Current resource managers do not have adequate node allocation and distribution strategies to efficiently schedule jobs on multi-core multi-threaded systems. Clusters composed of multiple cores per processor as well as multiple processors per node are presenting new challenges to users when trying to run their program efficiently on this class of machines. New allocation algorithms and their respective user interfaces need to be devised to ensure minimum contention for cache and memory, reduced on-chip contention, etc. as well as evaluate trade-offs between resource contentions.
Susanne M. Balle, Daniel J. Palermo