Process migration in virtual memory multicomputer systems

In a DVM environment, process migration can be achieved by transferring the process control block (PCB) of a process to the destination node. The process, at the new host node, requests the pages that belong to its working set on demand. Hence, process migration does not require that the data or the code of the process be transferred at migration time. Consequently, the time and the complexity involved in migrating a process is reduced to the time and complexity of migrating ihe PCB of the process. This applies for new processes as well as to processes that have begun their execution. Under DVM, the pages of a process are dynamically distributed over several nodes in the system. Therefore, a migrating process is likely to find part of its virtual space allocated to the same node to which the process has migrated. The paper addresses the typical questions when to invoke migration, which process to migrate, and where lo send the selected process This paper presents two approaches for solving the above questions. We define a load index as a function of the number of jobs at the node, the page fault rate, and the remaining unused memory. The load index is used to determine whether migration is needed or not. Tkace driven simulation of a dynamic distributed system is used to evaluate the performance of the various process migration techniques.