kern/85820: 1.5 times slower performance with SCHED_ULE than SCHED_4BSD

[O. Hartmann]

Ceri Davies wrote:
> On Sun, Oct 08, 2006 at 09:54:53PM +0400, Maxim Konovalov wrote:
>   
>> On Sun, 8 Oct 2006, 17:20-0000, Ceri Davies wrote:
>>
>>     
>>> Synopsis: 1.5 times slower performance with SCHED_ULE than SCHED_4BSD
>>>
>>> State-Changed-From-To: open->closed
>>> State-Changed-By: ceri
>>> State-Changed-When: Sun Oct 8 17:19:36 UTC 2006
>>> State-Changed-Why:
>>> ULE is no longer the default scheduler, and no longer has a maintainer.
>>> This is an interesting test case though.
>>>       
>> I think better mark ULE bugs as suspended.  I have plans to take them
>> over.
>>     
>
> I don't intend to sweep them all.  I just didn't see a problem statement
> in this PR, and figured that it was due to the fact that ULE was default
> at the time the PR was raised.
>
> Feel free to reopen it if you disagree.
>
> Ceri
>   
Well, maybe I'm the only 'stupid' in this round, so be patient with me.
I never realized, that SCHED_ULE has been discarded from being the
default scheduler in FreeBSD 6.X, as I can remember, SCHED_ULE has been
promoted to be the default scheduler due to its improvements to SMP.
Yesterday I read about some bugs, but I read this in this list and
therefore I do not exactly know whether these problems are AMD64 related
or also be an issue on i386. Well, let me tell you some remarkeable
experiences I made since yesterday.
Besides my lab's work I uitilize a FreeBSD 6.2-PRE/AMD64 box with a
single core Athlon64 3500+ at 2,2 GHz. This box got really slow when
walking through 6.1-STABLE up to now and I tried to figure out why.
At my lab's desk I use an i386 based box running a P4 at 3,0 GHz with HT
capabilities, enabled both in BIOS and in the system, so this box is
configured as a SMP machine. On this box, also using SCHED_ULE, I did
never realize the significant performance drop as seen on AMD64.
After reading about the scheduler problems, I first changed SCHED_ULE to
SCHED_4BSD on my AMD64 box and realized a quite impressive performance
boost (something strange and not quantified, because I see this only
when working with Firefox, Thunderbird and disk access, which seem to be
smoother and quite faster than before).
On the lab's box, I changed also scheduler back to SCHED_4BSD, but I did
not realize this massive peroformance impact as I did on the AMD64 box.
Either this problem is highly related to UP kernel configs and/or UP
systems as well or it is only related to the 64Bit FreeBSD or only to
AMD CPUs, I do not know but would like to know more about that.

Well, 'feeling' a performance impact in both directions isn't a very
qualifying statement, so I would like to aks you whether you have some
suitable benchmarks I can perform on both boxes to confirm my
observations and give some numbers.
In the first days of FreeBSD 5.0 I remember myself having seen an
impressive performance boost using SCHED_ULE on a dual P3/933 Mhz box
acting as  a file- and databaseserver, but this advantage has obviously
gone over the past.

Well, Im not very close to the development issue, I'm only a FreeBSD
user for scientific purposes and therefore I would appreciate any public
informations on this subject. I guess many others around here still
being 'stuck' on a faulty SCHED_ULE, not knowing this scheduler is about
to be doomed and slowing down a box. Those guys also may act like me -
installing once a system and then migrating from one minor release to
the next via cvsupdate/buildworld (due to driver/hardware/support
issues) and still keeping a kernel config from days, when SCHED_ULE was
said to be the new ultimative scheduler even for SMP and UP boxes.

Well, I may be wrong, but it would be much more convenient having those
important informations being released more public , this will keep  guys
like me  from sending stupid PRs about slowed down boxes suspecting
other hardware or bugs elsewhere in the OS.

Thnaks in advance,
Oliver