Re: hpn_buffer_size

[Thor Lancelot Simon <tls%panix.com@localhost>]

On Fri, Apr 27, 2012 at 08:10:07PM +0200, Petar Bogdanovic wrote:
> Hi,
> 
> I'm trying to figure out, why two nearly identical netbsd-6 hosts have
> troubles using hpn.  They are connected through a 5000 Kb/s link and the
> following (not very scientific) test:
> 
>       dd if=FILE bs=1m count=2 | ssh HOST dd bs=64k of=/dev/null
> 
> yields different results when defaults are in place (430KB/s) and when
> HPNDisabled=yes and/or HPNBufferSize=2048 (590KB/s).

I do think you need to test with larger files.  But the other thing that
occurs to me is that you're now using a kernel that automatically sizes
the TCP socket buffers; HPN has a tunable that needs to be set right if
this is the case (RcvBufferPolling?  Something like that.) which you
might want to investigate.

In fact, I think I'd first make sure it were set appropriately for a
kernel that _does_ auto-size the socket buffers, then try setting it
the other way and using systctl to set the socket buffer sizes by hand
(you will need to do this on *both* ends and restart the ssh daemon
process after each of these changes) to see whether the problem is either
A) HPN not playing nice with our socket buffer autosizing code due to wrong
default settings, or perhaps B) our socket buffer autosizing code being
broken.

Thor