On Fri, Feb 05, 2010 at 12:04:50PM +0100, Sebastiano Pomata wrote:
> On Fri, 5 Feb 2010 10:39:03 +0100
> Claudio Jeker <cje...@diehard.n-r-g.com> wrote:
> 
> > On Fri, Feb 05, 2010 at 08:59:44AM +0100, Sebastiano Pomata wrote:
> > > On Fri, 5 Feb 2010 02:05:49 +0000 (UTC)
> > > Stuart Henderson <s...@spacehopper.org> wrote:
> > > 
> > > > On 2010-02-04, Sebastiano Pomata <sebastianopom...@tiscali.it>
> > > > wrote:
> > > > > As doublechecking, I tried with another fast server inside the
> > > > > wan network of our academy, and I'm getting almost the same
> > > > > results (while absolute speeds are different from before, the
> > > > > gap is almost the same in magnitude).
> > > > >
> > > > > I've read the page about tcptune, it's pretty clear now (values
> > > > > are almost the same I edited), still not having clear why on
> > > > > default OpenBSD the transfer rates are so low.
> > > > 
> > > > we try to have safe defaults for the varioous machines/arch that 
> > > > can run OpenBSD.
> > > > 
> > > > we would need some kind of auto-tuning to incrrease the defaults,
> > > > and don't have that yet.
> > > > 
> > > 
> > > I really can understand this, for the sake of system portability
> > > and so on. Anyway, I really hardly understand why, without touching
> > > any of the default settings, download rate from every server would
> > > never overcome the value of 400 kB/s. Is it all due to the tcp
> > > windows size?
> > > 
> > 
> > # sysctl net.inet.tcp | grep space
> > net.inet.tcp.recvspace=16384
> > net.inet.tcp.sendspace=16384
> > # ftp ftp://mirror.switch.ch/pub/OpenBSD/snapshots/i386/install47.iso
> > ...
> > Retrieving pub/OpenBSD/snapshots/i386/install47.iso
> > local: install47.iso remote: install47.iso
> > 150 Opening BINARY mode data connection for install47.iso (255594496
> > bytes).
> > 100% |**************************************************|   243 MB
> > 00:33    
> > 226 File send OK.
> > 255594496 bytes received in 33.38 seconds (7.30 MB/s)
> > 221 Goodbye.
> > 
> > Yep, only 400kB/s, no wait... The use of every should be considered
> > carefully. Yes, our default window size limits download speed. It is
> > known and there is work ongoing to resolve this in a better way then
> > just bumping the limit.
> > 
> 
> There's no polemical intent in my messages, I'm just trying to debug a
> situation where I'm totally clueless on what's going on.
> Claudio, I tried your tcp.{send,recv}space settings (default I think)
> and then downloaded your very same file, stopping at about 400 kB/s.
> Then increased the same parameters both to 262144 and download speed
> bumped at over 2.5 MB/s. I can understand that bumping the window size
> could not be the only solution, in fact I still get ridiculous upload
> speeds from the OpenBSD box.
> Did you make any other tuning to the default install to get that high
> download speed?

I did not tune anything on that box. The big difference is that the server
is easy reachable via direct peerings. So the bandwidth is 100Mbps (the
speed of my workstations interface) and the delay is < 1ms.
Sure if the delay * bandwidth product gets bigger then 16k then it is not
possible to use all the resources in a single stream. In that case, if you
so desperate to get more speed out of a single TCP stream, sure bump
the net.inet.tcp.recvspace.

-- 
:wq Claudio

Reply via email to