Wait-free cache-affinity thread scheduling

Abstract

Cache utilisation is often very poor in multithreaded applications, due to the loss of data access locality incurred by frequent context switching. This problem is compounded on shared memory multiprocessors when dynamic load balancing is introduced, as thread migration disrupts cache content. Batching, a technique for reducing the negative impact of fine grain multithreading on cache performance is introduced to mitigate this problem. Moreover, the related issue of contention for shared data within a thread scheduler for shared memory multiprocessors is considered. In this regard wait-free techniques are applied in lieu of conventional lock-based methods for accessing internal scheduler structures, alleviating to some extent serialisation and hence the degree of contention. Prototype schedulers which make use of the above ideas are described, and finally experimental results which illustrate the observed improvements are presented.