device forms a bandwidth adjacency matrix and a delay adjacency matrix based on the acquired data and the network topology.
Because the network interconnection relation among all the computing nodes in the computing network is a mesh graph structure, a plurality of directed adjacency matrixes are constructed by using the concept of graph theory aiming at the network communication capacity of the whole computing network, and each weight in the adjacency matrixes represents the network index of the network path so as to express the network performance of the whole network.
N Band is the bandwidth adjacency matrix:
Wherein matrix element N Band[i][j] identifies the network bandwidth in the node i to node j direction. When node i has no direct connection with node j, a bandwidth of zero is identified using 0.
N Lat is a delay adjacency matrix:
Wherein matrix element N Lat[i][j] identifies the network delay in the direction of node i to node j. When the node i is not directly connected with the node j, the infinity delay is marked by using the infinity.
In one embodiment, step 110 performs a shortest path calculation based on the delay adjacency matrix of the power network formed in step 1083. When a certain requirement is made on the bandwidth and the delay of the calculation force demand in the service calculation force demand sent by the calculation force demand node, the data among the unsuitable nodes is brought into the shortest path calculation process, the calculation efficiency is reduced, and the calculation result may not meet the requirement of the demand side on the bandwidth and the delay.
In another embodiment, the step 108 further includes:
Step 1085, setting the inter-node bandwidth index smaller than the bandwidth constraint value at the demand side in the bandwidth adjacency matrix of the computing power network to 0, setting the inter-node bandwidth index not smaller than the bandwidth constraint value at the demand side in the bandwidth adjacency matrix of the computing power network to 1, and further forming the bandwidth adjacency matrix of the computing power network after the change;
step 1087, when the bandwidth index between nodes in the bandwidth adjacent matrix of the changed computing power network is 0, setting the delay index between related nodes in the delay adjacent matrix of the computing power network to infinity, thereby forming the delay adjacent matrix of the changed computing power network; and using the time delay adjacency matrix of the changed computing power network for the subsequent shortest time delay path calculation.
In step 1085, the bandwidth adjacency matrix and the delay adjacency matrix of the computing power network formed in step 1083 are transformed, so as to mainly filter out the data between nodes which do not meet the requirements.
The bandwidth constraint value may be transmitted to the computing power scheduling device along with the computing power demand node when transmitting the traffic computing power demand. Accordingly, the network delay constraint values in step 112 are also sent to the power dispatch device along with the business power demand. Or the computing power dispatching equipment pre-stores a plurality of service types, and each service type correspondingly stores a bandwidth constraint value and a network delay constraint value. And when the power calculation scheduling equipment receives the service power calculation demands, acquiring corresponding bandwidth constraint values and network delay constraint values according to the service type query.
The bandwidth constraint value Band limit and the network delay constraint value Lat limit at the demand side are determined according to the service type at the demand side of the computing power network. Specifically, it is determined based on historical data and human experience.
The bandwidth constraint value on the demand side is the minimum constraint value. Paths in the power network with network path bandwidths less than the constraint value are not available for the power traffic due to bandwidth bottlenecks in the network. Therefore, the following changes are implemented on the original computing power network bandwidth adjacency matrix N Band to form an adjacency matrix conforming to the constraint of the computing power service bandwidth。
When the bandwidth Band i,j between the node i and the node j is smaller than the bandwidth constraint value, the network path between the node i and the node j does not accord with the cost-effective service bandwidth constraint, and the cost-effective scheduling is forbidden to be implemented through the network path, wherein=0. When the bandwidth Band i,j between the node i and the node j is larger than or equal to the bandwidth constraint value, the network path Fu Gesuan force service bandwidth constraint between the node i and the node j can implement the force scheduling through the network path, and at this time=1。
Based on the converted bandwidth adjacency matrixThe delay adjacent matrix N Lat is synchronously adjusted to be。
When the bandwidth between node i and node jWhen=1, the network path between node i and node j is available, and the network path delay parameterFor actual measured original network delayI.e. the original data is kept unchanged; when the bandwidth between node i and node jWhen=0, the network path between the node i and the node j is not available, and the network path delay parameter is corrected to ≡.
In this embodiment, not only the calculation efficiency is improved, but also a scheduling scheme meeting all the demands of the force demand nodes can be obtained when the shortest path is calculated in the subsequent step 110.
In step 110, the shortest delay path between the demand side and the supply side is calculated using the Dijkstra algorithm.
The Dijkstra algorithm is a Dijkstra algorithm. The algorithm is mainly characterized in that a greedy algorithm strategy is adopted from a starting point, and every time the algorithm traverses to adjacent nodes of the vertex which is nearest to the starting point and is not visited until the algorithm extends to a finishing point.
The shortest delay path includes a source node, several intermediate nodes, and a destination node, and is shown in path order. For example, the source node is 1 and the destination node is 20.
The shortest delay path may be: latency 1,2'→Latency2,3'→Latency3,8'. Fwdarw.
Latency 8,12'→Latency12,16'→Latency16,20'. The node sequence of the service scheduling path can be known as node 1, node 2, node 3, node 8, node 12, node 16 and node 20.
After the shortest delay path is obtained, step 112 is performed to determine, and the network delay sum among all the computing power nodes on the shortest delay path is calculated.
For example, latency Sum total =Latency1,2'+Latency2,3'+Latency3,8'
+Latency8,12'+Latency12,16'+Latency16,20’。
When the network delay sum meets the network delay constraint value (i.e., is not greater than the network delay constraint value), the shortest delay path may be output as a viable scheduling policy, and the computing power supply node as the destination node is taken as a computing power supply node with high matching degree. The power computing scheduling computing equipment distributes power computing supply nodes with high matching degree and shortest delay paths corresponding to the power computing supply nodes to the power computing demand nodes for service scheduling aiming at the service.
When steps 110 and 112 are performed, the computing power supply node serving as the destination node is selected from the computing power supply set, the computing power supply nodes are traversed in sequence according to the matching degree score, the shortest delay path is calculated for each computing power supply node, and whether the service network delay constraint Lat limit is met or not is judged. Once satisfied, outputting a high-matching-degree computing power supply node and a shortest delay path corresponding to the computing power supply node; otherwise, the next computing power supply node is obtained to perform the processes of step 110 and step 112 until the computing power supply node with high matching degree meeting the delay constraint of the service network and the shortest delay path corresponding to the computing power supply node are obtained.
When the power calculation network aims at the business power calculation requirement of a certain power calculation requirement node, the power calculation supply node with high matching degree and the shortest delay path corresponding to the power calculation supply node are distributed. And (3) carrying out the shortest-delay calculation power scheduling strategy on the business calculation power demand of the next calculation power demand node by the calculation power network according to the process (step 102-step 112, and the sequence shown in fig. 3).
And calculating a new calculation power scheduling strategy aiming at the next business calculation power demand. In forming the offer matrix, the resource information may be collected again by the resource collection engine and then the offer matrix may be reformed. The power resource adjustment may be performed according to the previously formed supply matrix, that is, the resource capacity of the corresponding power supply node after the power scheduling policy has been executed may be reduced (see fig. 3).
As shown in fig. 4, the computing power network task scheduling device in the embodiment of the present disclosure at least includes: a first module 1001, a second module 1002, a matching module 1003, a third module 1004, a calculating module 1005, and a judging module 1006.
The first module 1001 is configured to form a demand matrix of a demand side of the power computing network based on indexes of a CPU, a GPU, a memory, and a hard disk required by the power computing demand node in the power computing network, and form a supply matrix of a supply side of the power computing network based on indexes of the CPU, the GPU, the memory, and the hard disk that can be supplied by the power computing supply node in the power computing network.
The second module 1002 is configured to perform a dot product operation on the demand matrix and the supply matrix to obtain an evaluation matrix of the power calculation network; the evaluation matrix stores comprehensive CPU, GPU, memory and hard disk indexes of a plurality of calculation power supply nodes in the calculation power network;
The matching module 1003 is configured to score a supply-demand matching degree for each computing power supply node in the evaluation matrix based on the comprehensive CPU, GPU, memory, and hard disk index, and arrange each computing power supply node from high to low according to the supply-demand matching degree score, so as to form a computing power supply set;
The third module 1004 is configured to form a bandwidth adjacency matrix and a delay adjacency matrix of the power calculation network based on bandwidth and delay indexes between power calculation nodes in the power calculation network; the power calculation nodes in the power calculation network comprise power calculation supply nodes, power calculation demand nodes and network communication nodes which are positioned between the power calculation supply nodes and the power calculation demand nodes and are used for realizing communication between the power calculation supply nodes and the power calculation demand nodes;
a calculating module 1005, configured to calculate, using the computing power demand side as a source node and the first computing power supply node of the computing power supply set as a destination node, a shortest delay path between the demand side and the supply side based on a delay adjacency matrix of the computing power network using Dijkstra algorithm;
And the judging module 1006 is configured to judge whether the network delay sum of all the power supply nodes of the shortest delay path is not greater than the network delay constraint value of the demand side, if yes, output the shortest delay path calculated based on the current power supply node as the destination node, otherwise, sequentially traverse the power supply set, take the next power supply node as the destination node, trigger the calculating module to repeat the shortest delay path calculating process between the demand side and the supply side until the network delay sum of all the power supply nodes of the shortest delay path is not greater than the network delay constraint value of the demand side.
The device in the embodiment of the present disclosure is provided in an apparatus for scheduling an amount of power, and is capable of performing supply-demand matching on a service amount of power on a power network demand side based on information of an amount of power resource collected on a power network supply side and information of network capacity collected on a power network transmission layer, and calculating a power supply node with high matching degree and a shortest delay path between the power supply node and the power demand node. The device of the embodiment of the specification can flexibly output the optimal task scheduling strategy according to the dynamic property of the nodes in the power network and the diversity of the tasks, and realizes intensive and efficient utilization of the power network resources.
The first module 1001 includes:
The demand unit is used for determining CPU, GPU, memory and hard disk indexes required by the service according to the service type of the demand side of the power computing network;
the first forming unit is used for forming a demand matrix of the power calculation network demand side based on CPU, GPU, memory and hard disk indexes required by the service.
The second module 1002 includes:
The supply unit is used for acquiring the related data of the indicators of the CPU, the GPU, the memory and the hard disk represented by a plurality of calculation power supply nodes in the calculation power network, and respectively forming a CPU supply matrix, a GPU supply matrix, a memory supply matrix and a hard disk supply matrix;
The second forming unit is used for forming a supply matrix of the power calculation network supply side based on the CPU supply matrix, the GPU supply matrix, the memory supply matrix and the hard disk supply matrix.
The matching module 1003 includes:
The weight coefficient acquisition unit is used for determining the index weight coefficients of the CPU, the GPU, the memory and the hard disk related to the service according to the service type of the power calculation network demand side;
The computing unit is used for carrying out weighted average computation on the comprehensive CPU, GPU, memory and hard disk indexes of each computing power supply node in the evaluation matrix by utilizing the obtained CPU, GPU, memory and hard disk index weight coefficients to obtain supply and demand matching degree scores of each computing power supply node;
and the forming unit is used for arranging each calculation force supply node from high to low according to the supply-demand matching degree score to form a calculation force supply set.
The third module 1004 includes:
The acquisition unit is used for acquiring the network topology structure of the power calculation network and the bandwidth and delay indexes among all nodes in the power calculation network;
The matrix forming unit is used for forming a bandwidth adjacency matrix of the power calculation network based on the network topology structure of the power calculation network and bandwidth indexes among all nodes in the power calculation network; forming a delay adjacency matrix of the power calculation network based on a network topology structure of the power calculation network and delay indexes among nodes in the power calculation network; wherein, the bandwidth index between nodes without direct connection is 0, and the delay index between nodes without direct connection is infinity;
The first transformation unit is used for setting the inter-node bandwidth index of the bandwidth constraint value smaller than the bandwidth constraint value at the demand side in the bandwidth adjacent matrix of the computational power network to be 0, setting the inter-node bandwidth index of the bandwidth adjacent matrix of the computational power network not smaller than the bandwidth constraint value at the demand side to be 1, and further forming a bandwidth adjacent matrix of the varied computational power network;
the second transformation unit is used for setting the delay index between related nodes in the delay adjacent matrix of the computing power network to be infinite when the bandwidth index between the nodes in the bandwidth adjacent matrix of the computing power network after the change is 0, so as to form the delay adjacent matrix of the computing power network after the change; the time delay adjacency matrix of the changed computing power network is used for the subsequent shortest time delay path calculation.
The above-described embodiments are merely preferred embodiments of the present disclosure, and do not limit the scope of the disclosure, and various modifications and improvements made by those skilled in the art to the technical solutions of the disclosure should fall within the protection scope defined by the claims of the disclosure without departing from the design spirit of the disclosure.
Claims (10)
1. A computing power network task scheduling method, executed in a computing power scheduling device, characterized by comprising:
Forming a demand matrix of a demand side of the power calculation network based on indexes of a CPU, a GPU, a memory and a hard disk required by power calculation demand nodes in the power calculation network, and forming a supply matrix of a supply side of the power calculation network based on indexes of the CPU, the GPU, the memory and the hard disk which can be supplied by a plurality of power calculation supply nodes in the power calculation network;
performing dot product operation on the demand matrix and the supply matrix to obtain an evaluation matrix of the calculation network; the evaluation matrix stores comprehensive CPU, GPU, memory and hard disk indexes of a plurality of calculation power supply nodes in the calculation power network;
based on comprehensive CPU, GPU, memory and hard disk indexes, carrying out supply-demand matching degree scoring on each calculation power supply node in the evaluation matrix, and arranging each calculation power supply node from high to low according to the supply-demand matching degree scoring to form a calculation power supply set;
based on bandwidth and delay indexes among all computing nodes in the computing network, respectively forming a bandwidth adjacency matrix and a delay adjacency matrix of the computing network; the power calculation nodes in the power calculation network comprise power calculation supply nodes, power calculation demand nodes and network communication nodes which are positioned between the power calculation supply nodes and the power calculation demand nodes and are used for realizing communication between the power calculation supply nodes and the power calculation demand nodes;
Taking the force calculation demand side as a source node, taking the first force calculation supply node of the force calculation supply set as a destination node, and calculating the shortest delay path between the demand side and the supply side based on a delay adjacency matrix of a force calculation network by using Dijkstra algorithm;
and judging whether the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side, if so, outputting the shortest delay path calculated based on the current calculation force supply node as a destination node, otherwise, traversing the calculation force supply set in sequence, taking the next calculation force supply node as the destination node, and repeating the calculation flow of the shortest delay path between the demand side and the supply side until the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side.
2. The method of claim 1, wherein forming a demand matrix for the demand side of the power network based on the CPU, GPU, memory, hard disk metrics required by the power demand nodes in the power network comprises:
According to the service type of the power network demand side, determining the CPU, GPU, memory and hard disk indexes required by the service;
and forming a demand matrix of the demand side of the power calculation network based on the CPU, GPU, memory and hard disk indexes required by the service.
3. The method of claim 1, wherein forming a provisioning matrix for a provisioning side of the power network based on CPU, GPU, memory, hard disk metrics that can be provisioned by a number of power provisioning nodes in the power network comprises:
Acquiring related data of indexes of a plurality of computing power supply nodes representing a CPU, a GPU, a memory and a hard disk in a computing power network, and respectively forming a CPU supply matrix, a GPU supply matrix, a memory supply matrix and a hard disk supply matrix;
The CPU supply matrix, the GPU supply matrix, the memory supply matrix, and the hard disk supply matrix constitute a supply matrix on the power network supply side.
4. The method according to claim 1, wherein scoring the supply-demand matching degree of each computing power supply node in the evaluation matrix based on the comprehensive CPU, GPU, memory, and hard disk indexes comprises:
Determining index weight coefficients of a CPU, a GPU, a memory and a hard disk related to a service according to the service type of a power calculation network demand side;
And carrying out weighted average calculation on the comprehensive CPU, GPU, memory and hard disk indexes of each computing power supply node in the evaluation matrix by using the obtained CPU, GPU, memory and hard disk index weight coefficients to obtain the supply and demand matching degree score of each computing power supply node.
5. The method of claim 1, wherein forming the bandwidth adjacency matrix and the delay adjacency matrix of the power network based on the bandwidth and delay metrics between the power nodes in the power network, respectively, comprises:
acquiring a network topology structure of the power calculation network and bandwidth and delay indexes among nodes in the power calculation network;
Forming a bandwidth adjacency matrix of the computing power network based on a network topology structure of the computing power network and bandwidth indexes among nodes in the computing power network; forming a delay adjacency matrix of the power calculation network based on a network topology structure of the power calculation network and delay indexes among nodes in the power calculation network; the bandwidth index between the nodes without direct connection is 0, and the delay index between the nodes without direct connection is infinity.
6. The method of claim 5, wherein forming the bandwidth adjacency matrix and the delay adjacency matrix of the power network based on the bandwidth and delay metrics between the power nodes in the power network, respectively, further comprises:
Setting the inter-node bandwidth index smaller than the bandwidth constraint value at the demand side in the bandwidth adjacent matrix of the computing power network as 0, setting the inter-node bandwidth index not smaller than the bandwidth constraint value at the demand side in the bandwidth adjacent matrix of the computing power network as 1, and further forming the bandwidth adjacent matrix of the computing power network after change;
when the bandwidth index between nodes in the bandwidth adjacent matrix of the changed computing power network is 0, setting the delay index between related nodes in the delay adjacent matrix of the computing power network to infinity, and further forming the delay adjacent matrix of the changed computing power network;
and using the time delay adjacency matrix of the changed computing power network for the subsequent shortest time delay path calculation.
7. The method of claim 6, wherein the bandwidth constraint value and the network delay constraint value on the demand side are determined based on a traffic type on the demand side of the power network.
8. A computing power network task scheduling device provided in a computing power scheduling apparatus, comprising:
The first module is used for forming a demand matrix of a demand side of the power calculation network based on indexes of a CPU, a GPU, a memory and a hard disk required by the power calculation demand nodes in the power calculation network, and forming a supply matrix of a supply side of the power calculation network based on indexes of the CPU, the GPU, the memory and the hard disk which can be supplied by a plurality of power calculation supply nodes in the power calculation network;
the second module is used for carrying out dot multiplication operation on the demand matrix and the supply matrix to obtain an evaluation matrix of the calculation network; the evaluation matrix stores comprehensive CPU, GPU, memory and hard disk indexes of a plurality of calculation power supply nodes in the calculation power network;
The matching module is used for scoring the supply and demand matching degree of each computing power supply node in the evaluation matrix based on the comprehensive CPU, the GPU, the memory and the hard disk indexes, and arranging each computing power supply node from high to low according to the supply and demand matching degree score to form a computing power supply set;
the third module is used for respectively forming a bandwidth adjacency matrix and a delay adjacency matrix of the power calculation network based on bandwidth and delay indexes among all power calculation nodes in the power calculation network; the power calculation nodes in the power calculation network comprise power calculation supply nodes, power calculation demand nodes and network communication nodes which are positioned between the power calculation supply nodes and the power calculation demand nodes and are used for realizing communication between the power calculation supply nodes and the power calculation demand nodes;
the computing module is used for taking the computing force demand side as a source node, taking the first computing force supply node of the computing force supply set as a destination node, and computing the shortest delay path between the demand side and the supply side based on a delay adjacency matrix of a computing force network by using a Dijkstra algorithm;
And the judging module is used for judging whether the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side, if so, outputting the shortest delay path calculated based on the current calculation force supply node as a destination node, otherwise, traversing the calculation force supply set in sequence, taking the next calculation force supply node as the destination node, triggering the calculating module to repeat the shortest delay path calculating flow between the demand side and the supply side until the network delay sum of all the calculation force supply nodes of the shortest delay path is not more than the network delay constraint value of the demand side.
9. The apparatus of claim 8, wherein the matching module comprises:
The weight coefficient acquisition unit is used for determining the index weight coefficients of the CPU, the GPU, the memory and the hard disk related to the service according to the service type of the power calculation network demand side;
The computing unit is used for carrying out weighted average computation on the comprehensive CPU, GPU, memory and hard disk indexes of each computing power supply node in the evaluation matrix by utilizing the obtained CPU, GPU, memory and hard disk index weight coefficients to obtain supply and demand matching degree scores of each computing power supply node;
and the forming unit is used for arranging each calculation force supply node from high to low according to the supply-demand matching degree score to form a calculation force supply set.
10. The apparatus of claim 8, wherein the third module comprises:
The acquisition unit is used for acquiring the network topology structure of the power calculation network and the bandwidth and delay indexes among all nodes in the power calculation network;
The matrix forming unit is used for forming a bandwidth adjacency matrix of the power calculation network based on the network topology structure of the power calculation network and bandwidth indexes among all nodes in the power calculation network; forming a delay adjacency matrix of the power calculation network based on a network topology structure of the power calculation network and delay indexes among nodes in the power calculation network; wherein, the bandwidth index between nodes without direct connection is 0, and the delay index between nodes without direct connection is infinity;
The first transformation unit is used for setting the inter-node bandwidth index of the bandwidth constraint value smaller than the bandwidth constraint value at the demand side in the bandwidth adjacent matrix of the computational power network to be 0, setting the inter-node bandwidth index of the bandwidth adjacent matrix of the computational power network not smaller than the bandwidth constraint value at the demand side to be 1, and further forming a bandwidth adjacent matrix of the varied computational power network;
the second transformation unit is used for setting the delay index between related nodes in the delay adjacent matrix of the computing power network to be infinite when the bandwidth index between the nodes in the bandwidth adjacent matrix of the computing power network after the change is 0, so as to form the delay adjacent matrix of the computing power network after the change; the time delay adjacency matrix of the changed computing power network is used for the subsequent shortest time delay path calculation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410832822.XA CN118394486B (en) | 2024-06-26 | 2024-06-26 | Calculation network task scheduling method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410832822.XA CN118394486B (en) | 2024-06-26 | 2024-06-26 | Calculation network task scheduling method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN118394486A CN118394486A (en) | 2024-07-26 |
| CN118394486B true CN118394486B (en) | 2024-09-03 |
Family
ID=92006045
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410832822.XA Active CN118394486B (en) | 2024-06-26 | 2024-06-26 | Calculation network task scheduling method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118394486B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115545364A (en) * | 2021-06-29 | 2022-12-30 | 中国移动通信有限公司研究院 | Node evaluation method and device, electronic equipment and readable storage medium |
| CN116263701A (en) * | 2022-11-29 | 2023-06-16 | 中移(苏州)软件技术有限公司 | Computing power network task scheduling method and device, computer equipment and storage medium |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115730631A (en) * | 2021-08-30 | 2023-03-03 | 华为云计算技术有限公司 | Method and device for federal learning |
| CN116132403A (en) * | 2022-11-29 | 2023-05-16 | 浪潮通信技术有限公司 | Route distribution method and device of computing power network, electronic equipment and storage medium |
| CN117278566A (en) * | 2023-07-26 | 2023-12-22 | 中国移动通信集团山东有限公司 | Computing power node selection method and device, electronic equipment and readable storage medium |
| CN117395251A (en) * | 2023-09-27 | 2024-01-12 | 中国电信股份有限公司技术创新中心 | Resource scheduling method, device and computer readable storage medium |
| CN118069380B (en) * | 2024-04-25 | 2024-06-25 | 维能(深圳)大数据技术有限公司 | Computing power resource processing method |
-
2024
- 2024-06-26 CN CN202410832822.XA patent/CN118394486B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115545364A (en) * | 2021-06-29 | 2022-12-30 | 中国移动通信有限公司研究院 | Node evaluation method and device, electronic equipment and readable storage medium |
| CN116263701A (en) * | 2022-11-29 | 2023-06-16 | 中移(苏州)软件技术有限公司 | Computing power network task scheduling method and device, computer equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN118394486A (en) | 2024-07-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108566659B (en) | 5G network slice online mapping method based on reliability | |
| CN113098773B (en) | Data processing method, device and system | |
| CN112738820A (en) | Dynamic deployment method and device of service function chain and computer equipment | |
| CN102724103B (en) | Proxy server, hierarchical network system and distributed workload management method | |
| CN112148484B (en) | Coupling degree-based micro-service online distribution method and system | |
| CN111538570B (en) | Energy-saving and QoS guarantee-oriented VNF deployment method and device | |
| CN108111335B (en) | A kind of method and system of scheduling and link virtual network function | |
| CN112039965A (en) | Multitask unloading method and system in time-sensitive network | |
| CN112486690A (en) | Edge computing resource allocation method suitable for industrial Internet of things | |
| CN111385202A (en) | Route distribution method based on virtual network function | |
| Huang et al. | Enabling DNN acceleration with data and model parallelization over ubiquitous end devices | |
| CN114595049A (en) | Cloud-edge cooperative task scheduling method and device | |
| CN114071582A (en) | Service chain deployment method and device for cloud-edge collaborative Internet of things | |
| CN113094179B (en) | Job allocation method, job allocation device, electronic equipment and readable storage medium | |
| Cao et al. | A deep reinforcement learning approach to multi-component job scheduling in edge computing | |
| CN116700920A (en) | Cloud primary hybrid deployment cluster resource scheduling method and device | |
| CN115907038A (en) | Multivariate control decision-making method based on federated split learning framework | |
| CN113190342B (en) | Method and system architecture for multi-application fine-grained offloading of cloud-edge collaborative networks | |
| CN118394486B (en) | Calculation network task scheduling method and device | |
| CN109298932B (en) | OpenFlow-based resource scheduling method, scheduler and system | |
| Bensalem et al. | Towards optimal serverless function scaling in edge computing network | |
| CN118138923A (en) | Unloading method of dependent perception task in multi-core fiber elastic optical network | |
| CN117707763A (en) | Hierarchical calculation scheduling method, system, equipment and storage medium | |
| CN115499306A (en) | Method and device for constructing traffic scheduling model, electronic equipment and storage medium | |
| CN115361453A (en) | Load fair unloading and transferring method for edge service network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |