On Tue, Nov 20, 2007 at 09:50:56PM +0800, Herbert Xu ([EMAIL PROTECTED]) wrote:
> On Tue, Nov 20, 2007 at 04:35:09PM +0300, Evgeniy Polyakov wrote:
> >
> > On Tue, Nov 20, 2007 at 08:27:05AM -0500, Andrew Gallatin ([EMAIL
> > PROTECTED]) wrote:
> > > Hmm.. rather than a global tunable, what if it was a
> > > network driver managed tunable which toggled a flag in the
> > > lro_mgr features? Would that be better?
> >
> > What about ethtool control to set LRO_simple and LRO_ACK_aggregation?
>
> I have two concerns about this:
>
> 1) That same option can still be turned on by distros.
FC and Debian turn on hardware checksumm offloading in e1000 and I have
a card where this results in more than 10% performance _decrease_.
I do not know why, but Im able to run script which disables it via
ethtool.
> 2) This doesn't make sense because the code is actually in the
> core networking stack.
It depends. Software lro can be controlled by simple procfs switch, but
hardware one? I recall it was number of times pointed that hardware LRO
is possible and likely being implemented in some asics.
> I'm particular unhappy about 2) because I don't want be in a
> situation down the track where every driver is going to add this
> option so that they're not left behind in the arms race.
For software lro I agree, but this looks exactly like gso/tso case and
additional tweak for software gso. Having it per-system is fine, and I
believe no one should ever care that some distro will do bad/good things
with it. Actually we do have so much tricky options in procfs already
which can kill performance...
--
Evgeniy Polyakov
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
