linux buffers sizes and auto-tuning
[changed subject:]
On 2 Déc, Kevin Gibbs wrote:
> Nope the web100/net100 projects are completely different than the Linux
> autotuning.
>> Or Linux has other algorithms? How could I know it's enabled or not?
> Check out some of the links off the Iperf page about tuning.
I think the best explanation is this one:
<http://www.csm.ornl.gov/~dunigan/net100/auto.html>
> I am not sure
> off hand where and if you can globally disable the autotuning in the
> kernel.
Yes, you can. You just have to specify a "default" buffer size by hand.
You can easily see this in the send and receive code here and here:
<http://lxr.linux.no/ident?i=SOCK_SNDBUF_LOCK>
<http://lxr.linux.no/ident?i=SOCK_RCVBUF_LOCK>
In fact, I think there is not really "default" send or receive buffer
sizes in linux 2.4, since the default is... auto-tuning ! But there are
mininums and maximums, see net100 link above and kernel docs.
PS Chong: are you sure not having root access on one of the box, does
not limit tuning and thus performance in some way ?
PS Kevin: could your support staff migrate this mailing-list to a system
_with archives_ ? It's a pity all the information we currently exchange
gets lost.
Sympa for instance <http://www.sympa.org/> rocks
:-)
Cheers,
--
Marc Herbert
Ph.D. Intern
INRIA / SunLabs Europe
--
Mitch Kutzko | mitch --at-- dast.nlanr.net | mitch --at-- ncsa.uiuc.edu | 217-333-1199
http://hobbes.ncsa.uiuc.edu/