DLG parallel efficiency on bwulf at the UMD VDIL

Program "dlgmcp" simulates the driven lattice gas using the Monte Carlo method in parallel . Each process accesses an independent stream of pseudorandom numbers using the additive lagged Fibonacci generator in the sprng library. A simulation collects and averages data from several (8 in the case of the data below) independent runs.

The parallel code is in the class of "embarrassingly parallel" methods, the runs merely being divided among the processes.


If np is the number of processes, then np-1 processes do Monte Carlo runs while one process is reserved to collect and average data. That server process uses approximately zero cpu time.


<U/(6J)> is the average energy in dimensionless units
time = sum total elapsed time for all processes
np-1
(workers)
tworkers(sec) np
(all)
tall(sec) <U/(6J)> std err in
<(U/(6J)>
12199 24399 -1.4830.001333
22220 33346 -1.4830.001323
32193 43017 -1.4830.001390
42185 52734 -1.4830.001356
521996 2758 -1.4830.001377
62235 72794 -1.4830.001365

One sees that total processor time is approximately independent of the number of processors. Results (energy and std error of energy) are nearly independent of the number of processors. That results depend slightly on the number of processors probably stems from the dependence of random numbers on processors.


go back to MPI on beowulf at VDIL index.