Communication optimizations for parallel C programs

Abstract

This paper presents algorithms for reducing the communication overhead for parallel C programs that use dynamically allocated data structures. The framework consists of an analysis phase called possible-placement analysis, and a transformation phase called communication selection. The fundamental idea of possible-placement analysis is to find all possible points for insertion of remote memory operations. Remote reads are propagated upwards, whereas remote writes are propagated downwards. Based on the results of the possibleplacement analysis, the communication selection transformation selects the``best'' place for inserting the communication and determines if pipelining or blocking of communication should be performed. The framework has been implemented in the EARTH-McCAT optimizing C compiler, and experimental results are presented for five pointer-intensive benchmarks running on the EARTH-MANNA distributed-memory parallel processor. These experiments show that the communication optimization can provide performance improvements of up to 160 over the unoptimized benchmarks. Academic Press

SyncChecker: Detecting Synchronization Errors between MPI Applications and Libraries

“…One popular method for optimizing MPI applications is to overlap the computation of MPI applications and the communication of underlying MPI libraries since it can hide costly communication latency [4], [5], [6], [7]. Such overlapping can be achieved by invoking nonblocking MPI functions or RDMA (Remote Direct Memory Access) functions.…”

Evaluating support for global address space languages on the Cray X1

“…The straightforward approach to applying this transformation is to convert one-sided blocking get/put operations into an initiation call and a corresponding synchronization call, then perform code motion to separate the two as far as possible while inserting independent computation or communication code in between. Several studies [30,17,6] have proposed global communication scheduling techniques that attempt to find an optimal arrangement for all non-blocking memory accesses. Other variants of this optimization such as message strip mining [27] and software prefetching [18] are also useful in reducing an application's stall times due to communication latencies.…”

A performance analysis of the Berkeley UPC compiler

“…Several studies ( [5], [22]) present compiler algorithms that perform possible-placement analysis, both on basic blocks and whole programs. Code motion of communication operations in UPC needs to be supplemented with an analysis to ensure that the new schedule of operations does not violate the memory consistency model of the language.…”