Question about throughput

[Doug Ledford]

Mads Bondo Dydensborg wrote:
> I have tried to analyze what happens, and
> it seems that the aic7xxx driver will only issue a single read command at
> a time. When writing, multiple write commands will be issued.

The driver will issue as many reads to the device as the mid level and higher
level code sends to it.  Under normal conditions, large numbers of reads are
queued by the block device read-ahead code in the upper block layers, not by
the SCSI subsystem.  So, in order for linux to do the same thing for your
device, you would have to be accessing it through the block device layers
(aka, your drive appears as any other normal disk drive and you aren't using a
raw device to access it) and you would have to be doing sequential reads.  If
you meet those two items, then linux should be using the tagged queueing to
access your drive with multiple reads outstanding.

> Obviously, using the SCI network is not the same as the local case. There
> is a greater latency for operations.
> 
> My question is this; are there any way to force the aic7xxx driver to
> issue more read commands to a device?
> 
> Any help, hints, suggestions, comments, flames are welcome. Anything,
> really. :-) I must defend my results in 3 weeks, and would _really_ like
> to have solved this problem first.

I would try making a filesystem on the SCSI network device, then mounting that
filesystem, then run the bonnie program on that filesystem to see how it
performs under those conditions.  I assume that's not what you are doing now
;-)  Using bonnie on a mounted filesystem on a local disk usually results in
sequential reads being better than writes until writes eclipse the 85MB/s
mark, at which point they start to pass reads (on a dual PII-500 I think it
was, and the reads were running out of CPU power for searching through the
buffer cache before each read command was issued as I recall).

-- 

 Doug Ledford <dledford at redhat.com>  http://people.redhat.com/dledford
      Please check my web site for aic7xxx updates/answers before
                      e-mailing me about problems


To Unsubscribe: send mail to majordomo at FreeBSD.org
with "unsubscribe aic7xxx" in the body of the message