LSim: Fine-Grained Simulation Framework for Large-Scale Performance Evaluation

Abstract

As large-scale workloads with massive parallelism emerge, the demand for large-scale systems such as datacenters and supercomputers is rising sharply. To accurately design a large-scale system, architects heavily rely on performance modeling at design phases. However, modeling a large-scale workload without a large-scale system is a challenging problem. This paper presents LSim, a framework for large-scale performance evaluation. Based on the captured behavior within small-scale workload traces, LSim extrapolates the behavior of the workload on a large-scale system. To do so, we propose two techniques: (1) representative trace model and (2) function latency model to synthesize a trace and to predict the latency of functions in the synthesized trace, respectively.