[PATCH] Improving sftp client performance

[mouring at etoh.eviladmin.org]

On Sat, 5 Jan 2002, Tobias Ringstrom wrote:

> On Sat, 5 Jan 2002, Tobias Ringstrom wrote:
>
> > On Fri, 4 Jan 2002 mouring at etoh.eviladmin.org wrote:
> >
> > > No it was not in the CVS tree yet.  There was still discussions as to the
> > > implementation.  I have a version that does not do more more then two
> > > overlapping writes.
> >
> > I tried to vary the number of overlapping writes, but there is something
> > (not so) funny going on that I must understand first.  It seems to have
> > something to do with pipes/lo in Linux.  I'll be back...
>
> Eureka!  I found the real problem.  I had a feeling somthing was not right
> when I needed such a big number of overlapping requests.
>
> The problem is the Nagle algorithm.  Adding TCP_NODELAY to both ssh and
> sshd makes them fly.  Adding a flag to ssh that turns off Nagle would be
> easy.  I wonder what the best way is to to turn it off on the server side.
> One way is of course for the client to turn it off remotely.  Another way
> would be to turn it off per subsystem using the config file.  Yet another
> way would be for the subsystem to send a signal to its sshd.  What do you
> think?
>

This is a current topic else where on the two OpenSSH lists.   I believe
we don't want TCP_NODELAY for tty sessions, but maybe we can disable it
on all non-tty sessions.  That would handled the client side.   As to the
server side there is talk about using TCP_NODELAY everywhere.

Thoughts Markus? Steve? Damien?

- Ben