Re: [Bug #11308] tbench regression on each kernel release from 2.6.22 -> 2.6.28

[Ingo Molnar <mingo-X9Un+BFzKDI@public.gmane.org>]

* Ingo Molnar <mingo-X9Un+BFzKDI@public.gmane.org> wrote:

> > 4% on my machine, but apparently my machine is sooooo special (see 
> > oprofile thread), so maybe its cpus have a hard time playing with 
> > a contended cache line.
> >
> > It definitly needs more testing on other machines.
> >
> > Maybe you'll discover patch is bad on your machines, this is why 
> > it's in net-next-2.6
> 
> ok, i'll try it on my testbox too, to check whether it has any effect 
> - find below the port to -git.

it gives a small speedup of ~1% on my box:

   before:      Throughput 3437.65 MB/sec 64 procs
   after:       Throughput 3473.99 MB/sec 64 procs

... although that's still a bit close to the natural tbench noise 
range so it's not conclusive and not like a smoking gun IMO.

But i think this change might just be papering over the real 
scalability problem that this workload has in my opinion: that there's 
a single localhost route/dst/device that millions of packets are 
squeezed through every second:

 phoenix:~> ifconfig lo
 lo        Link encap:Local Loopback  
           inet addr:127.0.0.1  Mask:255.0.0.0
           UP LOOPBACK RUNNING  MTU:16436  Metric:1
           RX packets:258001524 errors:0 dropped:0 overruns:0 frame:0
           TX packets:258001524 errors:0 dropped:0 overruns:0 carrier:0
           collisions:0 txqueuelen:0 
           RX bytes:679809512144 (633.1 GiB)  TX bytes:679809512144 (633.1 GiB)

There does not seem to be any per CPU ness in localhost networking - 
it has a globally single-threaded rx/tx queue AFAICS even if both the 
client and server task is on the same CPU - how is that supposed to 
perform well? (but i might be missing something)

What kind of test-system do you have - one with P4 style Xeon CPUs 
perhaps where dirty-cacheline cachemisses to DRAM were particularly 
expensive?

	Ingo