HPN SSH

[Chris Rapier]

Peter W. Osel wrote:
> Hello,
> 
> I know that this has been asked before, just wanted to mention that I,
> too, would like to see the HPN SSH functionality incorporated in the
> standard OpenSSH.

As the developer of the HPN-SSH patch I'd like to say that for *most* 
people the new version of OpenSSH will probably meet most of their 
needs. The developers increased in the internal window size to 2MB 
(which from what I can tell leads to a 1MB effective size in BDP 
equations). This *significantly* increases performance over the previous 
size.

However, for people on faster networks this is still a bit small and 
HPN-SSH still has a role there (ex. On a patch between Switzerland and 
the US OpenSSH4.6 peaked at 0.69MB/s, OpenSSH peaked at 17.7MB/s and 
HPN-SSH at 58.3 MB/s(though it was CPU bound at this point))

> Would there be technical disadvantages integrating the changes?

Initially there were performance problems with LAN transfers between two 
HPN enabled hosts. I think the latest version of the code (hpn12v19) has 
dealt with that. The larger issue is that for the majority of users - 
who tend to be poorly connected - the patch is probably overkill. That 
and the integration of the NONE cipher switch likely makes it dead in 
the water.

Also, the HPN patch code is a bit of a moving target at this point. 
HPN13 is undergoing testing now and represents a more fundamental 
departure from the code base than even HPN12 did. I believe we've 
reached the limit of what buffer tuning can provide and we've been 
exploring other directions for optimization. Hopefully we'll have 
something available for a public test by January.

> I know we are all pretty busy, but I would certainly spend time to help,
> e.g. with testing, documentation, etc.

I'll be continuing development on the HPN patch for sometime. I've 
brought on a couple more people. We have some interesting plans for the 
next couple of years but I believe, and I think the OpenSSH developers 
agree with me on this, that there probably won't be a wholesale 
integration in the foreseeable future. Honestly, this is likely for the 
best. It gives me a bit more freedom to try out different techniques 
which the devel team, if they find any of them useful, can integrate 
when they can.

Chris Rapier
http://www.psc.edu/networking/projects/hpn-ssh