better performance during stress

better performance during stress

[Ran Shalit]

Hello,


I would like to ask if anyone familiar with results showing better
performance during stress.
We expected (and seen) usually degraded performance during stress, so
such result is a surprise.
I use cyclictest and see that both average and maximum have lower
value during stress.
The chip is i7.

Thank you.
Ran

Re: better performance during stress

[Daniel Bristot de Oliveira]

On 12/04/2017 06:06 PM, Ran Shalit wrote:
> Hello,
> 
> 
> I would like to ask if anyone familiar with results showing better
> performance during stress.
> We expected (and seen) usually degraded performance during stress, so
> such result is a surprise.
> I use cyclictest and see that both average and maximum have lower
> value during stress.
> The chip is i7.

Stress runs a busy loop in user-space, hence the preempt/irqs are 
always enabled. Moreover, it avoids letting the processor going to idle, 
avoiding "exit from idle" latencies.

Using a stress thread per CPU will probably give you better results than 
just letting the system going idle.

You should use rteval as a workload:

https://wiki.linuxfoundation.org/realtime/documentation/howto/tools/rteval

-- Daniel
> Thank you.
> Ran
> --
> 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
> 

Re: better performance during stress

[Ran Shalit]

On Mon, Dec 4, 2017 at 9:11 PM, Daniel Bristot de Oliveira
<daniel@bristot.me> wrote:
>
>
> On 12/04/2017 06:06 PM, Ran Shalit wrote:
>> Hello,
>>
>>
>> I would like to ask if anyone familiar with results showing better
>> performance during stress.
>> We expected (and seen) usually degraded performance during stress, so
>> such result is a surprise.
>> I use cyclictest and see that both average and maximum have lower
>> value during stress.
>> The chip is i7.
>
> Stress runs a busy loop in user-space, hence the preempt/irqs are
> always enabled.

I understand the comments that comes later, but not this one above.
Can you please explain?

Thanks


>Moreover, it avoids letting the processor going to idle,
> avoiding "exit from idle" latencies.
>
> Using a stress thread per CPU will probably give you better results than
> just letting the system going idle.
>
> You should use rteval as a workload:
>
> https://wiki.linuxfoundation.org/realtime/documentation/howto/tools/rteval
>
> -- Daniel
>> Thank you.
>> Ran
>> --
>> 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
>>

Re: better performance during stress

[Daniel Bristot de Oliveira]

On 12/04/2017 09:29 PM, Ran Shalit wrote:
>> Stress runs a busy loop in user-space, hence the preempt/irqs are
>> always enabled.
> I understand the comments that comes later, but not this one above.
> Can you please explain?
> 

A "latency" you see on cyclictest is a side effect of either IRQs being
masked, or the preemption being disabled.

However, a process always returns to user-space with both interrupts and
preemption enabled. So, a task in user-space will not cause latencies.

As stress runs a busy loop in user-space... it will not cause latencies.

-- Daniel