Beyond Profiling: Scaling Profiling Data Usage to Multiple Applications

TL;DR

This paper proposes a method to leverage profiling data from certain applications to improve the performance of different, unseen applications on the same hardware, enhancing overall system efficiency.

Contribution

It introduces a generalized approach for using profiling information across multiple applications to optimize hardware resource usage, demonstrated through cache management.

Findings

19% average reduction in last-level cache blocks

2.8% average increase in cache misses

Effective cross-application profiling data utilization

Abstract

Profiling techniques are used extensively at different parts of the computing stack to achieve many goals. One major goal is to make a piece of software execute more efficiently on a specific hardware platform, where efficiency spans criteria such as power, performance, resource requirements, etc. Researchers, both in academia and industry, have introduced many techniques to gather, and make use of, profiling data. However, one thing remains unchanged: making application A run more efficiently on machine 1. In this paper, we extend this criteria by asking: can profiling information of application A on machine 1 be used to make application B run more efficiently on machine 1? If so, then this means as machine 1 continues to execute more applications, it becomes better and more efficient. We present a generalized method for using profiling information gathered from the execution of programs from a limited corpus of applications to improve the performance of software from outside our corpus. As a proof of concept, we apply our technique to the specific problem of selecting the most efficient last-level-cache with which to execute an application. We were able to turn off an average of 19% of last-level-cache blocks for selected programs from PARSEC benchmark suite and only saw an average 2.8% increase in the rate of last-level cache misses.