Re: Poor UDP performance using 2.6.21-rc5-rt5

[dasperry@comcast.net (dave_sperry@ieee.org)]

Thanks for all the input Ingo, Here's a list of all the permutations I've tried:

setup                    Thruput      CPU% from cyclesoak
2.6.21-rc5 vanilla       935          29%

2.6.21-rc5-rt5           711          50% //basically all of 1 cpu

2.6.21-rc5-rt8           733          52%

2.6.21-rc5-rt8           824          64%
   netperf @50        
   hardirq @50
   softirq @50

2.6.21-rc5-rt8           937          74%
   netperf @51
   hardirq @50
   softirq @50

2.6.21-rc5-rt8           106          8%
   netperf @51
   hardirq @49
   softirq @50

2.6.21-rc5-rt8           233          14%
   netperf @51
   hardirq @49
   softirq @48

2.6.21-rc5-rt8           67           5%
   netperf @batch
   hardirq @batch
   softirq @batch

2.6.21-rc5-rt8           331           OFF
   netperf @batch
   hardirq @batch
   softirq @batch
   cyclesoak off   


Any thoughts?

-Dave



 -------------- Original message ----------------------
From: Ingo Molnar <mingo@elte.hu>
> 
> * Dave Sperry <dave_sperry@ieee.org> wrote:
> 
> > I checked the clock source and in both the vanilla and rt cases and 
> > they were both acpi_pm
> 
> ok, thanks for double-checking that.
> 
> > Here's the oprofile for my vanilla case:
> 
> i tried your workload and i think i managed to optimize it some more: i 
> have uploaded the -rt8 kernel with these improvements included - could 
> you try it? Is there any measurable improvement relative to -rt5?
> 
> one more thing to improve netperf performance is to do this before 
> running it:
> 
>   chrt -f -p 50 $$
> 
> this will put netperf on the same priority level as the net hardirq and 
> the net softirq (which both default to SCHED_FIFO:50), and should result 
> in a (much) reduced context-switch rate.
> 
> Or, if networking is not latency-critical, then you could move the net 
> hardirq and softirq threads to SCHED_BATCH, and run netperf under 
> SCHED_BATCH as well, using:
> 
>   chrt -b -p 0 $$
> 
> and figuring out the active softirq hardirq thread PIDs and "chrt -b" 
> -ing them too.
> 
> 	Ingo