Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 11 Mar 2004 10:37:09 -0500
From:      Kenneth Culver <culverk@sweetdreamsracing.biz>
To:        Ian FREISLICH <if@hetzner.co.za>
Cc:        Kris Kennaway <kris@obsecurity.org>
Subject:   Re: I like SCHED_4BSD
Message-ID:  <20040311103709.kgcok0c0s0ksooco@www.sweetdreamsracing.biz>
In-Reply-To: <E1B1Lh0-0007Rm-00@hetzner.co.za>
References:  <E1B1Lh0-0007Rm-00@hetzner.co.za>

next in thread | previous in thread | raw e-mail | index | archive | help
Quoting Ian FREISLICH <if@hetzner.co.za>:

>> > o Better interactivity -- No mouse jerkiness, no sluggish screen updates =
>> when
>> >   switching between virtual desktops, etc.
>> >=20
>> > o Better scheduling!  I'm serious here.  Watching top under SCHED_ULE, I'm
>> >   seeing 10, 15, 20 seconds go by where ALL processes are sleeping.=20
>> >   Processes seem to be spending inordinate amounts of time in the "kserel"
>> >   state.  This, of course, doesn't happen with SCHED_4BSD.
>>
>> That this observation may well be bogus, because you're trying to
>> measure the scheduler using a tool that is itself run by the
>> scheduler, so the process stats you see may not be representative of
>> what is really happening on your system.
>
> That point is taken, however, on my old SMP box _ULE adds an extra
> ~5000 seconds of wall clock time to a 'make buildworld -j8' compared
> to _BSD 'make world -j8'.
>
> I did some tests about 3 weeks ago using bind-dlz and postgresql-7.4.1
> on this same machine trying libc_r and libkse using both schedulers.
> ULE results with either thread library were dismal ~.33 that of
> BSD: ULE gave about 17 authoritive lookups a second with BSD giving
> about 50.  The KSE library only dropped the rate by about 2 lookups
> per second on both schedulers which is essentially noise, but I had
> expected an improvement.
>
> Then there is the question about ULE and processor affinity which
> I posed about a week ago which nobody has answered.  I'll pose it
> here again:
>
> ------------------------
> I thought I'd give ULE a spin again since according to Robert Watson
> and David O'Brien's recents posts it appears that it might not to
> slow down my system.
>
> I noticed this oddity:
>
> dnetc in order to utilize both CPUs forks a child which also
> processes.  SCHED_4BSD seems to be aware that the second CPU is
> idle and the child or parent (don't really care which) migrates to
> the second CPU so that all CPU time is occupied on both CPUs.
>
> SCHED_ULE doesn't seem to migrate processes to an idle CPU and as
> a result one CPU in my system is 100% idle.
>
> last pid:   677;  load averages:  2.06,  1.69,  0.89    up 0+00:09:57 
>  09:22:29
> 40 processes:  3 running, 37 sleeping
> CPU states:  0.0% user, 50.0% nice,  0.4% system,  0.4% interrupt, 49.2% idle
> Mem: 25M Active, 25M Inact, 22M Wired, 56K Cache, 28M Buf, 111M Free
> Swap: 512M Total, 512M Free
>
>   PID USERNAME  PRI NICE   SIZE    RES STATE  C   TIME   WCPU    CPU COMMAND
>   624 ianf      139   20  1072K   884K RUN    0   4:00 48.44% 48.44% dnetc
>   636 ianf      139   20  1072K   884K CPU0   0   3:26 48.44% 48.44% dnetc
>
> Do processes have CPU afinity and is that afinity inherited by their
> children?  Is this a wise thing to do since as demonstrated here
> it is possible that all the CPU hogs may land up on 1 processor
> thereby pessimising runtime?
> ------------------------
>
> In my opinion ULE is not yet ready for prime-time.   Perhaps it
> doesn't affect UP or is even more efficient with UP, but I thought
> it was originally tabled as the new scheduler to make the SMP case
> more efficient, which it certainly does not.
>
> Ian
>
Even on UP, in my totally subjective tests, under a load, 4BSD seems a lot
faster. I haven't taken any real measurements of various times for buildworld
and all that, but things just "feel" faster with BSD.

Ken



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20040311103709.kgcok0c0s0ksooco>