Inconsistent IO performance
smithi at nimnet.asn.au
Sat Aug 14 08:02:18 UTC 2010
On Fri, 13 Aug 2010, Kevin Oberman wrote:
> > Date: Sat, 14 Aug 2010 11:06:58 +0800
> > From: TJ Varghese <tj at tjvarghese.com>
> > You're using a laptop with 2 HDDs, so does that mean you're using the
> > Ultrabay for the 2nd HDD?
> > Perhaps anything connected to that drops down to ATA33 (pure speculation on
> > my part) since it was designed for optical drives ...dmesg/atacontrol logs
> > would be useful here.
> > You may want to try dd with the of=/dev/null instead to remove the 2nd
> > variable and benchmark solely the read speed of the 1st hdd.
> For this test I get a very consistent 34.75 MB. (1000 10M
> blocks). Distressingly low when there is almost no seek activity.
> Nope, it is running at UDMA100 using a SATA-PATA converter. (The ICH6
> controller is SATA.) But that was a good idea.
For some sort of comparison, on a T23 (1133MHz P3-M, ICH3 PATA, Fujitsu
MHV2120AH 120GB 5400rpm UDMA100 drive, single device on channel, 768MB
RAM), I do better than that: # dd if=/dev/ad0 of=/dev/null count=1024
bs=10M gets, over two runs, just either side of 40,500,000 bytes/s.
Same result, just slightly better, with bs=512M count=20. (I also tried
bs=1024M count=10, but that was really dumb with 768MB RAM :)
But with if=/dev/ad0s4 (starts about 70% into the disk) just either side
of 29,200,000 bytes/s. All tests show 0.5-0.7% CPU, so hardly a factor,
and with that usage powerd kept the CPU at 733MHz.
Could be as you suggest, the SATA to PATA converter is a bottleneck?
I see mine has acoustic management on also; let us know if it matters?
Oh, and that's on 8.0R, about to be brought up to 8.1-S, if relevant.
More information about the freebsd-stable