Program transformation and runtime support for threaded MPI execution on shared-memory machines

Abstract

Parallel programs written in MPI have been widely used for developing high-performance applications on various platforms. Because of a restriction of the MPI computation model, conventional MPI implementations on shared memory machines map each MPI node to an OS process, which can suffer serious performance degradation in the presence of multiprogramming. This paper studies compile-time and runtime techniques for enhancing performance portability of MPI code running on multiprogrammed shared memory machines. The proposed techniques allow MPI nodes to be executed safely and efficiently as threads. Compile-time transformation eliminates global and static variables in C code using node-specific data. The runtime support includes an efficient and provablycorrect communication protocol that uses lock-free data structure and takes advantage of address space sharing among threads. The experiments on SGI Origin 2000 show that our MPI prototype called TMPI using the proposed techniques is competitive with SGI's native MPI implementation in a dedicated environment, and that it has significant performance advantages in a multiprogrammed environment.

Leveraging MPI’s One-Sided Communication Interface for Shared-Memory Programming

“…Previous work on efficiently supporting MPI on shared-memory systems has concentrated mostly on mapping an MPI rank to a system-level or user-level thread [4][5][6][7]. This approach allows MPI ranks to share memory inside an operating system process, but it requires program transformation or knowledge on the part of the programmer to handle global and static variables appropriately.…”

Automatic MPI to AMPI Program Transformation Using Photran

“…TMPI [16] uses multithreading for performance enhancement of multi-threaded MPI programs on shared memory machines. More recent work in FG-MPI [9] shares the same idea with AMPI by exploiting fine grained decomposition using threads.…”

Automatic Handling of Global Variables for Multi-threaded MPI Programs

“…Adaptive MPI (AMPI) [4], [5], FG-MPI [6], Phoenix [7] and TMPI (Threaded MPI) [8] exemplify this approach. One advantage of this approach is that it allows automatic adaptive overlap of communication and computationwhen one MPI thread is blocked to receive a message, another MPI thread on the same processor can be scheduled to be executed.…”

“…TMPI [8] uses multithreading for performance enhancement of multi-threaded MPI programs on shared-memory machines. Because it targets a shared-memory scenario, its main goal in implementing MPI tasks via threads is to provide a common address space to the threads such that memory copy in message-passing is avoided.…”