On Wed, May 22, 2002 at 06:28:18AM -0700, Wolfgang Rupprecht wrote:
> 
> I've just noticed a strange performance anomaly in the disk transfer
> speed (under -current/x86 w. Promise ATA-100 and Seagate Barracuda
> IV's, with wd1 being an IDE "slave" of wd0).
> 
> If I set up a dd transfer from a raw disk partition to /dev/null I get
> ~40Mbytes/sec.  In theory ATA-100 can support 2 of these xfers going
> full bore, and I was expecting to see the transfer speed drop a bit.
> What I wasn't expecting was to see the disk xfer drop in less than
> half to ~13MBytes/sec.
> 
> Transfer from wd0 only:
> 
>              wd0             wd1            cpu
>    KB/t t/s MB/s   KB/t t/s MB/s us ni sy in id
> 
>   64.00 640 39.98   0.00   0 0.00  2  0  5  0 93
>   63.22 643 39.67   0.00   0 0.00  0  0  1  0 99
>   64.00 639 39.91   0.00   0 0.00  3  0  2  0 95
>   64.00 639 39.91   0.00   0 0.00  0  0  1  0 99
> 
> Transfer from wd1 only:
> 
>              wd0             wd1            cpu
>    KB/t t/s MB/s   KB/t t/s MB/s us ni sy in id
> 
>    0.00   0 0.00  64.00 640 39.98  1  0  0  0 99
>    0.00   0 0.00  64.00 635 39.67  2  0  7  0 91
>    0.00   0 0.00  64.00 639 39.91  1  0  6  0 93
>    8.00   2 0.00  64.00 640 39.98  4  1  2  0 93
> 
> Transfer from both wd0 and wd1:
> 
>              wd0             wd1            cpu
>    KB/t t/s MB/s   KB/t t/s MB/s us ni sy in id
> 
>   64.00 200 12.50  64.00 209 13.0   2  1  0  0 97
>   64.00 210 13.12  64.00 211 13.1   0  0  2  0 98
>   64.00 222 13.86  64.00 221 13.8   0  0  0  0 100
>   63.69 177 11.02  64.00 270 16.8   3  0  1  0 96
> 
> Anyone have a good model for what is going on here?

IDE doesn't do disconnect/reselect, so when you have 2 concurent one drive
is idle while the other one is working. So in your test it's expected
to have the per-drive transfer rate drop by half.
I'm not sure why it's more than half. Maybe the extra delay between 2 transfers
prevents the drive from doing some kind of "streaming" ?

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
--