Exploiting spatial parallelism in Ethernet-based cluster interconnects.

Stavros Passas; George Kotsis; Sven Karlsson; Angelos Bilas

doi:10.1109/IPDPS.2008.4536145

Exploiting spatial parallelism in Ethernet-based cluster interconnects.

Stavros Passas, George Kotsis, Sven Karlsson, Angelos Bilas

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

Abstract

In this work we examine the implications of building a single logical link out of multiple physical links. We use MultiEdge to examine the throughput-CPU utilization tradeoffs and examine how overheads and performance scale with the number and speed of links. We use low- level instrumentation to understand associated overheads, we experiment with setups between 1 and 8 1-GBit/s links, and we contrast our results with a single 10-GBit/s link. We find that: (a) Our base protocol achieves up-to 65% of the nominal aggregate throughput, (b) Replacing the interrupts with polling significantly impacts only the multiple link configurations, reaching 80% of nominal throughput, (c) The impact of copying on CPU overhead is significant, and removing copying results in up-to 66% improvement in maximum throughput, reaching almost 100% of the nominal throughput, (d) Scheduling packets over heterogeneous links requires simple but dynamic scheduling to account for different link speeds and varying load.

Original language	English
Title of host publication	Proceedings of Workshop on Communication Architecture for Clusters, CAC 2008
Publisher	IEEE Computer Society Press
Publication date	2008
ISBN (Print)	978-1-4244-1693-6
DOIs	https://doi.org/10.1109/IPDPS.2008.4536145
Publication status	Published - 2008
Event	2008 IEEE International Symposium on Parallel and Distributed Processing - Miami, United States Duration: 14 Apr 2008 → 18 Apr 2008 Conference number: 22 https://ieeexplore.ieee.org/xpl/conhome/4519061/proceeding

Conference

Conference	2008 IEEE International Symposium on Parallel and Distributed Processing
Number	22
Country/Territory	United States
City	Miami
Period	14/04/2008 → 18/04/2008
Internet address	https://ieeexplore.ieee.org/xpl/conhome/4519061/proceeding

Access to Document

10.1109/IPDPS.2008.4536145

OpenUrl availability

Full text

Cite this

@inproceedings{8fcadbf19c594c258d4192eb1d1cc953,

title = "Exploiting spatial parallelism in Ethernet-based cluster interconnects.",

abstract = "In this work we examine the implications of building a single logical link out of multiple physical links. We use MultiEdge to examine the throughput-CPU utilization tradeoffs and examine how overheads and performance scale with the number and speed of links. We use low- level instrumentation to understand associated overheads, we experiment with setups between 1 and 8 1-GBit/s links, and we contrast our results with a single 10-GBit/s link. We find that: (a) Our base protocol achieves up-to 65% of the nominal aggregate throughput, (b) Replacing the interrupts with polling significantly impacts only the multiple link configurations, reaching 80% of nominal throughput, (c) The impact of copying on CPU overhead is significant, and removing copying results in up-to 66% improvement in maximum throughput, reaching almost 100% of the nominal throughput, (d) Scheduling packets over heterogeneous links requires simple but dynamic scheduling to account for different link speeds and varying load.",

author = "Stavros Passas and George Kotsis and Sven Karlsson and Angelos Bilas",

year = "2008",

doi = "10.1109/IPDPS.2008.4536145",

language = "English",

isbn = "978-1-4244-1693-6",

booktitle = "Proceedings of Workshop on Communication Architecture for Clusters, CAC 2008",

publisher = "IEEE Computer Society Press",

note = "2008 IEEE International Symposium on Parallel and Distributed Processing, IPDPS 2008 ; Conference date: 14-04-2008 Through 18-04-2008",

url = "https://ieeexplore.ieee.org/xpl/conhome/4519061/proceeding",

}

Passas, S, Kotsis, G, Karlsson, S & Bilas, A 2008, Exploiting spatial parallelism in Ethernet-based cluster interconnects. in Proceedings of Workshop on Communication Architecture for Clusters, CAC 2008. IEEE Computer Society Press, 2008 IEEE International Symposium on Parallel and Distributed Processing, Miami, Florida, United States, 14/04/2008. https://doi.org/10.1109/IPDPS.2008.4536145

TY - GEN

T1 - Exploiting spatial parallelism in Ethernet-based cluster interconnects.

AU - Passas, Stavros

AU - Kotsis, George

AU - Karlsson, Sven

AU - Bilas, Angelos

N1 - Conference code: 22

PY - 2008

Y1 - 2008

N2 - In this work we examine the implications of building a single logical link out of multiple physical links. We use MultiEdge to examine the throughput-CPU utilization tradeoffs and examine how overheads and performance scale with the number and speed of links. We use low- level instrumentation to understand associated overheads, we experiment with setups between 1 and 8 1-GBit/s links, and we contrast our results with a single 10-GBit/s link. We find that: (a) Our base protocol achieves up-to 65% of the nominal aggregate throughput, (b) Replacing the interrupts with polling significantly impacts only the multiple link configurations, reaching 80% of nominal throughput, (c) The impact of copying on CPU overhead is significant, and removing copying results in up-to 66% improvement in maximum throughput, reaching almost 100% of the nominal throughput, (d) Scheduling packets over heterogeneous links requires simple but dynamic scheduling to account for different link speeds and varying load.

AB - In this work we examine the implications of building a single logical link out of multiple physical links. We use MultiEdge to examine the throughput-CPU utilization tradeoffs and examine how overheads and performance scale with the number and speed of links. We use low- level instrumentation to understand associated overheads, we experiment with setups between 1 and 8 1-GBit/s links, and we contrast our results with a single 10-GBit/s link. We find that: (a) Our base protocol achieves up-to 65% of the nominal aggregate throughput, (b) Replacing the interrupts with polling significantly impacts only the multiple link configurations, reaching 80% of nominal throughput, (c) The impact of copying on CPU overhead is significant, and removing copying results in up-to 66% improvement in maximum throughput, reaching almost 100% of the nominal throughput, (d) Scheduling packets over heterogeneous links requires simple but dynamic scheduling to account for different link speeds and varying load.

U2 - 10.1109/IPDPS.2008.4536145

DO - 10.1109/IPDPS.2008.4536145

M3 - Article in proceedings

SN - 978-1-4244-1693-6

BT - Proceedings of Workshop on Communication Architecture for Clusters, CAC 2008

PB - IEEE Computer Society Press

T2 - 2008 IEEE International Symposium on Parallel and Distributed Processing

Y2 - 14 April 2008 through 18 April 2008

ER -

Exploiting spatial parallelism in Ethernet-based cluster interconnects.

Abstract

Conference

Access to Document

OpenUrl availability

Fingerprint

Cite this