基于算力网络的异构算力请求路由算法  

Routing algorithm for heterogeneous computing force requests based on computing first network

作　　者：章刚 黎曦 ZHANG Gang;LI Xi(School of Information and Artificial Intelligence,Nanchang Institute of Science and Technology,Nanchang 330108,China;Nanchang Key Laboratory of Internet of Things Information Visualization Technology,Nanchang Institute of Science and Technology,Nanchang 330108,China)

机构地区：[1]南昌工学院信息与人工智能学院,江西南昌330108 [2]南昌工学院南昌市物联网信息可视化技术重点实验室,江西南昌330108

出　　处：《电信科学》2025年第2期95-110,共16页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.61572325);江西省教育厅科技研究项目(No.GJJ212510);南昌工学院人才引进项目(No.NGRCZX-21-07)。

摘　　要：由于算力请求具有特殊性和独特性,如何为一组异构算力请求寻找传输链路互不相交的有效路径集,使得该组请求能够到达各自目的算力节点,从而实现为该组请求分配算力资源,是当前算力网络面临的关键问题。首先,对异构算力请求的路由问题进行剖析,并通过建模将其转化为非确定性多项式(nondeterministic polynomial,NP)完全问题。针对该问题,提出一种优化型遗传算法。该算法从局部和全局两个层面进行设计:在局部层面,为保证快速收敛到目标解,采用单参数满足随机性策略初始化种群,使得种群广泛地分散在解空间中;采用多参数解(或路径)均衡选择策略进行选择操作,使得被选种群丰富多样;采用两层交叉策略进行交叉操作,目的是拓宽全域搜索广度;采用多参数随机单点变异策略进行变异操作,目的是深挖局域搜索能力。在全局层面,为保证路径不冲突,采用路径分发策略,通过构建求解矩阵,并借助评价函数、可行解随机选择、低需求优先避让原则等方法,确保最终找到一组可行解集。实验从异构算力请求的传输成功率、算法收敛延时比、算力网络负载均衡误差率等3个方面进行验证,相较于IGAGCT算法与RBDQN算法,该算法在传输成功率、算法收敛延时比和负载均衡方面分别平均优化了8.85%、15.51%、17.03%及10.41%、16.5%、16.81%。Due to the particularity and uniqueness of computing force requests,how to find an effective path set with non-intersecting transmission links for a group of heterogeneous computing force requests,so that the group of requests can reach their respective destination computing force nodes,and thus allocate computing force resources for the group of requests,is a key issue facing current computing first networks.Firstly,the routing problem of heterogeneous computing force requests was analyzed and was transformed into an nondeterministic polynomial(NP)-complete problem through modeling.An optimized genetic algorithm to address this issue was proposed.This algorithm was designed from both local and global perspectives:to ensure fast convergence to the target solution locally,a single parameter satisfying the randomness strategy was used to initialize the population,making it widely dispersed in the solution space;adopting a multi-parameter solution(or path)balanced selection strategy for selection operations,making the selected population rich and diverse;adopting a two-layer crossover strategy for crossover operations,with the aim of expanding the breadth of global search;adopting a multi parameter random single point mutation strategy for mutation operations,with the aim of deepening local search capabilities.To ensure that the paths did not conflict globally,a path distribution strategy was adopted.By constructing a solution matrix and utilizing evaluation functions,random selection of feasible solutions,and the principle of low demand priority avoidance,a set of feasible solution sets was ultimately found.The experiment verifies that algorithm has been optimized by an average of 8.85%,15.51%,and 17.03%in terms of transmission success rate,convergence delay ratio,and load balancing compared to the IGAGCT algorithm and RBDQN algorithm,and 10.41%,16.5%,and 16.81%,respectively from three aspects:heterogeneous request success rate,algorithms convergence delay rate,and load error rate of computing first networks.

关 键 词：算力网络 异构算力请求 遗传算法 算力路由 NP完全问题 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]