Performace issues with ssh through nat port

[rutt]

See this thread: http://www.vmware.com/community/click.jspa?searchID=2718957&messageID=542118

So, I got all that working as was able to ssh into my guest RHEL 4 guest through the NAT port 8889. Cool!

But once the guest gets busy, there seems to be a performance issue. I run a make with -j 4 under an emacs which is using X to talk to the host server. Response gets very very bad until the compile finishes.

Alternatively, I can use ssh from the guest console to make listen on the host:

ssh -R 22:localhost:8887 host

And then ssh into this port from the host:

ssh localhost -p 8887

Now this make under emacs works much better. You wouldn't think so since there is more than one trip over an ssh encrypted connection. But there it is.

I suspect this due to some issue in the vm itself, which needs to service both the interrupts of the guest and the nat forwarding.

Anyway, this might be the sort of issue the beta is meant to expose.

[rcardona2k]

I don't know if this is related but I opened an SR because I noticed nat traffic isn't returned from the Guest VM unless the host is sending data. In some way, nat forwarding works in a reciprocal manner. My example was, I enabled Microsoft Terminal Services RDP in a Windows VM, and the screen would not render unless I moved the mouse around. The data transfer from the guest to the host was obviously blocked. You could the see the screen paint in row "chunks", each time data was sent from the host-to-the-guest. It would stop when the host would not send data. I described the issue in the SR as a buffer holding issue.

I also experimented with changing the udp buffer timeout and I even added a 'fictious' tcp buffer setting in nat.conf and set them all to zero. Nothing changed the behavior I submitted in the SR.