Streaming nested data parallelism on multicores

Frederik Meisner Madsen; Andrzej Filinski

doi:10.1145/2975991.2975998

Streaming nested data parallelism on multicores

Frederik Meisner Madsen, Andrzej Filinski

Datalogisk Institut

1 Citationer (Scopus)

Abstract

The paradigm of nested data parallelism (NDP) allows a variety of semi-regular computation tasks to be mapped onto SIMD-style hardware, including GPUs and vector units. However, some care is needed to keep down space consumption in situations where the available parallelism may vastly exceed the available computation resources. To allow for an accurate space-cost model in such cases, we have previously proposed the Streaming NESL language, a refinement of NESL with a high-level notion of streamable sequences.

In this paper, we report on experience with a prototype implementation of Streaming NESL on a 2-level parallel platform, namely a multicore system in which we also aggressively utilize vector instructions on each core. We show that for several examples of simple, but not trivially parallelizable, text-processing tasks, we obtain single-core performance on par with off-the-shelf GNU Coreutils code, and near-linear speedups for multiple cores.

Originalsprog	Engelsk
Titel	Proceedings of the 5th International Workshop on Functional High-Performance Computing
Antal sider	8
Forlag	Association for Computing Machinery
Publikationsdato	8 sep. 2016
Sider	44-51
ISBN (Elektronisk)	978-1-4503-4433-3
DOI	https://doi.org/10.1145/2975991.2975998
Status	Udgivet - 8 sep. 2016
Begivenhed	International Workshop on Functional High-Performance Computing - Nara, Japan Varighed: 22 sep. 2016 → 22 sep. 2016 Konferencens nummer: 5 https://sites.google.com/site/fhpcworkshops/

Workshop

Workshop	International Workshop on Functional High-Performance Computing
Nummer	5
Land/Område	Japan
By	Nara
Periode	22/09/2016 → 22/09/2016
Internetadresse	https://sites.google.com/site/fhpcworkshops/

Adgang til dokumentet

10.1145/2975991.2975998

Citationsformater

@inproceedings{ae53687e6c3045c58ac0d23e195b37dd,

title = "Streaming nested data parallelism on multicores",

abstract = "The paradigm of nested data parallelism (NDP) allows a variety of semi-regular computation tasks to be mapped onto SIMD-style hardware, including GPUs and vector units. However, some care is needed to keep down space consumption in situations where the available parallelism may vastly exceed the available computation resources. To allow for an accurate space-cost model in such cases, we have previously proposed the Streaming NESL language, a refinement of NESL with a high-level notion of streamable sequences.In this paper, we report on experience with a prototype implementation of Streaming NESL on a 2-level parallel platform, namely a multicore system in which we also aggressively utilize vector instructions on each core. We show that for several examples of simple, but not trivially parallelizable, text-processing tasks, we obtain single-core performance on par with off-the-shelf GNU Coreutils code, and near-linear speedups for multiple cores.",

author = "Madsen, {Frederik Meisner} and Andrzej Filinski",

year = "2016",

month = sep,

day = "8",

doi = "10.1145/2975991.2975998",

language = "English",

pages = "44--51",

booktitle = "Proceedings of the 5th International Workshop on Functional High-Performance Computing",

publisher = "Association for Computing Machinery",

note = "International Workshop on Functional High-Performance Computing ; Conference date: 22-09-2016 Through 22-09-2016",

url = "https://sites.google.com/site/fhpcworkshops/",

}

TY - GEN

T1 - Streaming nested data parallelism on multicores

AU - Madsen, Frederik Meisner

AU - Filinski, Andrzej

N1 - Conference code: 5

PY - 2016/9/8

Y1 - 2016/9/8

N2 - The paradigm of nested data parallelism (NDP) allows a variety of semi-regular computation tasks to be mapped onto SIMD-style hardware, including GPUs and vector units. However, some care is needed to keep down space consumption in situations where the available parallelism may vastly exceed the available computation resources. To allow for an accurate space-cost model in such cases, we have previously proposed the Streaming NESL language, a refinement of NESL with a high-level notion of streamable sequences.In this paper, we report on experience with a prototype implementation of Streaming NESL on a 2-level parallel platform, namely a multicore system in which we also aggressively utilize vector instructions on each core. We show that for several examples of simple, but not trivially parallelizable, text-processing tasks, we obtain single-core performance on par with off-the-shelf GNU Coreutils code, and near-linear speedups for multiple cores.

AB - The paradigm of nested data parallelism (NDP) allows a variety of semi-regular computation tasks to be mapped onto SIMD-style hardware, including GPUs and vector units. However, some care is needed to keep down space consumption in situations where the available parallelism may vastly exceed the available computation resources. To allow for an accurate space-cost model in such cases, we have previously proposed the Streaming NESL language, a refinement of NESL with a high-level notion of streamable sequences.In this paper, we report on experience with a prototype implementation of Streaming NESL on a 2-level parallel platform, namely a multicore system in which we also aggressively utilize vector instructions on each core. We show that for several examples of simple, but not trivially parallelizable, text-processing tasks, we obtain single-core performance on par with off-the-shelf GNU Coreutils code, and near-linear speedups for multiple cores.

U2 - 10.1145/2975991.2975998

DO - 10.1145/2975991.2975998

M3 - Article in proceedings

SP - 44

EP - 51

BT - Proceedings of the 5th International Workshop on Functional High-Performance Computing

PB - Association for Computing Machinery

T2 - International Workshop on Functional High-Performance Computing

Y2 - 22 September 2016 through 22 September 2016

ER -