>
> Kanoj, earlier you had said the wall time looks bad.
> I don't understand. For the 1-way make, it took 13:28 == 808 seconds,
> of which 742sec were in user mode (> 90%) ... which is typical of
> large compilation workloads.
13:28 seconds for a 1 way make is just not an acceptable number.
Refer to John's recent mail about AIM7 performance too.
>
> Going to -j4, elapsed time went from 808sec -> 254sec (4:14)
> which is a scaling of ~3.2 times ... if -j4 had had an
> elapsed time of 202sec (3:22) then it would have been perfect
> scaling. But add in the fact that -j4 vs. NO-J would introduce
> additional complexity in the make itself, which now spawns
> more processes, etc. ... this shows in the increased system time
> for -j4; of course the extra system time also counts increased
> lock contention. For this particular workload (primarily user-mode,
> little-or-no-shared parallelism), -j4 did pretty good.
>
> Finally, -j6 & -j8 may not discover any extra parallelism
> in the makes ... one gross way to check this is to watch "top"
> and see how many "cc" processes are running on an average.
> If it stays at 4, then -j4 is all that'll work.
>
j6 is the best, as you can make out from the results. j4 and j8
both do worse. (I guess that's why make -j6 is accepted as a good
stress test in the linux world).
I have absolutely no issues with how the compile time scales with
higher cpu (ie j-value) counts.
Kanoj
|