Re: issues multitasking during disk io

[Marcus Watts <mdw_(_at_)_umich_(_dot_)_edu>]

"resfilter" <resfilter_(_at_)_uniserve_(_dot_)_com> writes:
...
> for anyone to say that it's normal that extracting a single source archive
> can cause a program like ps to be barely able to execute, just because i'm
> using ide.. is simply rediculous.
> 
> i'm not imagining this:  i tried freebsd, and the problem went away.  i
> tried linux, and the problem went away. i moved the drive, operating system
> and all, over to a pentium 133, and the problem went away.
...

When you were using tar, you were also using gunzip/bunzip2.  Both of
those (especially the latter) are cpu hogs, and you're going through
the filesystem for all of your disk I/O.  I don't know how to separate
the effects of all that from the effects of the underlying disk
hardware in the output that you got.

I'd be very interested to see the results of this:
	dd </dev/rwd0c >/dev/null bs=32k count=100000
with
	iostat 10
running before, during & after.  The dd command will read the
raw disk, with a reasonable sized I/O buffer.  That will minimize seeks
and command reinstruct times, and give you the best measure of I/O
transfer speeds.  Optimally, you should expect to see I/O speeds that
are near the theoretical the drive is capable (given disk rotation
speed and physical geometry [which is almost certainly *not* what the
BIOS thinks it is, and also probably not what "atactl" reported]).  You
should also expect to see CPU utilization that is near 0 while doing
this (ie, say, 98-99% idle.)

It kinda sounds like somehow you are doing programmed I/O not DMA to
get to your IDE drives.  I'm at a loss to account for how that could
happen; looks like openbsd found "intel piix" hardware and did the
right thing.  But, if you get lousy results from "dd", then this is
probably where you need to look next.  If you get good results from
"dd", then that would point to something else, perhaps pipes, tar,
bunzip2, the filesystem, -- and you can try those parts out separately
to try to narrow your problem down.

				-Marcus Watts