Optimizing memory management for optimistic simulation with reinforcement learning

Simulation is a powerful technique to explore complex scenarios and analyze systems related to a wide range of disciplines. To allow for an efficient exploitation of the available computing power, speculative Time Warp-based Parallel Discrete Event Simulation is universally recognized as a viable solution. In this context, the rollback operation is a fundamental building block to support a correct execution even when causality inconsistencies are a posteriori materialized. If this operation is supported via checkpoint/restore strategies, memory management plays a fundamental role to ensure high performance of the simulation run. With few exceptions, adaptive protocols targeting memory management for Time Warp-based simulations have been mostly based on a pre-defined analytic models of the system, expressed as a closed-form functions that map system's state to control parameters. The underlying assumption is that the model itself is optimal. In this paper, we present an approach that exploits reinforcement learning techniques. Rather than assuming an optimal control strategy, we seek to find the optimal strategy through parameter exploration. A value function that captures the history of system feedback is used, and no a-priori knowledge of the system is required. An experimental assessment of the viability of our proposal is also provided for a mobile cellular system simulation.