Causal order in parallel and distributed simulation systems

Abstract

The causal order captures the most basic and important event ordering in distributed systems. In our research, it is applied to solve major problems in two categories of parallel and distributed simulation (PADS) applications, i.e., consistency in distributed virtual environments (DVEs) and performance in Time Warp simulation. In order to achieve more strict consistency in DVEs, inconsistency problems caused by receiving order of concurrent messages are identified. To solve the problems, a causal order based event ordering scheme, the Direct Follow Order (DFO), which enforces uninterrupted delivery of directly coupled causal messages, i.e. , direct-follow messages, is proposed. The DFO is proved to be a stricter ordering scheme than the causal order. However, as discovered in our research, the DFO cannot be always ensured. The sufficiency and necessity conditions under which the DFO can be guaranteed on direct-follow messages are derived. Implementation issues of DFO are also discussed. Performance in Time Warp simulation is greatly determined by the effective advancement of individual logical processes (LPs). Effective advancement can be addressed from various aspects, among which fast cancellation schemes and measures to reduce rollbacks are of our major concerns. These drive our research into two parts: (1) By capturing state dependency and scheduling dependency among events, batch based cancellation scheme is able to carry out rollback optimal cancellation, in which recovery of LPs caused by a straggler message can be done at the cost of at most one rollback; and (2) In view of noteworthy communication latency and possible non-FIFO transmission in distributed simulations, LPs may suffer more rollbacks from their uncontrolled optimistic advancement. To alleviate this problem, the causal order based T i m e Warp simulation is proposed and relevant techniques making the causal order and the time-stamp order consistent are discussed.