Maintaining quality of service with dynamic fault tolerance in fat-trees

Authors:
Frank Olaf Sem-Jacobsen;Tor Skeie
Affiliations:
Department of Informatics, University of Oslo, Oslo, Norway and Networks and Distributed Systems, Simula Research Laboratory, Lysaker, Norway;Department of Informatics, University of Oslo, Oslo, Norway and Networks and Distributed Systems, Simula Research Laboratory, Lysaker, Norway
Venue:
HiPC'08 Proceedings of the 15th international conference on High performance computing
Year:
2008

Citing 10
Cited 0

Fat-trees: universal networks for hardware-efficient supercomputing

IEEE Transactions on Computers
Fault-tolerant routing in MIN-based supercomputers

Proceedings of the 1990 ACM/IEEE conference on Supercomputing
A Strategy to Compute the InfiniBand Arbitration Tables

IPDPS '02 Proceedings of the 16th International Parallel and Distributed Processing Symposium
A Strategy to Manage Time Sensitive Traffic in InfiniBand

IPDPS '02 Proceedings of the 16th International Parallel and Distributed Processing Symposium
K-ary N-trees: High Performance Networks for Massively Parallel Architectures

K-ary N-trees: High Performance Networks for Massively Parallel Architectures
Siamese-Twin: A Dynamically Fault-Tolerant Fat-Tree

IPDPS '05 Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS'05) - Papers - Volume 01
QsNetII: Defining High-Performance Network Design

IEEE Micro
Dynamic Fault Tolerance with Misrouting in Fat Trees

ICPP '06 Proceedings of the 2006 International Conference on Parallel Processing
Decoupling the Bandwidth and Latency Bounding for Table-based Schedulers

ICPP '06 Proceedings of the 2006 International Conference on Parallel Processing
Combining Source Routing and Dynamic Fault Tolerance

SBAC-PAD '06 Proceedings of the 18th International Symposium on Computer Architecture and High Performance Computing

Quantified Score

Hi-index	0.00

Visualization

Abstract

A very important ingredient in the computing landscape isUtility Computing Data Centres (UCDCs), large-scale computing systemsthat offer computational services to concurrently running jobsthrough virtual servers. As UCDC systems increase in size and the meantime between failure decreases, it is becoming an increasingly importantchallenge to expediently tolerate failures (dynamically), while distributingthe effects of the failure amongst the virtual servers according to theirservice level agreements. We propose and evaluate a strategy for offeringpredictable service in fat-trees experiencing faults, by reprioritisingpackets. The strategy is able to distribute the effect of network faults inorder to satisfy a number of quality-of-service demands. Which demandsto favour depends on the computer system and the characteristics of thejobs it is running, and in the presence of a moderate number of faults itis to some degree possible to meet the demands.