Adaptive Cache Warming for Faster Simulations

Abstract

The use of hardware-based virtualization allows modern simulators to very quickly fast-forward between sample points and regions of interest. This dramatically reduces the simulation time compared to traditional functional forwarding. However, as the fast-forwarding takes place through virtualized execution on the native hardware, it is unable to warm simulated structures, such as caches. As a result, sampled simulations taking advantage of virtualization for fast-forwarding find their execution time dominated by functional warming. To address the cost of warming, we present Adaptive Cache Warming (ACW), a new fast method that determines how much warming each sample/phase/application needs. ACW takes advantage of the virtualization-based fast-forwarding to search for the minimum warming time required during simulation. To determine when the cache is sufficiently warm, ACW uses heuristics based on the last-level cache's cold-set misses. Our results show that typical practice of conservatively warming last-level caches for around 100M instructions is a vast overkill for nearly all checkpoints. By using ACW, we can adapt the warming per-sample and speedup the simulation by 92--103× on average (512× speedup maximum) depending on cache size (2-32MB).