Chip multicore processors (CMPs) have become the default architecture for modern desktops and servers alike. CMPs consist of several processing cores on a single die making them ideal for multiprogramming and parallel workloads. The CMP cores, however, are not fully independent processors but rather share certain resources amongst each other. These shared resources can become points of contention for threads on neighboring cores leading to significant and often unpredictable performance impact. We propose a contention-mitigation solution via thread-level scheduling. Based on extensive studies into both the causes of contention as well as techniques to predict contention in CMPs we create an effective contention-mitigation scheduler. We highlight the vast design-space of scheduling algorithms and present a novel tool-set called AKULA for rapidly exploring this search space. We demonstrate the advantage of the AKULA tool-set by using it to prototype several new scheduling algorithms and evaluate their utility.
展开▼
