A Study on the Performance of Reproducible Computations

被引：0

作者：

Bombace, Nico ^{[1
]}

Weiland, Michele ^{[1
]}

机构：

[1] Univ Edinburgh, Bayes Ctr, EPCC, Edinburgh, Midlothian, Scotland

来源：

HIGH PERFORMANCE COMPUTING: ISC HIGH PERFORMANCE 2019 INTERNATIONAL WORKSHOPS | 2020年 / 11887卷

基金：

英国工程与自然科学研究理事会;

关键词：

Reproducibility; Binned doubles; Performance;

D O I：

10.1007/978-3-030-34356-9_33

中图分类号：

TP301 [理论、方法];

学科分类号：

081202 ;

摘要：

Parallel computations are intrinsically non-reproducible, due to a combined effect of non-deterministic parallel reductions and non-associative floating point operations. Different strategies have been proposed in literature to alleviate this issue or eliminate it altogether, however at present there is no study on the performance impact of associative floating point operations on large scale applications. In this work, we implement associative operations using binned doubles in MiniFE, and perform various performance tests on Cirrus and Fulhame, two state-of-the-art HPC systems.

引用

页码：441 / 451

页数：11

共 6 条

[1] Ahrens P., 2015, ACM Trans. Math. Softw, DOI DOI 10.1145/3389360
[2] [Anonymous], 2001, OPENMP APPL PROGR IN
[3] Heroux M.A., 2009, Improving performance via mini-applications
[4] FURTHER REMARKS ON REDUCING TRUNCATION ERRORS
KAHAN, W
[J]. COMMUNICATIONS OF THE ACM, 1965, 8 (01) : 40 - &
[5] In search of numerical consistency in parallel programming
Robey, Robert W.
Robey, Jonathan M.
Aulwes, Rob
[J]. PARALLEL COMPUTING, 2011, 37 (4-5) : 217 - 229
[6] Vandevoorde David., 2002, C TEMPLATES

← 1 →