[Xenomai] How to locate the max latency

[Xenomai] How to locate the max latency

[Zheng, Qi]

I am using xenomai 3.0.5 (https://xenomai.org/downloads/xenomai/stable/xenomai-3.0.5.tar.bz2) on the Intel SKL platform.
On console mode or accessed by vnc viewer (no i915 graphics interrupt), the max latency values are usually below 20us no matter low or high loading (by stress-ng).
But the max latency can reach up to 40us on GUI (gnome)  mode (dozens of i915 interrupt counts per second).
I tried enabling the ipipe trace function, but couldn't find valuable information.
Is there any other good ways to locate the max latency?
Or does anybody have experience on reducing the latency caused by graphics (such as i915) ?
Many thanks.


Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

[Xenomai] How to locate the max latency

[Zheng, Qi]

BTW, I already disabled the i915 power saving by adding "i915.enable_dc=0 i915.enable_rc6=0 i915.enable_execlists=0 i915.disable_power_well=0" in  the cmdline.

Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

From: Zheng, Qi
Sent: Tuesday, July 3, 2018 5:00 PM
To: xenomai@xenomai.org
Subject: How to locate the max latency

I am using xenomai 3.0.5 (https://xenomai.org/downloads/xenomai/stable/xenomai-3.0.5.tar.bz2) on the Intel SKL platform.
On console mode or accessed by vnc viewer (no i915 graphics interrupt), the max latency values are usually below 20us no matter low or high loading (by stress-ng).
But the max latency can reach up to 40us on GUI (gnome)  mode (dozens of i915 interrupt counts per second).
I tried enabling the ipipe trace function, but couldn't find valuable information.
Is there any other good ways to locate the max latency?
Or does anybody have experience on reducing the latency caused by graphics (such as i915) ?
Many thanks.


Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

Re: [Xenomai] How to locate the max latency

[Greg Gallagher]

Did youale any changes in menuconfig when building the kernel?

Greg

  Original Message  
From: qi.zheng@intel.com
Sent: July 4, 2018 3:27 AM
To: xenomai@xenomai.org
Subject: [Xenomai] How to locate the max latency

BTW, I already disabled the i915 power saving by adding "i915.enable_dc=0 i915.enable_rc6=0 i915.enable_execlists=0 i915.disable_power_well=0" in  the cmdline.

Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

From: Zheng, Qi
Sent: Tuesday, July 3, 2018 5:00 PM
To: xenomai@xenomai.org
Subject: How to locate the max latency

I am using xenomai 3.0.5 (https://xenomai.org/downloads/xenomai/stable/xenomai-3.0.5.tar.bz2) on the Intel SKL platform.
On console mode or accessed by vnc viewer (no i915 graphics interrupt), the max latency values are usually below 20us no matter low or high loading (by stress-ng).
But the max latency can reach up to 40us on GUI (gnome)  mode (dozens of i915 interrupt counts per second).
I tried enabling the ipipe trace function, but couldn't find valuable information.
Is there any other good ways to locate the max latency?
Or does anybody have experience on reducing the latency caused by graphics (such as i915) ?
Many thanks.


Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

_______________________________________________
Xenomai mailing list
Xenomai@xenomai.org
https://xenomai.org/mailman/listinfo/xenomai

Re: [Xenomai] How to locate the max latency

[Zheng, Qi]

Yes, most of my modifications on top of general x86_64 config were based on the https://gitlab.denx.de/Xenomai/xenomai/wikis/Configuring_For_X86_Based_Dual_Kernels.
Besides, I enabled the "CONFIG_PREEMPT", gained a little bit latency improvement by my test.
And I also confirmed there was no SMI during the test. 

Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

-----Original Message-----
From: Greg Gallagher [mailto:greg@embeddedgreg.com] 
Sent: Wednesday, July 4, 2018 7:11 PM
To: Zheng, Qi <qi.zheng@intel.com>; xenomai@xenomai.org
Subject: Re: [Xenomai] How to locate the max latency

Did youale any changes in menuconfig when building the kernel?

Greg

  Original Message  
From: qi.zheng@intel.com
Sent: July 4, 2018 3:27 AM
To: xenomai@xenomai.org
Subject: [Xenomai] How to locate the max latency

BTW, I already disabled the i915 power saving by adding "i915.enable_dc=0 i915.enable_rc6=0 i915.enable_execlists=0 i915.disable_power_well=0" in  the cmdline.

Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

From: Zheng, Qi
Sent: Tuesday, July 3, 2018 5:00 PM
To: xenomai@xenomai.org
Subject: How to locate the max latency

I am using xenomai 3.0.5 (https://xenomai.org/downloads/xenomai/stable/xenomai-3.0.5.tar.bz2) on the Intel SKL platform.
On console mode or accessed by vnc viewer (no i915 graphics interrupt), the max latency values are usually below 20us no matter low or high loading (by stress-ng).
But the max latency can reach up to 40us on GUI (gnome)  mode (dozens of i915 interrupt counts per second).
I tried enabling the ipipe trace function, but couldn't find valuable information.
Is there any other good ways to locate the max latency?
Or does anybody have experience on reducing the latency caused by graphics (such as i915) ?
Many thanks.


Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)

_______________________________________________
Xenomai mailing list
Xenomai@xenomai.org
https://xenomai.org/mailman/listinfo/xenomai

Re: [Xenomai] How to locate the max latency

[Jan Kiszka]

On 2018-07-03 10:59, Zheng, Qi wrote:
> I am using xenomai 3.0.5 (https://xenomai.org/downloads/xenomai/stable/xenomai-3.0.5.tar.bz2) on the Intel SKL platform.
> On console mode or accessed by vnc viewer (no i915 graphics interrupt), the max latency values are usually below 20us no matter low or high loading (by stress-ng).
> But the max latency can reach up to 40us on GUI (gnome)  mode (dozens of i915 interrupt counts per second).
> I tried enabling the ipipe trace function, but couldn't find valuable information.
> Is there any other good ways to locate the max latency?
> Or does anybody have experience on reducing the latency caused by graphics (such as i915) ?
> Many thanks.
>
First of all, how long did you run the tests? Maybe enabling graphics
just increases the probability to hit the peak, and the same is
conceptually possible without it.

It may also be the case that this is hardware-induced (GPU). Do you see
a similar increase in the latency (even if from a different ground
level) when using the PREEMPT-RT kernel?

Jan

-- 
Siemens AG, Corporate Technology, CT RDA IOT SES-DE
Corporate Competence Center Embedded Linux

Re: [Xenomai] How to locate the max latency

[Zheng, Qi]

1. I run the test one hour for a round, totally tested for 12 rounds.
     The data were consistent, on console mode, max latency were between 10us to 19us, on GUI mode, they were between 28us to 42us.
2. I tried the PREEMPT-RT kernel, yes, there was similar increase.


Best Regards

SSG->OTC->One Linux Platform Engineering Team (Shanghai)


-----Original Message-----
From: Jan Kiszka [mailto:jan.kiszka@siemens.com] 
Sent: Thursday, July 5, 2018 2:02 PM
To: Zheng, Qi <qi.zheng@intel.com>; xenomai@xenomai.org
Subject: Re: How to locate the max latency

On 2018-07-03 10:59, Zheng, Qi wrote:
> I am using xenomai 3.0.5 (https://xenomai.org/downloads/xenomai/stable/xenomai-3.0.5.tar.bz2) on the Intel SKL platform.
> On console mode or accessed by vnc viewer (no i915 graphics interrupt), the max latency values are usually below 20us no matter low or high loading (by stress-ng).
> But the max latency can reach up to 40us on GUI (gnome)  mode (dozens of i915 interrupt counts per second).
> I tried enabling the ipipe trace function, but couldn't find valuable information.
> Is there any other good ways to locate the max latency?
> Or does anybody have experience on reducing the latency caused by graphics (such as i915) ?
> Many thanks.
>
First of all, how long did you run the tests? Maybe enabling graphics just increases the probability to hit the peak, and the same is conceptually possible without it.

It may also be the case that this is hardware-induced (GPU). Do you see a similar increase in the latency (even if from a different ground
level) when using the PREEMPT-RT kernel?

Jan

--
Siemens AG, Corporate Technology, CT RDA IOT SES-DE Corporate Competence Center Embedded Linux

Re: [Xenomai] How to locate the max latency

[Jan Kiszka]

On 2018-07-05 08:55, Zheng, Qi wrote:
> 1. I run the test one hour for a round, totally tested for 12 rounds.
>      The data were consistent, on console mode, max latency were between 10us to 19us, on GUI mode, they were between 28us to 42us.
> 2. I tried the PREEMPT-RT kernel, yes, there was similar increase.
> 

Then you should probably discuss this with your colleagues from the
hardware design, if there are ways for software to avoid that increase
at all. If there are, we can look for how they can be activated on Linux
(but we likely need some hints from them).

Jan

> 
> Best Regards
> 
> SSG->OTC->One Linux Platform Engineering Team (Shanghai)
> 
> 
> -----Original Message-----
> From: Jan Kiszka [mailto:jan.kiszka@siemens.com] 
> Sent: Thursday, July 5, 2018 2:02 PM
> To: Zheng, Qi <qi.zheng@intel.com>; xenomai@xenomai.org
> Subject: Re: How to locate the max latency
> 
> On 2018-07-03 10:59, Zheng, Qi wrote:
>> I am using xenomai 3.0.5 (https://xenomai.org/downloads/xenomai/stable/xenomai-3.0.5.tar.bz2) on the Intel SKL platform.
>> On console mode or accessed by vnc viewer (no i915 graphics interrupt), the max latency values are usually below 20us no matter low or high loading (by stress-ng).
>> But the max latency can reach up to 40us on GUI (gnome)  mode (dozens of i915 interrupt counts per second).
>> I tried enabling the ipipe trace function, but couldn't find valuable information.
>> Is there any other good ways to locate the max latency?
>> Or does anybody have experience on reducing the latency caused by graphics (such as i915) ?
>> Many thanks.
>>
> First of all, how long did you run the tests? Maybe enabling graphics just increases the probability to hit the peak, and the same is conceptually possible without it.
> 
> It may also be the case that this is hardware-induced (GPU). Do you see a similar increase in the latency (even if from a different ground
> level) when using the PREEMPT-RT kernel?
> 
> Jan
> 
> --
> Siemens AG, Corporate Technology, CT RDA IOT SES-DE Corporate Competence Center Embedded Linux
> 

-- 
Siemens AG, Corporate Technology, CT RDA IOT SES-DE
Corporate Competence Center Embedded Linux