Memory and network bandwidth aware scheduling of multiprogrammed workloads on clusters of SMPs

Koukis, E; Koziris, N

dc.contributor.author	Koukis, E	en
dc.contributor.author	Koziris, N	en
dc.date.accessioned	2014-03-01T02:44:06Z
dc.date.available	2014-03-01T02:44:06Z
dc.date.issued	2006	en
dc.identifier.issn	15219097	en
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/31675
dc.subject	Linux	en
dc.subject	Memory bandwidth	en
dc.subject	Multiprogramming	en
dc.subject	Myrinet	en
dc.subject	Performance counters	en
dc.subject	Resource contention	en
dc.subject	Scheduling	en
dc.subject	SMP clusters	en
dc.subject.other	Myrinet	en
dc.subject.other	Performance counters	en
dc.subject.other	Resource contention	en
dc.subject.other	Symmetric Multiprocessors (SMPs)	en
dc.subject.other	Cluster analysis	en
dc.subject.other	Computer programming languages	en
dc.subject.other	Cost effectiveness	en
dc.subject.other	Data storage equipment	en
dc.subject.other	Multiprogramming	en
dc.subject.other	Resource allocation	en
dc.subject.other	Scheduling	en
dc.subject.other	Throughput	en
dc.subject.other	Multiprocessing systems	en
dc.title	Memory and network bandwidth aware scheduling of multiprogrammed workloads on clusters of SMPs	en
heal.type	conferenceItem	en
heal.identifier.primary	10.1109/ICPADS.2006.59	en
heal.identifier.secondary	http://dx.doi.org/10.1109/ICPADS.2006.59	en
heal.identifier.secondary	1655680	en
heal.publicationDate	2006	en
heal.abstract	Symmetric Multiprocessors (SMPs), combined with modern interconnection technologies are commonly used to build cost-effective compute clusters. However, contention among processors for access to shared resources, as is the main memory bus and the NIC can limit their efficiency significantly. In this paper, we first provide an experimental demonstration of the effect of resource contention on the total execution time of applications. Then, we present the design and implementation of an informed gang-like scheduling algorithm aimed at improving the throughput of multiprogrammed workloads on clusters of SMPs. Our algorithm selects the processes to be coscheduled so as not to saturate nor underutilize the memory bus or network link bandwidth. Its input data are acquired dynamically using hardware monitoring counters and a modified Myrinet NIC firmware, without any modifications to existing application binaries. Experimental evaluation shows throughput can improve up to 40-48% compared to the standard Linux 2.6 O(1) scheduler. © 2006 IEEE.	en
heal.journalName	Proceedings of the International Conference on Parallel and Distributed Systems - ICPADS	en
dc.identifier.doi	10.1109/ICPADS.2006.59	en
dc.identifier.volume	1	en
dc.identifier.spage	345	en
dc.identifier.epage	354	en