Execution-Driven Distributed Simulation of Parallel Architectures

A new methodology for the asynchronous discrete event-driven simulation of parallel computers is proposed. This methodology integrates sequential and distributed simulation in a unified paradigm and is applicable to the simulation of all classes of parallel computer architectures. In our own simulation work we accelerated simulations by more than an order of magnitude with parallel execution speedup efficiencies in the order of 60-70%. When simulating in parallel, our approach has the important benefit of testing the robustness of a simulated design by not hiding the asynchronous nature of the system being studied (our simulation model preserves the nondeterministic behavior of certain parallel executions). Unlike other distributed simulation methodologies, our execution model does not rely on a global view of virtual time to maintain coherent distributed event causality relations. The simulator correctly executes a given parallel application that observes a particular synchronization model of choice without a notion of virtual time. We then estimate a global virtual time in which the execution could have been carried out. We give a detailed description of our simulation methodology, the computational models that it can implement, and the conditions for its correctness. We also give some preliminary performance results obtained by implementing our parallelization technique to simulate a massively parallel machine on a CM-5 computer and on a heterogeneous network of workstations.