Shock resistant time warp
A. Ferscha, J. Johnson · 2003
In an attempt to cope with time-varying workload, traditional adaptive time warp protocols are designed to react in response to performance changes by altering control parameter configurations, like the amount of available memory, the size of the checkpointing interval, the frequency of GVT computation, fossil collection invocations, etc. We call those schemes "reactive" because all control decisions are undertaken based on historical performance information collected at runtime, and come into effect in future system states. We develop environment aware, self adaptive time warp LPs implementing a pro-active performance control scheme, addressing the timeliness of control decisions. Opposed to reactive TW schemes, our pro-active control mechanism based on a statistical analysis of the state history ...S/sub t-2/spl Delta//, S/sub t-/spl Delta//, S/sub t/ periodically collected in (real) time intervals of size /spl Delta/, forecasts a future LP state S/spl circ//sub t+/spl Delta//. A performance control decision CD(S/spl circ//sub t+/spl Delta//) is established, that is most appropriate for the expected future LP state, i.e. the state when the corresponding control activity would become effective. Depending on the forecast quality, a pro-active scheme will presumably exhibit performance superior to reactive schemes, at least for cases where state changes in the time frame /spl Delta/ are very likely. We study the ability of pro-active TW LPs, to adapt to sudden load changes, especially to abruptly occurring background workloads injected by other applications executing concurrently with the TW simulation on a network of workstations. Experimental results show that the protocol is able to capture abrupt changes in both computational and communication resource availability, justifying the title: shock resistant time warp.