Statistical Analysis of Communication Time on the IBM SP2
Theodore B. Tabe, Janis P. Hardwick, Quentin F. Stout · 2008
For parallel computers, the execution time of communication routines is an important determinate of users' performance. For one parallel computer, the IBM SP2, all of the higher-level communications routines show a drop in performance as the number of processors involved in the communication increases. Such a drop is unexpected and does not occur on most other parallel machines. While a few others have also recently studied the SP2's communication performance, they have reported only average performance, and failed to comment on the drop in performance or its causes [1, 9]. We generated a distribution of times for these routines and developed a simulator in an attempt to recreate the observed distribution. By studying distributions of communication times and by re ning the simulator, we were able to discern that the performance decrease is due to the variation in the communication times of the lowerlevel primitives upon which the higher-level communication routines are built. This variation is in turn caused by the deleterious e ects of interrupts generated by an operating system untuned to high-performance parallel computing.