[Openib-windows] SRP tuning

[Yossi Leybovich]

> -----Original Message-----
> From: Fabian Tillier [mailto:ftillier.sst at gmail.com] 
> Sent: Tuesday, December 05, 2006 9:08 PM
> To: Yossi Leybovich
> Cc: openib-windows at openib.org
> Subject: Re: [Openib-windows] SRP tuning
> 
> Hi Yossi,
> 
> On 12/5/06, Yossi Leybovich <sleybo at dev.mellanox.co.il> wrote:
> >
> > I set:
> >
> > SCSI_MAXIMUM_TRANSFER_SIZE = SP_UNINITIALIZED_VALUE.
> > SrbExtensionSize = 3552
> >
> > I played with the NumberOfPhysicalBreaks:
> > The StorPort driver set the of NumberOfPhysicalBreaksvalue 
> to 17 ( I 
> > don't know why , Can I configure?)
> 
> 17 is the worst case number of breaks to transfer 64KB of 
> data in a single I/O.  You should be fine increasing this (I 
> think the code already increases it based on the SRB extension size).
> 

DDK state
"...
If the port driver sets a value for this member, the miniport driver can
adjust the value lower but no higher. If this member is
SP_UNINITIALIZED_VALUE, the miniport driver must reset this member according
to the HBA's scatter/gather capacity, with zero representing no
scatter/gather support. 
..."

SO I don't think we can return larger s/g list size.

> > But even if I return 16/17 as my supported 
> NumberOfPhysicalBreaks it 
> > keep post IOs with 8 s/g list size (32K) [when I set it to 
> 4 ,storPort 
> > send IOs with 4 s/g , So the value did effect when I reduce it]
> >
> > Any idea why ?
> 
> What are you using to generate the I/O traffic?   Are you sure it's
> issuing larger than 32K I/O sizes?  I've used iometer before 
> with good results.
> 
I use Iometer as well , and test it with 64K -256K msg size.
Did you saw StorPort using SRB with 64K IO data?

> -Fab
>