Reliability in Warehouse-Scale Computing: Why Low Latency Matters

Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

82 Downloads (Pure)

Abstract

Warehouse sized buildings are nowadays hosting several types of large computing systems: from supercomputers to large clusters of servers to provide the infrastructure to the cloud. Although the main target, especially for high-performance computing, is still to achieve high throughput, the limiting factor of these warehouse-scale data centers is the power dissipation. Power is dissipated not only in the computation itself, but also in heat removal (fans, air conditioning, etc.) to keep the temperature of the devices within the operating ranges. The need to keep the temperature low within a minimal power envelope and to maintain high throughput and high reliability poses hard challenges.

In this work, we show that by moving part of the computation to accelerators, not only we reduce the latency of operations, but also make the system more energy efficient and reliable.
Original languageEnglish
Title of host publicationProceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale
Publication date2015
Pages2-6
Publication statusPublished - 2015
EventMEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale - Tallinn, Estonia
Duration: 10 Nov 201511 Nov 2015
http://www.median-project.eu/events/median-finale

Workshop

WorkshopMEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale
CountryEstonia
CityTallinn
Period10/11/201511/11/2015
Internet address

Cite this

Nannarelli, A. (2015). Reliability in Warehouse-Scale Computing: Why Low Latency Matters. In Proceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale (pp. 2-6)
Nannarelli, Alberto. / Reliability in Warehouse-Scale Computing: Why Low Latency Matters. Proceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale. 2015. pp. 2-6
@inproceedings{44dfbe141b834d9aa21b7ad7aff5840c,
title = "Reliability in Warehouse-Scale Computing: Why Low Latency Matters",
abstract = "Warehouse sized buildings are nowadays hosting several types of large computing systems: from supercomputers to large clusters of servers to provide the infrastructure to the cloud. Although the main target, especially for high-performance computing, is still to achieve high throughput, the limiting factor of these warehouse-scale data centers is the power dissipation. Power is dissipated not only in the computation itself, but also in heat removal (fans, air conditioning, etc.) to keep the temperature of the devices within the operating ranges. The need to keep the temperature low within a minimal power envelope and to maintain high throughput and high reliability poses hard challenges.In this work, we show that by moving part of the computation to accelerators, not only we reduce the latency of operations, but also make the system more energy efficient and reliable.",
author = "Alberto Nannarelli",
year = "2015",
language = "English",
pages = "2--6",
booktitle = "Proceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale",

}

Nannarelli, A 2015, Reliability in Warehouse-Scale Computing: Why Low Latency Matters. in Proceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale. pp. 2-6, MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale, Tallinn, Estonia, 10/11/2015.

Reliability in Warehouse-Scale Computing: Why Low Latency Matters. / Nannarelli, Alberto.

Proceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale. 2015. p. 2-6.

Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

TY - GEN

T1 - Reliability in Warehouse-Scale Computing: Why Low Latency Matters

AU - Nannarelli, Alberto

PY - 2015

Y1 - 2015

N2 - Warehouse sized buildings are nowadays hosting several types of large computing systems: from supercomputers to large clusters of servers to provide the infrastructure to the cloud. Although the main target, especially for high-performance computing, is still to achieve high throughput, the limiting factor of these warehouse-scale data centers is the power dissipation. Power is dissipated not only in the computation itself, but also in heat removal (fans, air conditioning, etc.) to keep the temperature of the devices within the operating ranges. The need to keep the temperature low within a minimal power envelope and to maintain high throughput and high reliability poses hard challenges.In this work, we show that by moving part of the computation to accelerators, not only we reduce the latency of operations, but also make the system more energy efficient and reliable.

AB - Warehouse sized buildings are nowadays hosting several types of large computing systems: from supercomputers to large clusters of servers to provide the infrastructure to the cloud. Although the main target, especially for high-performance computing, is still to achieve high throughput, the limiting factor of these warehouse-scale data centers is the power dissipation. Power is dissipated not only in the computation itself, but also in heat removal (fans, air conditioning, etc.) to keep the temperature of the devices within the operating ranges. The need to keep the temperature low within a minimal power envelope and to maintain high throughput and high reliability poses hard challenges.In this work, we show that by moving part of the computation to accelerators, not only we reduce the latency of operations, but also make the system more energy efficient and reliable.

M3 - Article in proceedings

SP - 2

EP - 6

BT - Proceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale

ER -

Nannarelli A. Reliability in Warehouse-Scale Computing: Why Low Latency Matters. In Proceedings of MEDIAN Finale - Workshop on Manufacturable and Dependable Multicore Architectures at Nanoscale. 2015. p. 2-6