[linux-cifs-client] extremely slow throughput writing to remote shares

[Jeremy Allison]

On Mon, May 22, 2006 at 09:19:55PM +0200, Christian Pernegger wrote:
> >My experiments with cifs writing over gigabit showed reasonable utilization
> >but would be much better [...]
> 
> That sounds great, but more like further optimizaton of something
> which is basically working. I _had_ transfer rates in the same
> ballpark with FTP and was hoping for someone to tell me what might
> have broken / regressed / grossly misconfigured.
> 
> > The cifs improvement (around 2.6.15) for using iovecs on write (which
> >eliminated a large memory allocation on the client, which could sometimes
> >cause delays while the client mm hunted for eight free contiguous pages)
> >also helps.
> 
> Hmm, the machine is on 2.6.15 at the moment, has 2.6.16 a reasonable
> chance of fixing things? How much of an impact would this improvement
> have, assuming both participating machines have at least 1,5GB free
> RAM?
> 
> Normal FS writes go into a buffers, to be written out when pdflush
> feels like it. I have a feeling this whole caching mechanism is
> somehow bypassed here.

Hmmm - sounds like the FILE_WRITE_THROUGH -> O_SYNC bug I just
fixed in the server for 3.0.23. But looking in the CIFSFS code I never
see the FILE_WRITE_THROUGH flag being set....

Jeremy.