Re: siis/atacam/ata/gmirror 8.0-BETA3 disk performance

Thank you for your informative response.

Alexander Motin wrote:
> Derek (freebsd lists) wrote:
>> I've been testing the new siis driver, and I have found no
>> appreciable performance change, using dd as a measure of "raw"
>> performance.
> 
> On linear read/write, without port multipliers used, performance of siis 
> driver is not so much differs from legacy driver. The most of it's 
> benefits are NCQ, FIS-based switching and command queuing affect mostly 
> highly parallel random workload.
> 

That's good news, in that the behaviour is expected.

>> I get about 40MB/s read, and 30MB/s write.  See attached
>> bench-*.txt files.
> 
> It's too small, indeed. Actually, there are two different kinds of siis 
> compatible devices: SiI3124 and SiI3132/SiI3531. SiI3124 is more 
> expensive PCI-X card, sometimes it goes with built-in PCIe x4 bridge. To 
> operate effectively it needs effective bus. Inserting it to usual PCI or 
> PCIe x1 slot kills any hope. SiI3132/SiI3531 same time are cheap PCIe x1 
> cards. Nobody have ever seen them giving more 130-150MB/s, even looking 
> that PCIe x1 should give 2.5Gb/s.
> 

I've got the SiI3124, plugged into a PCI bus.  While I know PCI 
is really slow by today's standards, I was hoping to see 
performance closer to the bus limits.

My rates are 30% of the maximum "theoretical" bus utilization 
(33Mhz*32-bits).  The figures you give for PCIe x1 are around 40% 
of maximum (on the low end of the figures) utilization, so we're 
in the same neighbourhood.

> To completely load gmirror on read operations, you may need to run two 
> dd's same time. Also make sure, that your gmirror runs in round-robin 
> mode. Default split mode, which should help with linear read, is IMHO 
> ineffective, at least with default MAXPHYS and slice values.
> ...

I'll play around with this as well, I'm not too interested in 
tweaking stuff much in production, but it would be interesting to 
see just how this performs.

In production, I plan to use the gmirror for swap, and zfs for 
everything else.  It seems that for swap usage, linear read isn't 
a great model.

Thanks again!
- Derek