A Library-Based Tool to Improve CPU Assignment for Multicore Processor-Based Pervasive Servers

This paper proposes SPLiT (Scalable Performance Library Tool) as the methodology to improve performance of applications on multicore processors, through CPU, and cache optimizations. SPLiT analyzes applications based on cycle counts, and cache misses, and predicts behavior of the applications according to the analysis. With these acquired knowledge on the target application, SPLiT improves CPU resource allocation determined by operating systems. SPLiT, and SPLiT library are designed to achieve resource optimizations based on the combination of hardware information collected by Operating System kernel, and software information collected by applications. Additionally, all the features of SPLiT are immediately available with the small modifications to the application source codes, and the modifications are basically to implant SPLiT library calls into the source code. This simple requirement diminishes programmers' difficulties for performance tuning of multicore applications. Through empirical experiments, we validated the efficiency of SPLiT, and observed that performance of the web application was improved by 26%.