Aspects of GPU perfomance in algorithms with random memory access

Alexander V. Kashkovsky; Anton A. Shershnev; Pavel V. Vashchenkov

doi:10.1063/1.5007505

Aspects of GPU perfomance in algorithms with random memory access

Alexander V. Kashkovsky, Anton A. Shershnev, Pavel V. Vashchenkov

Лаборатория неравновесных течений ФФ

Результат исследования: Публикации в книгах, отчётах, сборниках, трудах конференций › статья в сборнике материалов конференции › научная › рецензирование

2 Цитирования (Scopus)

Аннотация

The numerical code for solving the Boltzmann equation on the hybrid computational cluster using the Direct Simulation Monte Carlo (DSMC) method showed that on Tesla K40 accelerators computational performance drops dramatically with increase of percentage of occupied GPU memory. Testing revealed that memory access time increases tens of times after certain critical percentage of memory is occupied. Moreover, it seems to be the common problem of all NVidia's GPUs arising from its architecture. Few modifications of the numerical algorithm were suggested to overcome this problem. One of them, based on the splitting the memory into "virtual" blocks, resulted in 2.5 times speed up.

Язык оригинала	английский
Название основной публикации	Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017
Подзаголовок основной публикации	Dedicated to the 60th Anniversary of the Khristianovich Institute of Theoretical and Applied Mechanics SB RAS
Редакторы	Fomin
Издатель	American Institute of Physics Inc.
Число страниц	8
Том	1893
ISBN (электронное издание)	9780735415782
DOI	https://doi.org/10.1063/1.5007505
Состояние	Опубликовано - 26 окт. 2017
Событие	25th Conference on High-Energy Processes in Condensed Matter, HEPCM 2017 - Novosibirsk, Российская Федерация Продолжительность: 5 июн. 2017 → 9 июн. 2017

Серия публикаций

Название	AIP Conference Proceedings
Издатель	AMER INST PHYSICS
Том	1893
ISSN (печатное издание)	0094-243X

Конференция

Конференция	25th Conference on High-Energy Processes in Condensed Matter, HEPCM 2017
Страна/Tерритория	Российская Федерация
Город	Novosibirsk
Период	05.06.2017 → 09.06.2017

Доступ к документу

10.1063/1.5007505

Другие файлы и ссылки

Link to publication in Scopus

Цитировать

Kashkovsky, A. V., Shershnev, A. A., & Vashchenkov, P. V. (2017). Aspects of GPU perfomance in algorithms with random memory access. В Fomin (Ред.), Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017: Dedicated to the 60th Anniversary of the Khristianovich Institute of Theoretical and Applied Mechanics SB RAS (Том 1893). [030047] (AIP Conference Proceedings; Том 1893). American Institute of Physics Inc.. https://doi.org/10.1063/1.5007505

Kashkovsky, Alexander V. ; Shershnev, Anton A. ; Vashchenkov, Pavel V. / Aspects of GPU perfomance in algorithms with random memory access. Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017: Dedicated to the 60th Anniversary of the Khristianovich Institute of Theoretical and Applied Mechanics SB RAS. редактор / Fomin. Том 1893 American Institute of Physics Inc., 2017. (AIP Conference Proceedings).

@inproceedings{638d68b47dd948b2bc7e218d322338e0,

title = "Aspects of GPU perfomance in algorithms with random memory access",

abstract = "The numerical code for solving the Boltzmann equation on the hybrid computational cluster using the Direct Simulation Monte Carlo (DSMC) method showed that on Tesla K40 accelerators computational performance drops dramatically with increase of percentage of occupied GPU memory. Testing revealed that memory access time increases tens of times after certain critical percentage of memory is occupied. Moreover, it seems to be the common problem of all NVidia's GPUs arising from its architecture. Few modifications of the numerical algorithm were suggested to overcome this problem. One of them, based on the splitting the memory into {"}virtual{"} blocks, resulted in 2.5 times speed up.",

author = "Kashkovsky, {Alexander V.} and Shershnev, {Anton A.} and Vashchenkov, {Pavel V.}",

year = "2017",

month = oct,

day = "26",

doi = "10.1063/1.5007505",

language = "English",

volume = "1893",

series = "AIP Conference Proceedings",

publisher = "American Institute of Physics Inc.",

editor = "Fomin",

booktitle = "Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017",

note = "25th Conference on High-Energy Processes in Condensed Matter, HEPCM 2017 ; Conference date: 05-06-2017 Through 09-06-2017",

}

Kashkovsky, AV, Shershnev, AA & Vashchenkov, PV 2017, Aspects of GPU perfomance in algorithms with random memory access. в Fomin (ред.), Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017: Dedicated to the 60th Anniversary of the Khristianovich Institute of Theoretical and Applied Mechanics SB RAS. том. 1893, 030047, AIP Conference Proceedings, том. 1893, American Institute of Physics Inc., 25th Conference on High-Energy Processes in Condensed Matter, HEPCM 2017, Novosibirsk, Российская Федерация, 05.06.2017. https://doi.org/10.1063/1.5007505

Aspects of GPU perfomance in algorithms with random memory access. / Kashkovsky, Alexander V.; Shershnev, Anton A.; Vashchenkov, Pavel V.

Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017: Dedicated to the 60th Anniversary of the Khristianovich Institute of Theoretical and Applied Mechanics SB RAS. ред. / Fomin. Том 1893 American Institute of Physics Inc., 2017. 030047 (AIP Conference Proceedings; Том 1893).

Результат исследования: Публикации в книгах, отчётах, сборниках, трудах конференций › статья в сборнике материалов конференции › научная › рецензирование

TY - GEN

T1 - Aspects of GPU perfomance in algorithms with random memory access

AU - Kashkovsky, Alexander V.

AU - Shershnev, Anton A.

AU - Vashchenkov, Pavel V.

PY - 2017/10/26

Y1 - 2017/10/26

N2 - The numerical code for solving the Boltzmann equation on the hybrid computational cluster using the Direct Simulation Monte Carlo (DSMC) method showed that on Tesla K40 accelerators computational performance drops dramatically with increase of percentage of occupied GPU memory. Testing revealed that memory access time increases tens of times after certain critical percentage of memory is occupied. Moreover, it seems to be the common problem of all NVidia's GPUs arising from its architecture. Few modifications of the numerical algorithm were suggested to overcome this problem. One of them, based on the splitting the memory into "virtual" blocks, resulted in 2.5 times speed up.

AB - The numerical code for solving the Boltzmann equation on the hybrid computational cluster using the Direct Simulation Monte Carlo (DSMC) method showed that on Tesla K40 accelerators computational performance drops dramatically with increase of percentage of occupied GPU memory. Testing revealed that memory access time increases tens of times after certain critical percentage of memory is occupied. Moreover, it seems to be the common problem of all NVidia's GPUs arising from its architecture. Few modifications of the numerical algorithm were suggested to overcome this problem. One of them, based on the splitting the memory into "virtual" blocks, resulted in 2.5 times speed up.

UR - http://www.scopus.com/inward/record.url?scp=85034272764&partnerID=8YFLogxK

U2 - 10.1063/1.5007505

DO - 10.1063/1.5007505

M3 - Conference contribution

AN - SCOPUS:85034272764

VL - 1893

T3 - AIP Conference Proceedings

BT - Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017

A2 - Fomin, null

PB - American Institute of Physics Inc.

T2 - 25th Conference on High-Energy Processes in Condensed Matter, HEPCM 2017

Y2 - 5 June 2017 through 9 June 2017

ER -

Kashkovsky AV, Shershnev AA, Vashchenkov PV. Aspects of GPU perfomance in algorithms with random memory access. В Fomin, редактор, Proceedings of the XXV Conference on High-Energy Processes in Condensed Matter, HEPCM 2017: Dedicated to the 60th Anniversary of the Khristianovich Institute of Theoretical and Applied Mechanics SB RAS. Том 1893. American Institute of Physics Inc. 2017. 030047. (AIP Conference Proceedings). https://doi.org/10.1063/1.5007505

Aspects of GPU perfomance in algorithms with random memory access

Аннотация

Серия публикаций

Конференция

Доступ к документу

Другие файлы и ссылки

Fingerprint

Цитировать