A compiler abstraction for machine independent parallel communication generation

Abstract

In this paper, we consider the problem of generating effident, portable communication in compilers for parallel languages. We introduce the IRONMAN abstraction, which separates data transfer from its implementing communication paradigm. This is done by annotating the compiler-generated code with legal ranges for data transfer in the form of calls to the IRONMAN library. On each target platform, these library calls are instantiated to perform the transfer using the machine's optimal communication paradigm. We confirm arguments against generating message passing calls in the compiler based on our experiences using PVM and MPI --specifically, the observation that these interfaces do not perform well on machines that are not built with a message passing communication paradigm. The overhead for using IRONMAN, as opposed to a machine-specific back end, is demonstrated to be negligible. We give performance results for a number of benchmarks running with PVM, MPI, and machine-specific implementations of the IRONMAN abstraction, yielding performance improvements of up to 42% of communication time and 1-14% of total computation time.

An advanced compiler framework for non-cache-coherent multiprocessors

“…In fact, the more distributed memory machines have supported Put/Get in hardware, the more attention compiler researchers have paid to code optimizations in the context of shared-memory models. Although there has been some work on optimizations for NCC machines, most of the work did not handle extensive compiler issues, only focusing on a couple of individual optimization issues, such as reducing synchronization overhead in Put/Get-based parallel codes [15] or building a new communication library that is to be used for an array language [8].…”

“…For instance, some experimental studies [3], [13], [25] with microbenchmarking indicated that one-sided implementations are likely more efficient on NCC machines than two-sided implementations because they utilize the architectural features of the machines more efficaciously. 1 In addition, other studies [8], [15] concluded that 1) Put/Get is inherently less expensive than Send/ Receive for just data block copy and 2) their one-way communication protocol simplifies the compiler techniques for certain classes of computations, such as irregular computations and pointer chasing, because the problems of message marshaling, synchronization, and buffering are eliminated.…”