Research of Collectives Optimization on Modern Multicore Clusters

Abstract

With the rapid increase in HPC computing requirement,the multicore is widely deployed in HPC systems.To keep the efficiency of application in large scale systems,it is very important keep up the balance of communication to computation,thus multicore brings more requirement for communication systems.As collective communication is an important part in communication systems and is critical for the whole systems,thus it is important to research on the impacts of multicore environment on collective performance.This paper first analyzes how multicore impacts on collective communication.It is found out that multicore SMP clusters brings two conflict impacts,it not only has faster intra-socket communication path which can speed up the performance of collective communications,but also it brings memory/cache contention which might degrade the communication performance.Based on these aspects,this paper proposes multicore-aware collecitves optimization techniques,which includes:hierarchy-aware algorithms,limited-concurrency,NUMA-aware algorithms and cache-friendly optimization.These optimization methods are implemented in MPI_Barrier,MPI_Bcast and MPI_Alltoall.Experiments show that the proposed algorithms increase the performance and scalability of collective communication.