Re: Loopback performance from kernel 2.6.12 to 2.6.37

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 08 novembre 2010 à 12:04 +0100, Eric Dumazet a écrit :
> Le lundi 08 novembre 2010 à 11:58 +0100, Jesper Dangaard Brouer a
> écrit :
> > On Fri, 2010-11-05 at 21:29 +0100, Eric Dumazet wrote:
> > > Le vendredi 05 novembre 2010 à 11:49 +0100, Jesper Dangaard Brouer a
> > > écrit :
> > > > Hi Eric,
> > > > 
> > > > A colleague send me a link to someone who has done some quite extensive
> > > > performance measurements across different kernel versions.
> > > > 
> > > > I noticed that the loopback performance has gotten quite bad:
> > > > 
> > > > http://www.phoronix.com/scan.php?page=article&item=linux_2612_2637&num=6
> > > > 
> > > > I though you might be interested in the link.
> > > > 
> > > > See you around :-)
> > > 
> > > Hi !
> > > 
> > > Problem is : I have no idea what test they exactly use,
> > > do you have info about it ?
> > 
> > Its called the Phoronix test-suite, their website is:
> > http://www.phoronix-test-suite.com/?k=home
> > 
> > On my Ubuntu workstation their software comes as a software package:
> >  sudo aptitude install phoronix-test-suite
> > 
> > They seem to be related to the test/review site:
> > http://www.phoronix.com/
> > 
> > 
> > 
> > > This probably can be explained very fast.
> > 
> > The loopback test seems to be the only real networking test they do.
> > It looks like they just copy a very big fil via loopback, and record the
> > time it took... quite simple.
> > 
> > Their tests seems to be focused on CPU util/speed, graphics/games.
> > 
> > 
> > The thing that caught my attention, was that they seemed interested in
> > doing performance regression testing on all kernel versions...
> > 
> > So, I though, it would be great if someone else would do automated
> > performance regression testing for us :-),  Too bad they only have a
> > very simple network test.
> > 
> > 
> 

> 

CC netdev, if you dont mind.


Their network test is basically :

netcat -l 9999 >/dev/null &
time dd if=/dev/zero bs=1M count=10000 | netcat  127.0.0.1 9999

They say it takes 38 seconds on their "super fast" processor

On my dev machine, not super fast (E5540 @2.53GHz), I get 8 or 9
seconds, even if only one CPU is online, all others offline.

Go figure... maybe an artifact of the virtualization they use.

I suspect some problem with the ticket spinlocks and a call to
hypervisor to say 'I am spinning on a spinlock, see if you need to do
something useful', or maybe ACPI problem (going to/from idle)