Dynamic coscheduling on workstation clusters.pdfVIP

Dynamic Coscheduling on Workstation Clusters Patrick G Sobalvarro Scott Pakin William E Weihl Andrew A Chien 1 DEC Systems Research Center 2 University of Illinois at UrbanaChampaign March Submitted for Publication Abstract Coscheduling has b een shown to b e a critical factor in achieving ecient parallel execution in time shared environments However the most common approach gang scheduling has limitations in scaling can compromise go o d interactive resp onse and requires that communicating pro cesses b e identied in advance We explore a technique called dynamic coscheduling DCS which pro duces emergent coscheduling of the pro cesses constituting a parallel job Exp eriments are p erformed in a workstation environment with high p erformance networks and autonomous timesharing schedulers for each CPU The results demonstrate that DCS can achieve eective robust coschedulin g for a range of workloads and background loads Empirical comparisons to implicit scheduling and unco ordinated scheduling are presented Under spinblo ck synchronizati on DCS reduces job resp onse times by up to over implicit scheduling while maintaining fairness and under spinning synchronizatio n DCS reduces job resp onse times by up to two decimal orders of magnitude over unco ordinated scheduling The results suggest that DCS is a promising avenue for achieving co ordinated parallel scheduling in an environment that co exists with autonomous no de schedulers Intro duction Co ordinated scheduling of parallel jobs across the no des of a