Department of Computer Science & Engineering

Modern architecture research relies heavily on detailed pipeline simulation. Simulating the full execution of an industry standard benchmark can take weeks to months to complete. To address this issue researchers have recently proposed using Simulation Points (SimPoints) to increase the efficiency and accuracy of simulation. SimPoints are a small subset of execution intervals that are representative of a program's phases, so that when combined they accurately represent the complete execution of the program. These SimPoints are found using clustering techniques based on the structure of the executed code. The prior work on SimPoints assumed a somewhat advanced simulation environment that supported check-points, had perfect warm-up ability, and focused only on simulating single threaded architectures. In this research we explore what happens when you break these assumptions and provide techniques that are capable of dealing with more complex simulation environments. We focus on environments that have the ability to fast-forward, but to not checkpoint, and we extend the use of SimPoints to multi-threaded architectures. We present results comparing our improved SimPoint techniques with past SimPoint work and in addition provide a comparison to Random Sampling.

Pre-2018 CSE ID: CS2002-0727