Re: Reverse port forward channel performance

[Gearoid Murphy <gearoid.murphy@xxxxxx>]

Hi Again,

The problem was with my implementation, I'm using a multi-threaded 
implementation, one thread for reading, another for writing. A mutex is 
used to ensure serial access to the channel, the read thread used a 
non-blocking read to check the recv buffer, then it slept for 20 
milliseconds.
When I reduced the size of this sleep to 500 microseconds, the 
throughput jumped to ~8MB.

My apologies for this distraction.

 - Gearoid

On 30/01/12 11:37, Gearoid Murphy wrote:
> Hi Aris,
>
> I'm not sure if it's significant but the logs from the port forwarding 
> implementation show that the function 'channel_rcv_change_window' 
> makes changes to the window size but in the reverse port fowarding 
> implementation, changes to the window size are executed via the 
> 'grow_window'
> function.
>
>  - Gearoid
>
> On 27/01/12 20:42, Aris Adamantiadis wrote:
> > Hi Gearoid,
> >
> > It's possible that channel windows are not initialized the same way 
> > in forward and reverse forwarding. This could result in bad 
> > performances if the window is too small.
> >
> > I'll check if I can find anything.
> >
> > Aris
> >
> > Le 27/01/12 18:16, Gearoid Murphy a écrit :
> > > Hi,
> > >
> > > We've been observing poor throughput (230KB/s) on libssh channels
> > > generated from reverse port forwarding (ssh_forward_accept) via an
> > > OpenSSH server.
> > >
> > > Conversely, when using a channel generated through a normal port
> > > forwarding call (ssh_channel_open_forward), we observe far greater
> > > throughput (8MB/s).
> > >
> > > Has anyone observed or experienced similar issues?, is there some
> > > configuration aspect of the channel setup which needs to be modified?
> > >
> > > libssh 0.5.2-1
> > > Linux 3.0.0-15-generic x86_64 x86_64 x86_64 GNU/Linux
> > > openssh-server 1:5.8p1
> > >
> > > Thanks
> > > - Gearoid