Issues in Process Migration

Process migration means moving a process in the middle of its execution from one processor or host to another, for a variety of reasons. Usually, processes are migrated with the aim of balancing the work load across the cluster so that the capacity of underutilized nodes are also exploited. The idea of process migration is borne out of the fact that in the over whelming cases of non migratory scenarios, the capacity of most nodes are under utilized. Allocation can be static or dynamic. By its very nature, process migration presents a difficult task with many complex issues to be resolved transparency, scheduling and allocation policies, interaction with file system, naming, and scaling. In this paper we look at some well known implementations Condor, Sprite, MOSIX and MPVM. We see that facilitating process migration adds considerable complexity to the kernel where it is implemented at the kernel level (Sprite, MOSIX) and hinders transparency if implemented at the user level. We also look at some analytical results regarding process migration benefits for load sharing. We see that the cost of migrating a performance comes mainly from the cost of transferring state. We conclude that migration, in the general case, should be used only sparingly, and often as a last resort.