Re: Tweak Latency on Intel ATOM

["Max Müller" <mxmr@gmx.net>]

Clark Williams schrieb:
> On Thu, 11 Feb 2010 08:54:29 +0100
> Max Müller <mxmr@gmx.net> wrote:
>
>   
>> Clark Williams schrieb:
>>     
>>> On Tue, 9 Feb 2010 07:41:58 +0000 (UTC)
>>> Max Miller <mxmr@gmx.net> wrote:
>>>
>>>   
>>>       
>>>> Hello,
>>>>
>>>> im am using the PREEMPT-RT patch on linux 2.6.29.6. It runs on a MSI965GSE
>>>> industial board with Intel ATOM CPU (N270, 1,6GHz) and i945GSE Northbridge. 
>>>>
>>>> I got about 45µs as maximum and 13µs as average latency when hyperthreading is
>>>> disabled. With enabled Hyperthreading the maximum latency increses to about
>>>> 100µs. I measured the latency with cyclictest. 
>>>>
>>>> What can i do to get better maximum latency? Can I do somthing in the kernel
>>>> configuration or are there some kernel bootoptions? Or is it still impossible
>>>> with this CPU to get better results?
>>>>
>>>> Thanks in advance,
>>>> Max Miller 
>>>>
>>>>
>>>>     
>>>>         
>>> Make sure you turn off any power management settings in the BIOS and
>>> turn off the irqbalance and cpuspeed services on the Linux side.
>>>
>>> What cyclictest command are you using to measure latency?
>>>
>>> Clark
>>>   
>>>       
>> I run cyclictest as follows:
>>
>> cyclictest -n -t3 -p99
>>     
>
> You might want to try the new cyclictest option --smp (which is really
> the options -t, -a -n) and I'd back the priority down to -p95 just to
> keep out of the way of watchdog and migration threads. In general, when
> I run it on a multi-core box I use:
>
> 	$ cyclictest --smp -m -p95 -d0
>
> Lately on AMD boxes I use --numa, which makes calls into libnuma to
> allocated memory on local nodes for the measurement threads.
>
> If you want to get fancy and look at at the history of the run, you can
> use the -h <n> option to keep a histogram of <n> buckets (1 bucket == 1
> microsecond). 
>
>   
>> For generating additional system load i run (one to several instances):
>>
>> while true; do echo "blah" > /dev/null; done &
>>
>> Then i watch the max. latency from the thread with the highest priority.
>> Sometimes i add the parameter '-h' to generate a history. In this 
>> history i can
>> see that the most latency times are under 20µs, only  about 5ppm are 
>> worse than 30µs.
>> Am i doing this correctly?
>>     
>
> You're seeing some nice numbers there (any max latency under 100us is
> pretty good). 
>
> I have a python program I've been developing named 'rteval' which kicks
> off a kernel compile and a scheduler benchmark called 'hackbench', then
> runs cyclictest with the histogram option. After the run it generates a
> report on how well cyclictest did with the loads in place. If you're
> interested, you can get rteval from my kernel.org git repo:
>
> $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/clrkwllms/rteval.git
>
> It's not 100% complete, but it's getting there. 
>
>   
>> The only powersave setting in the BIOS is "Intel speedstep" which i 
>> disabled.
>>
>>
>> I will check with disabled "irqbalance and cpuspeed services" disabled 
>> and will report later.
>>
>>
>> What should the adequate max. latency on this system?
>>
>>
>>     
>
> I'd say you're doing pretty good keeping under 50us. You might want to
> try it under a heavier load than the shell script you've been running.
> If you don't want to fool with rteval, try kicking off a kernel compile
> in another window like this:
>
> $ while true; do make -j4 clean bzImage modules; done
>
> and then run cyclictest. A kernel compile with parallel jobs (-j) is a
> good overall load of computation and I/O.
>
>   

I tested now like you told me with irqbalance and cpuspeed services 
disabled. I hope i made the right for disabling irqbalance, i used the 
kernel parameter acpi_no_irqbalance. Is this correct? Unfortunately the 
results were nearly equal as before.

For measureing latency i did now the following:
-compile kernel (for high system load)
-running cyclictest -n -m -t3 -p94 (for having some running high 
priority threads)
-running cyclictest -n -m -h80 -p95 -l6000000 (for latency measurement)

I will also test your python program the next days.

I have now about 50µs worst case latency and about 15µs average latency.

Greetings,
Max
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html