Re: Read buffer size in clientloop.c

On Tue, Jul 07, 2009 at 12:57:16PM +0200, Corinna Vinschen wrote:
> when trying to optimize socket transfer rates under Cygwin, it turned
> out that the underlying WinSock implementation is surprisingly sensitive
> to buffer sizes.  The latest Cygwin from CVS is now setting the socket
> receive/send buffers (SO_RCVBUF/SO_SNDBUF) to 64K, rather than keeping
> them at their default values of 8K which thwarts data transfers a lot.
> 
> While testing I still had the problem that for some reason the ssh read
> transfer rates were only a third up to a half of the write transfer
> rates.  It turned out that the culprit was ssh itself.  In clientloop.c,
> it defines read buffers of the size 8K.  Setting them to 64K under
> Cygwin raises the read transfer rates comparable to the write rates.
> 
> It occured to me that we talked about this already 3 years ago in a
> thread about HPN-SSH, just the performance numbers were different:
> http://marc.info/?l=openssh-unix-dev&m=114414372902485&w=2
> 
> So the question I have is again this:
> 
> Would it be ok to raise the buffers in client_process_net_input() and
> client_process_input() to 64K, maybe only on Cygwin?  Or to make the
> buffer sizes a configurable option?

If there's a measureable performance gain (which it sounds like there is)
then I personally have no objection to making it a compile time option.
Damien?  Maybe something like this?

Index: clientloop.c
===================================================================
RCS file: /home/dtucker/openssh/cvs/openssh/clientloop.c,v