Increasing the server wsize/rsize

[Greg Banks]

On Thu, 2006-08-03 at 17:35, Neil Brown wrote:
> On Tuesday July 25, gnb at melbourne.sgi.com wrote:
> > On Tue, 2006-07-25 at 02:35, Chuck Lever wrote:
> > > On 7/23/06, Greg Banks <gnb at melbourne.sgi.com> wrote:
> > > > On Tue, 2006-07-04 at 20:12, Greg Banks wrote:
> > > > > On Tue, 2006-07-04 at 00:24, Chuck Lever wrote:
> > > > > > Only a small nit:  when adding these #defines, can you include
> > > > > > parenthesis around the definitions [eg. (6) and (4)] as you did with
> > > > > > the newly added #defines?
> > > > >
> > > > > Sure, see below.
> > > > >
> > > >
> > > > Are there any further comments on this patch?  If not I'll
> > > > formally submit the latest version to Neil.
> > > 
> > > Greg-
> > > 
> > > Did you decide not to move the auth header definitions?
> > 
> > No, I just forgot.  DoH!  New patch below.
> > 
> > --
> > Make knfsd support rsize/wsize of up to 1MB, over TCP.
> > The limit over UDP remains at 32K.  Also, make some of
> > the apparently arbitrary sizing constants clearer.
> > 
> 
> So where are we up to with this?  Should I consider this a formal
> submission yet???

I think it's ready and I was going to submit it in the next batch
of patches, as soon as I finish doing all the testing akpm wants.
So please consider it formally submitted.

> One thought I have is that this will unconditionally require 1 Meg of
> RAM to be allocated and pinned for each nfsd thread, which is slightly
> more than the current 40K or so.  Do we really want to do that?

Good point.  I'm spoiled by working on machines where the
minimum memory fit is 4 GiB ;-)

> Part of me say that this upper limit should be configurable and should
> default to something comfortable.
> Another part of me says that having tuning knobs is usually more of a
> pain than a gain.

Yep ;-)

> Maybe the auto-tuning should be based on ram size
>  < 256Meg -> default to 32K
>  < 2Gig   -> default to 512K
>  > 2Gig   -> default to 1Meg
> 
> or something like that.

Sure, works for me.  It would be a small incremental patch.

> Also, should svc_max_payload be 'static inline' in a header?
> It isn't large and is only called 7 times (I think).

Agreed.  I think I moved it into net/sunrpc/svc.c sometime during
the long conversation with Chuck.

Greg.
-- 
Greg Banks, R&D Software Engineer, SGI Australian Software Group.
I don't speak for SGI.