Re: Improved UDP performance using 2.6.21-rc5-rt10

Re: Improved UDP performance using 2.6.21-rc5-rt10

[dave_sperry@ieee.org]

rt10 shows a big improvement over rt8

##################################################
etup                    Thruput      CPU% 
Nvidia
2.6.21-rc5-rt8           938          65%
   netperf @51
   hardirq @50
   softirq @50

Nvidia
2.6.21-rc5-rt10           938          32%
   netperf @51
   hardirq @50
   softirq @50

very nice for a "work in progress" 

I ran the netperf test a few times and one of the times the netperf test would no longer talk to the server. I had to reboot the rt client to get netperf to talk again. I did have cycle soak running which may have contributed to the problem.

Also when netperf is not running the hard IRQs associated with the NIC use a constant 2% load.


Thanks,
Dave


 -------------- Original message ----------------------
From: Ingo Molnar <mingo@elte.hu>
> 
> * David Sperry <dave_sperry@ieee.org> wrote:
> 
> > > there are a few other things i'm working on to improve this. I've 
> > > uploaded -rt9 which is the current state of affairs. Note that using 
> > > -rt9 you'll likely only see IRQ-8406 overhead in the system, because 
> > > i've added an optimization to do process the softirq-net-tx workload 
> > > in the hardirq thread if the priority of the two is the same (which 
> > > is the default behavior). But -rt9 is still work in progress that is 
> > > not fully finished yet: in some cases i'm seeing 'fluctuating 
> > > performance' problems on forcedeth that werent there before.
> > 
> > I tried -rt9 and saw some odd 'fluctuating performance'. I'll try it 
> > again tomorrow when I am much closer to the box's power button.
> 
> could you try -rt10? I've improved hardirq/softirq threading performance 
> some more, and have tuned the forcedeth driver too.
> 
> 	Ingo

Re: Improved UDP performance using 2.6.21-rc5-rt10

[Ingo Molnar]

* dave_sperry@ieee.org <dasperry@comcast.net> wrote:

> rt10 shows a big improvement over rt8
> 
> ##################################################
> etup                    Thruput      CPU% 
> Nvidia
> 2.6.21-rc5-rt8           938          65%
>    netperf @51
>    hardirq @50
>    softirq @50
> 
> Nvidia
> 2.6.21-rc5-rt10           938          32%
>    netperf @51
>    hardirq @50
>    softirq @50

cool! Btw., could you run the same but without running cyclesoak, and 
only watching the 'idle' metric in 'top' or vmstat? I found in my 
experiments that cyclesoak sometimes interfers with the workload - and 
the idle stats are precise and accurate on -rt. (they are not on 
vanilla)

	Ingo

Re: Improved UDP performance using 2.6.21-rc5-rt10

[dave_sperry@ieee.org]

Ingo, Here are some more results:

Nvidia w/cyclesoak
 2.6.21-rc5-rt10           938          32%
    netperf @51
    hardirq @50
    softirq @50

Nvidia w/top
 2.6.21-rc5-rt12           938          31%
    netperf @51
    hardirq @50
    softirq @50

Hopefully I'll run some more tests in the morning

Thanks,
Dave
 -------------- Original message ----------------------
From: Ingo Molnar <mingo@elte.hu>
> 
> * dave_sperry@ieee.org <dasperry@comcast.net> wrote:
> 
> > rt10 shows a big improvement over rt8
> > 
> > ##################################################
> > etup                    Thruput      CPU% 
> > Nvidia
> > 2.6.21-rc5-rt8           938          65%
> >    netperf @51
> >    hardirq @50
> >    softirq @50
> > 
> > Nvidia
> > 2.6.21-rc5-rt10           938          32%
> >    netperf @51
> >    hardirq @50
> >    softirq @50
> 
> cool! Btw., could you run the same but without running cyclesoak, and 
> only watching the 'idle' metric in 'top' or vmstat? I found in my 
> experiments that cyclesoak sometimes interfers with the workload - and 
> the idle stats are precise and accurate on -rt. (they are not on 
> vanilla)
> 
> 	Ingo