A multithreaded communication engine for multicore architectures

Abstract

The current trend in clusters leads towards an increase of the number of cores per node. As a result, an increasing number of parallel applications is mixing message passing and multithreading as an attempt to better match the underlying architecture's structure. This naturally raises the problem of designing efficient, multithreaded implementations of MPI. In this paper, we present the design of a multithreaded communication engine able to exploit idle cores to speed up communications in two ways: it can move CPUintensive operations out of the critical path (e.g. PIO transfers offload), and is able to let rendezvous transfers progress asynchronously. We have implemented these methods in the PM2 software suite, evaluated their behavior in typical cases, and we have observed good performance results in overlapping communication and computation. Classical non-multithreaded communication engines such as OpenMPI [3], MPICH2 [1] or MVAPICH2 [4] do not fully take benefit from multicore architectures: com

A scalable and generic task scheduling system for communication libraries

“…In a previous paper [2] we showed that this permits to overlap communication and computation even for small messages. The submission of requests to a network works as follows: the state of each core is evaluated in order to find an idle core that could process the task.…”

“…These synchronization issues ofte leave holes in thread scheduling. We showed [2] that it is possible to exploit these holes to make the communication library progress. Rendezvous handshake or packet submission to the NICs may be offloaded to idle cores so as to perform these operations in the background and thus to allow communication and computation to overlap.…”

“…Moreover, we have shown that the communication library itself may benefit from multi-threading [2], for example to make non-blocking communication actually progress in background and overlap with computation.…”

GePSeA: A General-Purpose Software Acceleration Framework for Lightweight Task Offloading

“…In [23], the authors present a communication engine to exploit the cores in multi-core systems using various multi-threading techniques. The authors plan to integrate their engine with MPICH2 in the future.…”

Ocean-Atmosphere Application Scheduling within DIET

“…If there are 2 scenarios executing on the same node, the application behaves abnormally and both scenarios crash. The MPI used by Ocean-Atmosphere is OpenMPI, but it is going to be replaced by MadMPI 3 [15]. When the replacement of MPI will be done the problem might be solved, but we are currently investigating this problem because it can occur quite often.…”