Virtual Time Based Dynamic Load Management in the Time Warp Operating System

The Time Warp Operating System (TWOS) executes event-driven simulations in an optimistic style on parallel machines. Recently TWOS has been substantially improved by the addition of dynamic load management for the purpose of (a) handling fluctuations in a simulation’s performance, (b) dealing effectively with dynamic creation and destruction of processes, and (c) eliminating the burden on users of assigning processes to processors. Because TWOS uses optimistic synchronization, existing load management theory, which tends to be based on balancing processor utilization, is not applicable; TWOS instead balances load using a more general metric called effective utilization. In addition, TWOS introduces a new program unit called a phase, which is a process delimited by two simulation times. In TWOS, the phase, not the process, is the fundamental unit of scheduling, migration, and rollback, as well as the target of messages. This paper describes early results of our experiments with the TWOS dynamic load management facility. It covers the theory and mechanics of phase splitting and migration, and presents performance results for several combinations of benchmark, configuration, and load management policy.