The Multi-Program Performance Model: Debunking current practice in multi-core simulation

Composing a representative multi-program multi-core workload is non-trivial. A multi-core processor can execute multiple independent programs concurrently, and hence, any program mix can form a potential multi-program workload. Given the very large number of possible multi-program workloads and the limited speed of current simulation methods, it is impossible to evaluate all possible multi-program workloads. This paper presents the Multi-Program Performance Model (MPPM), a method for quickly estimating multi program multi-core performance based on single-core simulation runs. MPPM employs an iterative method to model the tight performance entanglement between co-executing programs on a multi-core processor with shared caches. Because MPPM involves analytical modeling, it is very fast, and it estimates multi-core performance for a very large number of multi-program workloads in a reasonable amount of time. In addition, it provides confidence bounds on its performance estimates. Using SPEC CPU2006 and up to 16 cores, we report an average performance prediction error of 2.3% and 2.9% for system throughput (STP) and average normalized turnaround time (ANTT), respectively, while being up to five orders of magnitude faster than detailed simulation. Subsequently, we demonstrate that randomly picking a limited number of multi-program workloads, as done in current practice, can lead to incorrect design decisions in practical design and research studies, which is alleviated using MPPM. In addition, MPPM can be used to quickly identify multi-program workloads that stress multi-core performance through excessive conflict behavior in shared caches; these stress workloads can then be used for driving the design process further.