Date: Thu, 8 Mar 2001 00:42:39 +1100 (EST) From: Bruce Evans <bde@zeta.org.au> To: Mark Murray <mark@grondar.za> Cc: Maxim Sobolev <sobomax@FreeBSD.ORG>, dougb@FreeBSD.ORG, current@FreeBSD.ORG Subject: Re: harvest_interrupt=YES slows down machine Message-ID: <Pine.BSF.4.21.0103080020001.2722-100000@besplex.bde.org> In-Reply-To: <200103071246.f27Ck6R59376@gratis.grondar.za>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, 7 Mar 2001, Mark Murray wrote:
> > > Apart from a ridiculously low maxusers (you have 10, I recommend 128),
> > > I'm not sure what the problem is.
> >
> > I do not see why it is "ridiculously low". Even GENERIC recommends 32, while
> > this is a small system intended to be used by only one person, so I do not see
> > any problem with it. I never had any `out of descriptors' or `can't fork'
> > during my routine operations on this box (2.5 years).
>
> Ok - then please leave it at 32.
I use 10 with no problems.
> If you want me to help you, please help me get good info.
>
> It's very important to me _know_ wether this is a boot slowdown or
> a generic - assertions are not good enough, I need hard facts.
>
> My own laptop (A Toshiba Libretto 110CT) does a make world in
> about 8 hours (and it always has). Interrupt harvesting has not
> made a noticable difference (I have not been keeping records, but
> an overnight build has not yet progressed into my breakfast).
Just do something that causes a lot of interrupts that go through the
random harvester. E.g.:
dd if=/dev/ad0 of=/dev/null
causes 7750 interrupts/sec here (on a Celeron 366 overclocked to
522). The random task takes 100% of the available cpu cycles. This
slows down cpu-bound processes by a factor of about 3.5. With a block
size of 64k instead of the default of 512, this causes only 300
interrupts/sec. The random task takes a measly 27% of the cpu to
process these. It can apparently only handle about 10 interrupts/second
with a reasonable overhead (1%).
Interrupt harvesting doesn't make much difference to makeworld because
makeworld is cpu-bound. I estimate it to be about 2% here (20 interrupts
per second for disk i/o to local disks. It would be a lot more for nfs).
Bruce
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.BSF.4.21.0103080020001.2722-100000>