NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems

Carl Pearson; I. Hsin Chung; Zehra Sura; Wen Mei Hwu; Jinjun Xiong

doi:10.1007/978-3-030-02465-9_32

NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems

Carl Pearson, I. Hsin Chung, Zehra Sura, Wen Mei Hwu, Jinjun Xiong

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

High-performance computing increasingly relies on heterogeneous systems with specialized hardware accelerators to improve application performance. For example, NVIDIA’s CUDA programming system and general-purpose GPUs have emerged as a widespread accelerator in HPC systems. This trend has exacerbated challenges of data placement as accelerators often have fast local memories to fuel their computational demands, but slower interconnects to feed those memories. Crucially, real-world data-transfer performance is strongly influenced not just by the underlying hardware, but by the capabilities of the programming systems. Understanding how application performance is affected by the logical communication exposed through abstractions, as well as the underlying system topology, is crucial for developing high-performance applications and architectures. This report presents initial data-transfer microbenchmark results from two POWER-based systems obtained during work towards developing an automated system performance characterization tool.

Original language	English (US)
Title of host publication	High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers
Editors	Michèle Weiland, Rio Yokota, Sadaf Alam, John Shalf
Publisher	Springer-Verlag Berlin Heidelberg
Pages	448-454
Number of pages	7
ISBN (Print)	9783030024642
DOIs	https://doi.org/10.1007/978-3-030-02465-9_32
State	Published - Jan 1 2018
Event	International Conference on High Performance Computing, ISC High Performance 2018 - Frankfurt, Germany Duration: Jun 28 2018 → Jun 28 2018

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	11203 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	International Conference on High Performance Computing, ISC High Performance 2018
Country	Germany
City	Frankfurt
Period	6/28/18 → 6/28/18

Keywords

Benchmark
CUDA
GPGPU
NVLink
Unified Memory

ASJC Scopus subject areas

Theoretical Computer Science
Computer Science(all)

Access to Document

10.1007/978-3-030-02465-9_32

Fingerprint Dive into the research topics of 'NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems'. Together they form a unique fingerprint.

Cite this

Pearson, C., Chung, I. H., Sura, Z., Hwu, W. M., & Xiong, J. (2018). NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems. In M. Weiland, R. Yokota, S. Alam, & J. Shalf (Eds.), High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers (pp. 448-454). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11203 LNCS). Springer-Verlag Berlin Heidelberg. https://doi.org/10.1007/978-3-030-02465-9_32

NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems. / Pearson, Carl; Chung, I. Hsin; Sura, Zehra; Hwu, Wen Mei; Xiong, Jinjun.

High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers. ed. / Michèle Weiland; Rio Yokota; Sadaf Alam; John Shalf. Springer-Verlag Berlin Heidelberg, 2018. p. 448-454 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11203 LNCS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Pearson, C, Chung, IH, Sura, Z, Hwu, WM & Xiong, J 2018, NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems. in M Weiland, R Yokota, S Alam & J Shalf (eds), High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11203 LNCS, Springer-Verlag Berlin Heidelberg, pp. 448-454, International Conference on High Performance Computing, ISC High Performance 2018, Frankfurt, Germany, 6/28/18. https://doi.org/10.1007/978-3-030-02465-9_32

Pearson C, Chung IH, Sura Z, Hwu WM, Xiong J. NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems. In Weiland M, Yokota R, Alam S, Shalf J, editors, High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers. Springer-Verlag Berlin Heidelberg. 2018. p. 448-454. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-030-02465-9_32

Pearson, Carl ; Chung, I. Hsin ; Sura, Zehra ; Hwu, Wen Mei ; Xiong, Jinjun. / NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems. High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers. editor / Michèle Weiland ; Rio Yokota ; Sadaf Alam ; John Shalf. Springer-Verlag Berlin Heidelberg, 2018. pp. 448-454 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{d8c924d2a9b0464bbddcc14542f5e6e0,

title = "NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems",

abstract = "High-performance computing increasingly relies on heterogeneous systems with specialized hardware accelerators to improve application performance. For example, NVIDIA{\textquoteright}s CUDA programming system and general-purpose GPUs have emerged as a widespread accelerator in HPC systems. This trend has exacerbated challenges of data placement as accelerators often have fast local memories to fuel their computational demands, but slower interconnects to feed those memories. Crucially, real-world data-transfer performance is strongly influenced not just by the underlying hardware, but by the capabilities of the programming systems. Understanding how application performance is affected by the logical communication exposed through abstractions, as well as the underlying system topology, is crucial for developing high-performance applications and architectures. This report presents initial data-transfer microbenchmark results from two POWER-based systems obtained during work towards developing an automated system performance characterization tool.",

keywords = "Benchmark, CUDA, GPGPU, NVLink, Unified Memory",

author = "Carl Pearson and Chung, {I. Hsin} and Zehra Sura and Hwu, {Wen Mei} and Jinjun Xiong",

year = "2018",

month = jan,

day = "1",

doi = "10.1007/978-3-030-02465-9_32",

language = "English (US)",

isbn = "9783030024642",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer-Verlag Berlin Heidelberg",

pages = "448--454",

editor = "Mich{\`e}le Weiland and Rio Yokota and Sadaf Alam and John Shalf",

booktitle = "High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers",

note = "International Conference on High Performance Computing, ISC High Performance 2018 ; Conference date: 28-06-2018 Through 28-06-2018",

}

TY - GEN

T1 - NUMA-Aware Data-Transfer Measurements for Power/NVLink Multi-GPU Systems

AU - Pearson, Carl

AU - Chung, I. Hsin

AU - Sura, Zehra

AU - Hwu, Wen Mei

AU - Xiong, Jinjun

PY - 2018/1/1

Y1 - 2018/1/1

N2 - High-performance computing increasingly relies on heterogeneous systems with specialized hardware accelerators to improve application performance. For example, NVIDIA’s CUDA programming system and general-purpose GPUs have emerged as a widespread accelerator in HPC systems. This trend has exacerbated challenges of data placement as accelerators often have fast local memories to fuel their computational demands, but slower interconnects to feed those memories. Crucially, real-world data-transfer performance is strongly influenced not just by the underlying hardware, but by the capabilities of the programming systems. Understanding how application performance is affected by the logical communication exposed through abstractions, as well as the underlying system topology, is crucial for developing high-performance applications and architectures. This report presents initial data-transfer microbenchmark results from two POWER-based systems obtained during work towards developing an automated system performance characterization tool.

AB - High-performance computing increasingly relies on heterogeneous systems with specialized hardware accelerators to improve application performance. For example, NVIDIA’s CUDA programming system and general-purpose GPUs have emerged as a widespread accelerator in HPC systems. This trend has exacerbated challenges of data placement as accelerators often have fast local memories to fuel their computational demands, but slower interconnects to feed those memories. Crucially, real-world data-transfer performance is strongly influenced not just by the underlying hardware, but by the capabilities of the programming systems. Understanding how application performance is affected by the logical communication exposed through abstractions, as well as the underlying system topology, is crucial for developing high-performance applications and architectures. This report presents initial data-transfer microbenchmark results from two POWER-based systems obtained during work towards developing an automated system performance characterization tool.

KW - Benchmark

KW - CUDA

KW - GPGPU

KW - NVLink

KW - Unified Memory

UR - http://www.scopus.com/inward/record.url?scp=85066126248&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85066126248&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-02465-9_32

DO - 10.1007/978-3-030-02465-9_32

M3 - Conference contribution

AN - SCOPUS:85066126248

SN - 9783030024642

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 448

EP - 454

BT - High Performance Computing - ISC High Performance 2018 International Workshops, Revised Selected Papers

A2 - Weiland, Michèle

A2 - Yokota, Rio

A2 - Alam, Sadaf

A2 - Shalf, John

PB - Springer-Verlag Berlin Heidelberg

T2 - International Conference on High Performance Computing, ISC High Performance 2018

Y2 - 28 June 2018 through 28 June 2018

ER -