> On 7 Apr 96 at 1:10, Gerard Roudier wrote:
>
> <snip>
>
> > It'is not possible to see the advantage of Tagged Queue with the benchmarks
> > like Bonnie, Iozone, Test suite and Hdparm.
> > I have done some simulations of multi-threaded hard disk load.
> > I often have observed a gain less than 5 %.
> > However, I have measured a gain of 25 % elapsed time while copying lots of
> > files between 2 partitions of the same HD (Dec 1 giga byte).
> > It seems that HD firmware is very important for Tagged Queue.
>
> I've just enabled Taged Queueing on my AHA2940 (aic7xxx), and timed
> doing several full kernel compiles with it enabled and disabled.
> Enabling Tagged Queues actually seems to lower performance by around
> 3 or 4%. Are there any benchmarks out there to see how much the speed
> is actually increasing/decreasing by?
>
> The disk was a ST410800N, one of two on the internal SCSI bus. The
> machine was rebooted before every test, and only had a single user
> running on it.
>
Do the following simple test:
- Use a driver that can queue at least 4 commands to a device.
(I dont know if the aic7xxx driver can do that)
- Create 4 files of about 25 MB in the same partition.
- Run the following script:
time cat File1 >/dev/null &
time cat File2 >/dev/null &
time cat File3 >/dev/null &
time cat File4 >/dev/null &
- And wait and see
With my HDs (IBMS12 and CONNER 1080S), speed is improved
of about 3 to 4% when I enable Tagged Command Queuing.
When you have heavy disk load with tagged queue enabled, you may get
timeouts since the device reorder commands as it want.
Only ncrBsd2Linux and the Buslogic driver are protected against such
timeouts.
In order to avoid timeouts due to command reordering with the aic7xxx
driver, edit drivers/scsi/sd.c et set SD_TIMEOUT to about 30 seconds.
Regards, Gerard.