Performance Of Collective Communication Operations Research Articles

Model-based selection of optimal MPI broadcast algorithms for multi-core clusters

The performance of collective communication operations determines the overall performance of MPI applications. Different algorithms have been developed and implemented for each MPI collective operation, but none proved superior in all situations. Therefore, MPI implementations have to solve the problem of selecting the optimal algorithm for the collective operation depending on the platform, the number of processes involved, the message size(s), etc. The current solution method is purely empirical.Recently, an alternative solution method using analytical performance models of collective algorithms has been proposed and proved both accurate and efficient for one-process-per-CPU configurations. The method derives the analytical performance models of algorithms from their code implementation rather than from high-level mathematical definitions, and estimates the parameters of the models separately for each algorithm. The method is network and topology oblivious and uses the Hockney model for point-to-point communications. In this paper, we extend that selection method to the case of clusters of multi-core processors, where each core of the platform runs a process of the MPI application.We present the proposed approach using Open MPI broadcast algorithms, and experimentally validate it on three different clusters of multi-core processors, Grisou, Gros and MareNostrum4.

Network Performance Aware MPI Collective Communication Operations in the Cloud

This paper examines the performance of collective communication operations in message passing interfaces (MPI) in the cloud computing environment. The awareness of network topology has been a key factor in performance optimizations for existing MPI implementations. However, virtualization in the cloud environment not only hides the network topology information from the users, but also causes traffic interference and dynamics to network performance. Existing topology-aware optimizations are no longer feasible in the cloud environment. Therefore, we develop novel network performance aware algorithms for a series of collective communication operations including broadcast, reduce, gather and scatter. We further implement two common applications, N-body and conjugate gradient (CG). We have conducted our experiments with two complementary methods (on Amazon EC2 and simulations). Our experimental results show that the network performance awareness results in 25.4 and 28.3 percent performance improvement over MPICH2 on Amazon EC2 and on simulations, respectively. Evaluations on N-body and CG show 41.6 and 14.3 percent respectively on application performance improvement.

Optimization of Collective Communication Operations in MPICH

We describe our work on improving the performance of collective communication operations in MPICH for clusters connected by switched networks. For each collective operation, we use multiple algorithms depending on the message size, with the goal of minimizing latency for short messages and minimizing bandwidth use for long messages. Although we have implemented new algorithms for all MPI (Message Passing Interface) collective operations, because of limited space we describe only the algorithms for allgather, broadcast, all-to-all, reduce-scatter, reduce, and allreduce. Performance results on a Myrinet-connected Linux cluster and an IBM SP indicate that, in all cases, the new algorithms significantly outperform the old algorithms used in MPICH on the Myrinet cluster, and, in many cases, they outperform the algorithms used in IBM's MPI on the SP. We also explore in further detail the optimization of two of the most commonly used collective operations, allreduce and reduce, particularly for long messages and nonpower-of-two numbers of processes. The optimized algorithms for these operations perform several times better than the native algorithms on a Myrinet cluster, IBM SP, and Cray T3E. Our results indicate that to achieve the best performance for a collective communication operation, one needs to use a number of different algorithms and select the right algorithm for a particular message size and number of processes.