DMCA
An Adaptive Runtime Library for OpenMP on (2005)
BibTeX
@MISC{Smps05anadaptive,
author = {Hyperthreaded Smps and Yun Zhang and Mihai Burcea and Victor Cheng and Ron Ho and Michael Voss},
title = {An Adaptive Runtime Library for OpenMP on},
year = {2005}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
Hyperthreaded (HT) and simultaneous multithreaded (SMT) processors are now available in commodity workstations and servers. This technology is designed to increase throughput by executing multiple concurrent threads on a single physical processor. These multiple threads share the processor’s functional units and on-chip memory hierarchy in an attempt to make better use of idle resources. Most OpenMP applications and runtime libraries have been written assuming an Symmetric Multiprocessor (SMP), not an SMT, model. Threads executing on the same physical SMT processor have interactions on data locality and resource sharing that do not occur on traditional SMPs. This work focuses on tuning the behavior of an OpenMP runtime library to improve the performance of applications executing on SMPs with SMT processors. First we implement and evaluate three techniques that have been asserted as important on HT processors: the use of pause instructions in spin loops, incorporation of per-thread stack offsets in slave threads, and lock padding. After implementing these techniques in the Omni OpenMP runtime library, we discover that both pause instructiosn and lock padding show noticible improvements, while stack offsets show negligible gains. We then propose three adaptive loop schedulers: a region-based scheduler, a loop-based scheduler and hardware counter based scheduler. These novel schedulers determine effective hierarchical schedules for individual parallel regions or loops to exploit the unique nature of SMPs with HT processors. We compare the performance of our three proposed schedulers against several standard schedulers using the SPEC and NAS OpenMP benchmark suites. We show that schedule selection must be done at the granularity of loops not regions, and that our loop-based adaptive schedulers outperform all other approaches on average, increasing speedup on average by over 25 % when all thread contexts are used. I.
Keyphrases
adaptive runtime library ht processor commodity workstation region-based scheduler effective hierarchical schedule slave thread physical smt processor traditional smps multiple thread share openmp runtime library data locality stack offset single physical processor individual parallel region negligible gain symmetric multiprocessor loop-based adaptive scheduler loop-based scheduler several standard scheduler on-chip memory hierarchy schedule selection multiple concurrent thread adaptive loop scheduler spin loop omni openmp runtime library lock padding show noticible improvement novel scheduler resource sharing pause instructiosn hardware counter lock padding pause instruction runtime library openmp application processor functional unit smt processor unique nature idle resource per-thread stack offset na openmp benchmark suite
