Online Linear Regression of Sampling Data from Performance Event Counters

Sampling is a common way to collect execution information with performance event counters. However, the sampling data generated from performance event counters tend to be massive if sampling with high frequencies. Storing and processing a large amount of sampling data require much disk space and computing time. In this paper, we propose the online linear regression method to reduce the size of the sampling data. Our idea is to fit the sampling data with a series of straight lines. Each line represents the variation trend of the sample values within the corresponding section. Then we store the parameters of the lines instead of the sample values, resulting in a reduction of the sampling data size. The SPEC CPU 2006 benchmarks are tested to verify the proposed online linear regression method. The experimental result shows the online linear regression method can reduce the sampling data size effectively with a low overhead, and retain the variation characteristic of the sampling data with a normalized estimated standard deviation less than 0.1.