Poor disk drive performance with FC3 RC3
sp at scali.com
Thu Nov 4 19:16:20 UTC 2004
On Thu, 2004-11-04 at 13:59 -0500, David Zeuthen wrote:
> On Thu, 2004-11-04 at 13:40 -0500, Steffen Persvold wrote:
> > What is the purpose of the "haldaemon" and why is it destroying my
> > system performance ?
> hal polls for media change on your optical drive; is your optical
> drive /dev/hdb by any chance? E.g. on the same cable as /dev/hda - I've
> had reports yesterday about hdparm giving bad numbers in this case.
> Basically hal does this (every two seconds)
> open("/dev/hdc", O_RDONLY|O_NONBLOCK|O_EXCL|O_LARGEFILE) = 0
> ioctl(0, CDROM_DRIVE_STATUS, 0x7fffffff) = 1
> Can some kernel person comment on why this is killing performance on
> the /dev/hda harddrive when the optical drive at /dev/hdb is polled?
Yes, for some reason the cdrom is attached to the same IDE channel as my
harddrive (cdrom = /dev/hdb). Not sure why Dell implemented it that way
though. Back in the good old days I learned that for best performance
you shouldn't really have two devices on the same IDE channel unless you
absolutely had to... However I don't think there is any way of changing
it, and a lot of Linux users have Dell laptops ;) ...
And your analysis matches with the behavior I see on the drive light.
When I start hdparm, the drive light stays solid for a while, but then
suddenly it stops and depending on how long the light stays solid my
transfer rate differs (and this is dependent on when I start the test
vs. the polling interval for hald). I guess the ioctl() to /dev/hdb
interrupts the /dev/hda transfer for some reason.
More information about the test