Application Transparent and Efficient Mixed State-Saving in Speculative Simulation Platforms

Davide Cingolani

Abstract:
In this thesis we focus on devising a reverse code generator and on its integration with speculative simulation platforms, and specifically for Parallel Discrete Event Simulation (PDES). Speculative simulation platforms usually require a considerable storage allocation to model the reality and to store partial results, which further experience several update per time; nonetheless, the optimistic slant increases the likelihood of rollbacks. Simulation processes exhibit CPU-intensive burst loads too, representing a perfect field of application to test our approach in all the possible cases at once. Simulation is a problem-solving technique to cope with complex mathematical models generally conceived from real (or hypothetical) phenomena, which are non-trivially reproducible otherwise. Simulation applications handle a considerable number of parallel/distributed objects interacting together by message passing. Each object is a logical entity which relies on the virtual time concept, processing the incoming messages. The event-based simulation enforces equivalency between those messages and the relative triggered events. Unlike other structured processes, speculative simulation adopts optimistic heuristics, which allow to perform scheduled events even if they are not safe. Event safety straightway depends on actual processing events order with respect to global causality relationship they have been sent with. Optimistic approach looses event processing constraints and exploits much better computational resources, nevertheless it might bring the system to violate the causal order, bringing the simulation to an inconsistent state. The more complex simulation model is, the more likely it requires to rollback out-of-order events. So far the most consolidated way is to employ checkpointing techniques, which though exhibit considerable memory overhead and time latency as the simulation model gets more complex.

BibTeX Entry: