An Experimental Study Of L2-Cache Aware Processallocation In Hyper-Threadedsystems

Hyper-threaded systems show an increase in popularity in modern computers due to the performance improvements provided by executing two threads on a single core. However, hyper-threaded systems are unpredictable in terms of execution time due to a state called cache contention, which occurs when two threads simultaneously claim ownership over the shared cache and constantly push out each other’s data out of the cache. In this thesis, we investigate scheduling applications according to their cache usage as means to mitigate the L2 cache contention between threads of a core. We execute a set of applications, including feature detect and synthetic load applications on threads of a core, and measure the cache utilization. We then analyze the cache usage of our applications by creating resource models and empirically investigate how to schedule the applications to avoid L2 cache contention effectively. The results of this work are that we created a static schedule that partly avoids cache contention. Scheduling applications using this static schedule can improve their execution time up to 50% and up to 30% fewer cache misses than in congested scenarios