Charm++ parallel programming system

[jyeom <jyeom AT cs.vt.edu>]

I profiled my code using open|speedshop, and noticed that CmiWallTimer() was talking 5% of the sum of the time spent by every core. (convcore.c, line 987)

I was running Charm on top of mvapich2 compiled by intel compiler on a infiniband-sandbridge cluster.

The time seems especially spent for "_absoluteTime?currenttime:currenttime - inittime_wallclock;"

_absoluteTime is set once at initialization by "+useAbsoluteTime".

In fact, it took about 20% initially in some case when I call the routine at multiple places in my application.

After I removed all the calls in my application, it dropped to 5%.

Since it is now called by the runtime processes via CcdRaiseCondition() and CcdCallBacks() in conv-conds.c (line 521, 532), I do not believe it would affect the overall performance much. The call overlaps the application's computation in other cores in the SMP mode. Abhinav recommended to bring this to here just in case.

Jae-Seung