Network RAM Based Process Migration for HPC Clusters

Process migration is critical to dynamic balancing of workloads on cluster nodes in any high performance computing cluster to achieve high overall throughput and performance. Most existingprocess migration mechanisms are however unsuccessful in achieving this goal proper because theyeither allow once-only migration of processes or have complex implementations of address space transfer that degrade process migration performance. We propose a new process migration mechanism for HPC clusters that allows multiple migrations of each process by using the network RAM feature ofclusters to transfer the address spaces of processes upon their multiple migrations. We show experimentally that the superiority of our proposed mechanism in attaining higher performance compared to existing comparable mechanisms is due to effective management of residual data dependencies

Keywords:

High Performance Computing (HPC) Clusters , Process Migration , Network RAM , Load Balancing , Address Space Transfer