2016 USENIX Annual Technical Conference (USENIX ATC '16)

The efficiency of modern multiprogrammed multicore machines is heavily impacted by traffic due to data sharing and contention due to competition for shared resources. In this paper, we demonstrate the importance of identifying latency tolerance coupled with instructionlevel parallelism on the benefits of colocating threads on the same socket or physical core for parallel efficiency. By adding hardware counted CPU stall cycles due to cache misses to the measured statistics, we show that it is possible to infer latency tolerance at low cost. We develop and evaluate SAM-MPH, a multicore CPU scheduler that combines information on sources of traffic with tolerance for latency and need for computational resources. We also show the benefits of using a history of past intervals to introduce hysteresis when making mapping decisions, thereby avoiding oscillatory mappings and transient migrations that would impact performance. Experiments with a broad range of multiprogrammed parallel, graph processing, and data management workloads on 40-CPU and 80-CPU machines show that SAMMPH obtains ideal performance for standalone applications and improves performance by up to 61% over the default Linux scheduler for mixed workloads.