摘要：The amount of task-level parallelism (TLP) in a runtime workload is useful information for determining the efficient usage of multiprocessors. This paper presents mechanisms for dynamically estimating the amount of TLP in runtime workloads. Modifications are made to the operating system (OS) to collect information about processor utilization and task activities, from which the TLP can be calculated. By effectively utilizing the time stamp counter (TSC) hardware, the task activities can be monitored with fine time resolution, which enables the TLP to be estimated with fine granularity. We implemented the mechanisms on a recent version of Linux. Evaluation results indicate that the mechanisms can estimate the TLP accurately for various workloads. The overheads imposed by the mechanisms are small.
Loading...
