how to test the performance of PIO and DMA
alvin at Maggie.Linux-Consulting.com
alvin at Maggie.Linux-Consulting.com
Mon Feb 4 21:19:16 PST 2002
hi mark
ditto on your comments... nobody should be using PIO mode...
for (other's) info ....
usually you can see a "*" next to the current drive settings
from "hdparm -iv /dev/hda"
- look for *udma5 ( for "100MByte/sec transfer" disks )
even if PIO is supported on the disk.. you dont need it or use it..
c ya
alvin
http://www.Linux-1U.net/Disks ... IDE vs SCSI ...
- note teh PIO and MultiWord DMA speed ...
and note the current Ultra-xxx dma modes..
On Mon, 4 Feb 2002, Mark Hahn wrote:
> > > > Title: how to test the performance of PIO and DMA
> > >
> > > use hdparm for simple easy tests and watch your resulting
> > > transfer speed
> >
> > You may also want to monitor your CPU load: there is a huge difference
> > in CPU overhead between PIO and DMA modes.
>
> like night and day. PIO is a dinosaur, and any vaguely modern machine
> doing PIO is criminally misconfigured.
>
> > The standard hdparm test is
> > too brief to see this,
>
> I have no idea why you say this: when I run
> /usr/bin/time hdparm -t /dev/hda
> on the desktop I'm sitting at, I see around 24 MB/s and 9% CPU
> in dma mode, versus 8.9 MB/s and 55% CPU after hdparm -d0
> (note that the %CPU numbers are extremely approximate: the actual
> ide-dma overhead is in the 3-5% range.)
>
> > Moreover, I think that DMA gives you CRC checking for data going over
> > the IDE cable while PIO does not. DMA is a clear winner, unless you
>
> udma, actually - plain old ide-dma doesn't do the CRC.
> all machines should be using udma on ide, and should be using
> ultra2 on SCSI (which adds CRC's as well.)
>
More information about the Beowulf
mailing list