OpenSSH GSoC Project

[Damien Miller]

On Wed, 25 Mar 2009, Salvador Fandino wrote:

> > From: Damien Miller <djm at mindrot.org>
> > To: Salvador Fandino <sfandino at yahoo.com>
> > Cc: Ben Lindstrom <mouring at eviladmin.org>; openssh-unix-dev at mindrot.org
> > Sent: Wednesday, March 25, 2009 1:18:54 PM
> > Subject: Re: OpenSSH GSoC Project
> > 
> > Why bother? With the current packet size the overhead is something like
> > 30KB per GB of data transferred, but increasing this makes the protocol
> > perform worse for multiple concurrent operations (which we admittedly 
> > underutilise now) and less responsive to early aborts, etc.
>
> The problem is not overhead but latency, at least in theory.
>
> With current configuration, we have 32KB packet-length * 64
> max-packets-queued = 2MB of data on the "air", so for instance, if
> the communication line supports 4MB/s, but the roundtrip is 1 second,
> only 50% of the bandwidth will be used (AFAIK, the underlaying SSH
> transport does not use any ACK mechanism internally and so is not
> affected by round-trip, and TCP can handle that delays also).

Sure, but this can be trivially changed from either the commandline
or the code so it can hardly be used as an argument that sftp is
inherently slow.

It might be a good project for GSoC to make this automatically
tune itself.

-d