Inverse Amdahl's Law



The speedup curves for some of the applications deviate quite a bit from the linear speedup case. The reason given for such a behavior is that the data sets used for the simulation runs were small, and therefore the startup costs for these applications occupy a substantial portion of the running time. Also, the Splash-2 researchers claim that with bigger data sets the performance would scale better. An experiment that would support such a claim and also make the methodological process sound would be to separate the time spent in the sequential and parallel sections of the program.

The available concurrency could deviate from the ideal PROCS fold parallelism either because of sequential work replicated across all processors, time spent in locks and mutually exclusive sections of the code, or due to lack of sufficient parallelism throughout the computation. By separating out the time spent in sequential section while presenting speedup characteristics of the application, we have a greater understanding of the parallel program. In particular, a program that has (0.8*PROCS) tasks in the parallel section could be differentiated from a program that spends 20% of the time in sequential code fragments with PROCS fold parallelism in the parallel phase.


next up previous
Next: Single Cycle Assumption Up: Concurrency and Load Previous: Concurrency and Load

Randy Wang    rywang.public@gmail.com    Sun Mar 19 22:18:59 PST 1995