Survey of TeraGrid Job Distribution: Toward Specialized Serial Machines as TeraGrid Resources

Abstract

As we proceed towards the age of petascale computing, it is important to be aware that even today more than half of national cyberinfrastructure users are serial users who run single processor code; more over, coarse-grained parallel application do not necessarily benefit from a high-speed low-latency interconnect. While a majority of compute resources on the TeraGrid today are massive parallel machines with high-speed low-latency interconnects like Myrinet or Infiniband, optimized to run large fine-grained parallel applications that use hundreds of processors/cores in parallel, usage patterns indicate that there is still considerable demand that could be just as effectively met by large computational resources with no special high-speed or low-latency interconnects. Research allocations involving large serial applications or coarse-grained parallel applications could be allocated to these machines, thus possibly leading to decreased wait times for massive parallel and large serial jobs. This change in focus would also lower the financial barrier for potential new resource providers to the national cyberinfrastructure, by allowing them to reallocate funds from the interconnect to additional computational capacity.