Y. Ajima, K. Kumon, S. Sumimoto, M. Nagatsuka, T. Shimizu, and Y. Oinaga (Japan)
Super computing, Scheduling, Backfilling, Mesh.
Many modern highly scalable supercomputers employ meshor toroidal interconnects. Such topologies contribute to the system scalability of tens of thousands of nodes but typically require the submesh allocation of jobs. This restriction introduces fragmentation that suppresses utilization of a system. A novel job scheduling scheme Coverage Overlay (CO) backfilling is proposed to combine conservative backfilling with submesh allocation. CO backfilling is intended to schedule advance reservations of submeshes completely. A novel metric Waitload is also introduced to evaluate system performance under the moderately loaded condition and simulate the CO scheme using publicly available parallel workload logs. The simulation results show the difference of system utilization in the range of 10-20%, 27-69% reduced mean response time, and 30-95% reduced mean slowdown compared with the flat allocation.
Important Links:
Go Back
