HETSIM: Simulating Large-Scale Heterogeneous Systems using a Trace-driven, Synchronization and Dependency-Aware Framework

Early-stage design-space exploration (DSE) and performance/power evaluation of large-scale heterogeneous systems, such as those composed of chip multiprocessors (CMPs) coupled with fixed-function logic, have gained importance in the dark silicon era. Prior work has explored trace-based simulation techniques, that offer good trade-offs between simulation accuracy and speed, to simulate CMPs with up to 100s of threads, and similar methods are used for accelerators. However, there is lack of a unified framework for fast simulation of large-scale heterogeneous systems. In this work, we propose HETSIM, a trace-driven framework for estimating performance and power of accelerators, CMPs and heterogeneous systems with 1000s of cores. We present results on (i) a CMP and (ii) a heterogeneous accelerator, demonstrating average speed ups of 5.5× and 16.1× over detailed gem5 models with deviations in simulated time and power consumption of 4.6-28.1% and 1.7-3.3%, respectively.