UPMARC Workshop on Task-Based Parallel Programming

David Black-Schaffer, UPMARC.

Abstract. Modern multicore processors share cache and bandwidth across all cores, leading to performance degradation for cache- or bandwidth-sensitivite tasks. To understand this impact, we need to know the task's sensitivity to its shared resource allocation. However, this is quite difficult to model due to the complexities of the hardware, including out-of-order execution, hardware prefetching, memory system queues, and the details of cache replacement policies. To overcome these difficulties we have developed techniques that allow us to "steal" shared resources while measuring performance, thereby precisely capturing the shared resource sensitivity, while including all the effects of the real hardware. This data can then be used to accurately predict the impact of resource sharing. This talk will cover these technologies in the context of profiling applications and how they could work together to enable accurate shared resource modeling for task-based runtimes.