Reasearch And Implementation Of Clock Synchronization And Rate Adaptation Mechanism On HPC Interconnect Simulation Environment

Development of next-generation supercomputer as a national key development strategy,relevant universities and research institutes in supercomputer research and development have done a lot of research.However,as a huge computer system,the supercomputer has huge research and development costs and a long development cycle.Therefore,it is necessary to build a system-level simulation environment that runs parallel computing benchmark programs.The simulation environment simulates the hardware and software environment of a supercomputer,provides a test environment and evaluates device availability and related performance for the developed supercomputer-related devices(such as high-performance network card devices)in the early stages of development.Simulation environment is an efficient approach to lower the development costs and shorten development cycles for supercomputer.In this paper,we propose the construction scheme of HPC interconnect simulation environment on the basis of extensive research.In addition,we conduct an in-depth study of clock synchonization issues in simulation environment.To solve the problem of Network Time Protocol,we propose a dynamic NTP clock synchronization algorithm and manage to realize the clock synchronization of simulation environment.Aiming at the problem that CPU computing speed and network speed unmatch in the simulation environment,we propose using CGroup adjustment processes running speed on CPU and using OVS(Open vSwicth)adjustment network speed,which solves the problem of speed unmatch.Finally,we build a simulation environment of High Performace Computing which can support running parallel computing benchmark program.Using the dynamic NTP clock synchronization algorithm to achieve the simulation environment clock synchronization,experimental results show that the dynamic NTP clock synchronization algorithm can reach the accuracy of synchronization of node clock to the level of sub-millisecond.At the same time,it can effectively reduce the number of clock synchronization messages in the network.We use OVS and CGroup to adjust the network speed and CPU speed,and give the match range of CPU speed and network speed through experimental tests,which provides a reference for the speed match.