next up previous contents
Next: Parallel Workload Modeling Up: Scheduling Parallel Workloads : Previous: Focus

Methodology

We used CVM [3], a software based distributed-shared memory system, to extract traces for the various applications. The traces were time-stamped sequence of events that were reported happen in each of the process of a job. To get the traces, we instrumented the CVM source code. We used the gettimeofday() routine to figure out instants at which the events occur. Hence, although we measure the timing information to the granularity of microseconds, the actual accuracy may be a magnitude of order larger. Then we implemented a discrete event simulator on which to re-run these traces, using both the co-scheduling and implicit scheduling techniques. The simulator was allowed to re-order the sequence of events from different processors, since only the necessary dependencies like lock requests can follow lock releases, need to be preserved and the actual ordering will depend on the exact run-time environment. We finally compared the performance statistics that we got from the simulator.



Suman Banerjee
Tue May 20 22:29:25 EDT 1997