SHF: SMALL: LUCID: Low-overhead, Unobtrusive Cache Contention Detection and Repair

University Of Pennsylvania, Philadelphia PA

Investigators

Abstract

Title: SHF: SMALL: LUCID: Low-overhead, Unobtrusive Cache Contention Detection and Repair The last decade has overseen a historic shift in processor design towards multicore architectures. From servers to tablets, phones and watches, processor architects are using multiple cores to achieve performance and energy targets. Unfortunately, writing parallel code that performs well on these processors is far from easy. The intellectual merits of this project include developing new algorithms for finding and fixing performance bugs in parallel code by integrating support from across the computing stack including hardware, compilers and runtime systems. The project's broader significance and importance is to improve the performance and energy efficiency of parallel code, providing benefits for the users of all multicore computing devices. This project also seeks to encourage participation from underrepresented groups in the next generation of computer scientists. This project leverages recent advances in hardware performance counters that allow for low-overhead, unobtrusive profiling of cache contention in real systems. This performance counter information drives new techniques for finding and automatically fixing cache contention without programmer intervention. The research explores integrating contention detection and repair with managed language runtimes. The researchers plan to openly distribute the systems built for this project to facilitate examination by other researchers and to integrate the research results into education.