[Openib-windows] File transfer performance options

[Fabian Tillier]

Hi Paul,

On 9/6/06, Paul Baxter <paul.baxter at dsl.pipex.com> wrote:
> Lots of good advice and info snipped for brevity, but thanks Fab!
>
> > Anyhow, using SRP will likely require a cluster file system of some
> > sort if you plan on having multiple hosts share the target.  This adds
> > another dimension as cluster file systems can depend on specific HBA
> > functionality which may or may not exist in various SRP target
> > implementations.
>
> We have a single windows application wanting to read from four independent
> Linux nodes each with their RAID array available to the windows app.

Will the Linux nodes be writing to the RAID array too, or will the
disk access be limited to only that single Windows host?  If only the
Windows host will read the disk, you should be fine using SRP.  Each
Linux node would show up as an individual target.

> Hopefully this means I shouldn't need a cluster file system.

You should be fine without as long as the Linux nodes won't try to
write to the RAID array while the Windows host is using it.

> I'm not sure whether this makes the windows machine the SRP initiator and
> the four Linux nodes SRP targets or vice versa.

In this case, the Windows machine is the initiator - it issues all
requests for data, and the Linux nodes are the targets - they perform
the I/O requests.

> Pointers to SRP documentation gratefully received.

Here's a link to the spec.  This might lower level than what you
wanted.  I don't know if there are a lot of better docs out there.

http://www.t10.org/ftp/t10/drafts/srp2/srp2r00a.pdf

- Fab