On Thu, Jan 31, 2008 at 11:12:45PM -0500, Victor Duchovni wrote:
> On Fri, Feb 01, 2008 at 01:15:09PM +1300, Peter Gutmann wrote:
> > If anyone's interested, I did an analysis of this sort of thing in an
> > unpublished draft "Performance Characteristics of Application-level Security
> > Protocols", http://www.cs.auckland.ac.nz/~pgut001/pubs/app_sec.pdf.  It
> > compares (among other things) the cost in RTT of several variations of SSL 
> > and
> > SSH.  It's not the TCP RTTs that hurt, it's all the handshaking that takes
> > place during the crypto connect.  SSH is particularly bad in this regard.
> Thanks, an excellent reference! Section 6.2 is most enlightening, we were
> already considering adopting HPN fixes in the internal OpenSSH deployment,
> this provides solid material to motivate the work...

To be fair, the "handbrake" in SFTP isn't -- the clients and servers
should be using async I/O and support interleaving the transfers of many
files concurrently, which should allow the peers to exchange data as
fast as it can be read from disk.

The same is true of NFS, and keep in mind that SFTP is more of a remote
filesystem protocol than a file transfer protocol.

But nobody writes archivers that work asynchronously (or which are
threaded, since, e.g., close(2) has no async equivalent, and is required
to be synchronous in the NFS case).  And nobody writes SFTP clients and
server that work asynchronously.  But, we could, and we should.

And the handbrake in the SSHv2 connection protocol has its rationale as
well (namely to allow interactive sessions to be responsive).  As
described in Peter's paper, it can be turned off, effectively.  It's
most useful when mixing interactive sessions and X11 display forwarding
(and port forwarding which don't involve bulk data transfers).  It's
most useless when doing bulk transfers.  So use separate connections for
bulk transfers.


The Cryptography Mailing List
Unsubscribe by sending "unsubscribe cryptography" to [EMAIL PROTECTED]

Reply via email to