ULE/SCHED_SMP diff for 7.0, buildkernel & thanks.
Jeff Roberson
jroberson at chesapeake.net
Tue Jul 17 18:46:01 UTC 2007
With regards to buildkernel times; I do not want to sacrafice performance
on other benchmarks to improve buildkernel. The problem is that 4BSD is
as agressive as possible at scheduling work on idle cores. This behavior
that helps one buildworld hurts on other, in my opinion, more important
benchmarks.
For example: http://people.freebsd.org/~jeff/sysbench.png
ULE is 33% faster than SCHED_4BSD at this mysql test. This is a direct
result of prefering to idle to make more efficient scheduling decisions.
ULE is also faster at various networking benchmarks for similar reasons.
I also believe that while the real time may be slower on buildworld the
system and user time will be smaller by a degree greater than the delta in
real time. This means that while you're building packages you have a
little more cpu time leftover to handle other tasks. Furthermore, as the
number of cores goes up things start to tip in favor of ULE although this
is somewhat because it's harder for even 4BSD to keep them busy due to
disk bandwidth.
Thanks everyone for testing. Can someone confirm that they have tested
with x86 rather than amd64? I will probably commit later today.
Thanks,
Jeff
On Tue, 17 Jul 2007, Claus Guttesen wrote:
>> > sched_ule:
>> >
>> > -j 3 buildkernel: 13:23
>> > -j 4 buildkernel: 12:38
>> > -j 5 buildkernel: 12:41
>> > -j 6 buildkernel: 12:47
>> >
>> > sched_4bsd:
>> > -j 3 buildkernel: 11:43
>> > -j 4 buildkernel: 12:02
>> >
>> > So sched_ule seems to handle more processes slightly better than 4bsd
>> > albeit it does it slower. ule's sweet spot is -j 4 and 4bsd is -j 3.
>> >
>>
>> 4bsd vs ULE
>>
>> -j 3 buildkernel: 11:43 vs -j 3 buildkernel: 13:23
>>
>> -j 4 buildkernel: 12:02 vs -j 4 buildkernel: 12:38
>>
>>
>> ULE is always slower?
>
> In my case yes.
>
> --
> regards
> Claus
>
> When lenity and cruelty play for a kingdom,
> the gentlest gamester is the soonest winner.
>
> Shakespeare
> _______________________________________________
> freebsd-current at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-current
> To unsubscribe, send any mail to "freebsd-current-unsubscribe at freebsd.org"
>
More information about the freebsd-current
mailing list