Scheduling Pipelined , Multi - Threaded Programs in Linux

A process causes latency when it performs I/O or communication. Pipelined processes mitigate latency by concurrently executing multiple threads—sequences of operations—and overlapping computation, communication, and I/O. When more than one thread is ready to run, the scheduler determines which thread in fact runs. This paper presents techniques for scheduling pipelines, with the following three findings. First, using Linux kernel version 2.6 and the NPTL threads package, we observe a 3-6% performance improvement over kernel version 2.4 and the LinuxThreads package. Second, we test techniques that both take advantage of prior knowledge about whether a program is I/O-bound or compute-bound and raise and lower priorities before the pipeline begins working. These techniques, referred to as fixed scheduling, further improve performance by 5% in the case of the computebound columnsort algorithm. In the I/O-bound algorithm, fixed scheduling failed to yield better performance than the default scheduling. Third, we test simple, adjusting methods that do not take advantage of prior knowledge about whether a program in compute-bound or I/O-bound but rather adjust scheduling as the pipeline progresses. These techniques, called adjusting scheduling, fail to yield better performance than the default scheduling in any of our test cases. In addition, we suggest new scheduling calls and other operating-system improvements.