One of the reasons for the timing instability is because of the
usleep() calls. This allows some other process to schedule, and thus
disrupts the I-cache and uop caches. With the usleep() calls:
i7-4900MQ# schedtool -R -p 1 -e /tmp/fast_mix2_49
fast_mix: 213 fast_mix2: 280
fast_mix: 336 fast_mix2: 356
fast_mix: 174 fast_mix2: 392
fast_mix: 202 fast_mix2: 403
fast_mix: 152 fast_mix2: 280
fast_mix: 212 fast_mix2: 403
fast_mix: 213 fast_mix2: 403
fast_mix: 213 fast_mix2: 392
fast_mix: 202 fast_mix2: 403
fast_mix: 191 fast_mix2: 392
... and without the usleep calls:
i7-4900MQ# schedtool -R -p 1 -e /tmp/fast_mix2_49
fast_mix: 146 fast_mix2: 347
fast_mix: 157 fast_mix2: 90
fast_mix: 78 fast_mix2: 90
fast_mix: 78 fast_mix2: 89
fast_mix: 78 fast_mix2: 90
fast_mix: 78 fast_mix2: 90
fast_mix: 90 fast_mix2: 90
fast_mix: 79 fast_mix2: 90
fast_mix: 90 fast_mix2: 89
fast_mix: 79 fast_mix2: 90
I had originally added the usleep calls() in my test infrastructure to
more accurately disable the uop cache effects, since we are going to
be called from an interrupt handler, not in a loop. But anyway, this
is one of the reasons for the differences that you were seeing with
your benchmarking framework and mine, and why micro-benchmarking can
be so hard to get right. :-)
(BTW, this is your original mixer; I haven't tried playing with your
modified Skein-like core round yet.)
- Ted
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
