Unified memory in HOOMD-blue improves node-level strong scaling

Jens Glaser; Peter S. Schwendeman; Joshua A. Anderson; Sharon C. Glotzer

doi:10.1016/j.commatsci.2019.109359

Unified memory in HOOMD-blue improves node-level strong scaling

Jens Glaser, Peter S. Schwendeman, Joshua A. Anderson, Sharon C. Glotzer

Research output: Contribution to journal › Article › peer-review

6 Scopus citations

Abstract

Current supercomputer designs rely on increasing the compute density inside a node to maximize the performance of applications that tightly integrate the processors within a shared memory space. HOOMD-blue 2.5 enables molecular dynamics simulations that take advantage of multiple GPUs inside the same node which are connected via NVLINK. We describe the native implementation of CUDA unified memory in HOOMD-blue for strong scaling on this hardware, and provide performance benchmarks.

Original language	English
Article number	109359
Journal	Computational Materials Science
Volume	173
DOIs	https://doi.org/10.1016/j.commatsci.2019.109359
State	Published - Feb 15 2020
Externally published	Yes

Funding

This material is based upon work supported in part by the U.S. Army Research Laboratory and the U. S. Army Research Office under Grant No. W911NF-18-1-0167 (J.G.) and by the National Science Foundation, Division of Materials Research Award #DMR 1409620 (J.A. and S.C.G). This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725: INCITE Project MAT110 “Nucleation and growth of colloidal crystals,” and also an Early Science Project on OLCF's Summit supercomputer. Additional hardware support by NVIDIA Corp. to the Glotzer Group is gratefully acknowledged. This material is based upon work supported in part by the U.S. Army Research Laboratory and the U. S. Army Research Office under Grant No. W911NF-18-1-0167 (J.G.) and by the National Science Foundation , Division of Materials Research Award #DMR 1409620 (J.A. and S.C.G). This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725: INCITE Project MAT110 “Nucleation and growth of colloidal crystals,” and also an Early Science Project on OLCF’s Summit supercomputer. Additional hardware support by NVIDIA Corp. to the Glotzer Group is gratefully acknowledged.

Keywords

CUDA
GPUs
Molecular dynamics
NVLINK
Rigid bodies
Unified memory

Access to Document

10.1016/j.commatsci.2019.109359

Cite this

@article{75bb3f49e1284106af7dda9e946c6f40,

title = "Unified memory in HOOMD-blue improves node-level strong scaling",

abstract = "Current supercomputer designs rely on increasing the compute density inside a node to maximize the performance of applications that tightly integrate the processors within a shared memory space. HOOMD-blue 2.5 enables molecular dynamics simulations that take advantage of multiple GPUs inside the same node which are connected via NVLINK. We describe the native implementation of CUDA unified memory in HOOMD-blue for strong scaling on this hardware, and provide performance benchmarks.",

keywords = "CUDA, GPUs, Molecular dynamics, NVLINK, Rigid bodies, Unified memory",

author = "Jens Glaser and Schwendeman, {Peter S.} and Anderson, {Joshua A.} and Glotzer, {Sharon C.}",

note = "Publisher Copyright: {\textcopyright} 2019 Elsevier B.V.",

year = "2020",

month = feb,

day = "15",

doi = "10.1016/j.commatsci.2019.109359",

language = "English",

volume = "173",

journal = "Computational Materials Science",

issn = "0927-0256",

publisher = "Elsevier",

}

TY - JOUR

T1 - Unified memory in HOOMD-blue improves node-level strong scaling

AU - Glaser, Jens

AU - Schwendeman, Peter S.

AU - Anderson, Joshua A.

AU - Glotzer, Sharon C.

PY - 2020/2/15

Y1 - 2020/2/15

N2 - Current supercomputer designs rely on increasing the compute density inside a node to maximize the performance of applications that tightly integrate the processors within a shared memory space. HOOMD-blue 2.5 enables molecular dynamics simulations that take advantage of multiple GPUs inside the same node which are connected via NVLINK. We describe the native implementation of CUDA unified memory in HOOMD-blue for strong scaling on this hardware, and provide performance benchmarks.

AB - Current supercomputer designs rely on increasing the compute density inside a node to maximize the performance of applications that tightly integrate the processors within a shared memory space. HOOMD-blue 2.5 enables molecular dynamics simulations that take advantage of multiple GPUs inside the same node which are connected via NVLINK. We describe the native implementation of CUDA unified memory in HOOMD-blue for strong scaling on this hardware, and provide performance benchmarks.

KW - CUDA

KW - GPUs

KW - Molecular dynamics

KW - NVLINK

KW - Rigid bodies

KW - Unified memory

UR - http://www.scopus.com/inward/record.url?scp=85075372951&partnerID=8YFLogxK

U2 - 10.1016/j.commatsci.2019.109359

DO - 10.1016/j.commatsci.2019.109359

M3 - Article

AN - SCOPUS:85075372951

SN - 0927-0256

VL - 173

JO - Computational Materials Science

JF - Computational Materials Science

M1 - 109359

ER -

Unified memory in HOOMD-blue improves node-level strong scaling

Abstract

Funding

Keywords

Access to Document

Other files and links

Fingerprint

Cite this