[Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq XC702 Board with Xenomai 2.6.3

[Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq XC702 Board with Xenomai 2.6.3

[Botev, Roumen]

Hello, All,

I am running Xenomai's out of the box battery of tests (xeno-test) on a Xilinx Zynq XC702 evaluation board.

I am new to Xenomai and I am wondering if these results look good to people having more experience in running Xenomai on a Cortex-A9 hardware. In particular, I had higher expectations of the average and worst case latency numbers - I was expecting < 10 usec worst latency for the highest priority user mode task. Here is the summary of the latency test (the full output of the test is attached to this message):

RTT|  00:15:04  (periodic user-mode task, 1000 us period, priority 99)
RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat best|--lat worst
. . . . . . . . . .
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|      3.675|     14.049|     35.913|       0|     0|    00:15:04/00:15:04

Any comments and/or insights are appreciated.

Best regards,

Roumen




_____________________________________________________________________
This email message, including any attachments, may contain confidential and
proprietary information for the sole use of the intended recipient.  If you are
not the intended recipient, you are hereby notified that any use, copying or
dissemination of this message is strictly prohibited.  If you received this
message in error, please notify Brooks Automation, Inc. immediately by reply
email or by calling Brooks US Headquarters at +1 978-262-2400. Then delete this
message from your system, without making any copy or distribution.  Thank you.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: LatencyTest-Zynq XC702 - pw22.txt
URL: <http://www.xenomai.org/pipermail/xenomai/attachments/20140721/40c5b922/attachment.txt>

Re: [Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq XC702 Board with Xenomai 2.6.3

[Philippe Gerum]

On 07/21/2014 08:50 PM, Botev, Roumen wrote:
> Hello, All,
> 
> I am running Xenomai's out of the box battery of tests (xeno-test) on a Xilinx Zynq XC702 evaluation board.
> 
> I am new to Xenomai and I am wondering if these results look good to people having more experience in running Xenomai on a Cortex-A9 hardware. In particular, I had higher expectations of the average and worst case latency numbers - I was expecting < 10 usec worst latency for the highest priority user mode task. Here is the summary of the latency test (the full output of the test is attached to this message):
> 
> RTT|  00:15:04  (periodic user-mode task, 1000 us period, priority 99)
> RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat best|--lat worst
> . . . . . . . . . .
> ---|-----------|-----------|-----------|--------|------|-------------------------
> RTS|      3.675|     14.049|     35.913|       0|     0|    00:15:04/00:15:04
> 

Figures look ok particularly in dual core mode, regarding the scheduling
latency of a userland thread over a 1Khz period. A large part of such
latency is induced by cache effects, including those entailed by
switching memory contexts.

- scheduling latency for kernel threads (latency -t1) should show lower
figures, although these threads should be reserved to running
driver/RTDM tasks (i.e. not application level stuff)

- the lower the sampling frequency, the higher the latency, as the
regular linux system has more time to tread on a larger address space,
which in turn makes the real-time code a more frequent victim of cache
eviction. Running with -p 100 should get you different results (although
this may not apply to your use case anyway, just for kicks).

-- 
Philippe.

Re: [Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq XC702 Board with Xenomai 2.6.3

[Botev, Roumen]

Thank you very much for the feedback and the insights!

Indeed, for lower sampling periods I am seeing lower latency and jitter numbers -- this ties in with the cache eviction explanation. For what it's worth, here is the summary of the latency test with 100 usec sampling period (-p 100) for user / kernel mode tasks and IRQ, with default loading from xeno-test (dohell):

User mode task (-t 0):
RTH|----lat min|----lat avg|----lat max|-overrun|---msw|...
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|      3.384|      9.699|     30.918|       0|     0|    00:15:04/00:15:04
    
Kernel mode task (-t 1):
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|      3.430|      6.721|     26.238|       0|     0|    00:15:04/00:15:04

Timer IRQ (-t 2):
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|      1.560|      3.668|     14.137|       0|     0|    00:15:04/00:15:04




Best regards,
Roumen

-----Original Message-----
From: Philippe Gerum [mailto:rpm@xenomai.org] 
Sent: Thursday, July 24, 2014 5:16 AM
To: Botev, Roumen; xenomai@xenomai.org
Subject: Re: [Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq XC702 Board with Xenomai 2.6.3

On 07/21/2014 08:50 PM, Botev, Roumen wrote:
> Hello, All,
> 
> I am running Xenomai's out of the box battery of tests (xeno-test) on a Xilinx Zynq XC702 evaluation board.
> 
> I am new to Xenomai and I am wondering if these results look good to people having more experience in running Xenomai on a Cortex-A9 hardware. In particular, I had higher expectations of the average and worst case latency numbers - I was expecting < 10 usec worst latency for the highest priority user mode task. Here is the summary of the latency test (the full output of the test is attached to this message):
> 
> RTT|  00:15:04  (periodic user-mode task, 1000 us period, priority 99)
> RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat best|--lat worst
> . . . . . . . . . .
> ---|-----------|-----------|-----------|--------|------|-------------------------
> RTS|      3.675|     14.049|     35.913|       0|     0|    00:15:04/00:15:04
> 

Figures look ok particularly in dual core mode, regarding the scheduling
latency of a userland thread over a 1Khz period. A large part of such
latency is induced by cache effects, including those entailed by
switching memory contexts.

- scheduling latency for kernel threads (latency -t1) should show lower
figures, although these threads should be reserved to running
driver/RTDM tasks (i.e. not application level stuff)

- the lower the sampling frequency, the higher the latency, as the
regular linux system has more time to tread on a larger address space,
which in turn makes the real-time code a more frequent victim of cache
eviction. Running with -p 100 should get you different results (although
this may not apply to your use case anyway, just for kicks).

-- 
Philippe.



_____________________________________________________________________
This email message, including any attachments, may contain confidential and
proprietary information for the sole use of the intended recipient.  If you are
not the intended recipient, you are hereby notified that any use, copying or
dissemination of this message is strictly prohibited.  If you received this
message in error, please notify Brooks Automation, Inc. immediately by reply
email or by calling Brooks US Headquarters at +1 978-262-2400. Then delete this
message from your system, without making any copy or distribution.  Thank you.

Re: [Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq XC702 Board with Xenomai 2.6.3

[Gilles Chanteperdrix]

On 07/25/2014 10:34 AM, Andreas Glatz wrote:
> 
>>
>> Message: 5
>> Date: Thu, 24 Jul 2014 11:16:23 +0200
>> From: Philippe Gerum <rpm@xenomai.org>
>> To: "Botev, Roumen" <roumen.botev@brooks.com>,  "xenomai@xenomai.org"
>> 	<xenomai@xenomai.org>
>> Subject: Re: [Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq
>> 	XC702 Board with Xenomai 2.6.3
>> Message-ID: <53D0CEE7.6010208@xenomai.org>
>> Content-Type: text/plain; charset=ISO-8859-1
>>
>> On 07/21/2014 08:50 PM, Botev, Roumen wrote:
>>> Hello, All,
>>>
>>> I am running Xenomai's out of the box battery of tests (xeno-test)  
>>> on a Xilinx Zynq XC702 evaluation board.
>>>
>>> I am new to Xenomai and I am wondering if these results look good  
>>> to people having more experience in running Xenomai on a Cortex-A9  
>>> hardware. In particular, I had higher expectations of the average  
>>> and worst case latency numbers - I was expecting < 10 usec worst  
>>> latency for the highest priority user mode task. Here is the  
>>> summary of the latency test (the full output of the test is  
>>> attached to this message):
>>>
>>> RTT|  00:15:04  (periodic user-mode task, 1000 us period, priority  
>>> 99)
>>> RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat  
>>> best|--lat worst
>>> . . . . . . . . . .
>>> ---| 
>>> -----------| 
>>> -----------|-----------|--------|------|-------------------------
>>> RTS|      3.675|     14.049|     35.913|       0|     0|     
>>> 00:15:04/00:15:04
>>>
>>
>> Figures look ok particularly in dual core mode, regarding the  
>> scheduling
>> latency of a userland thread over a 1Khz period. A large part of such
>> latency is induced by cache effects, including those entailed by
>> switching memory contexts.
>>
> 
> 
> I tried the test on an omap4460 (A9; 2x1GHz) with Linux 3.14.1 and  
> Xenomai 2.6.3 and got similar results:
> 
> RTT|  00:15:04  (periodic user-mode task, 1000 us period, priority 99)
> RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat best|-- 
> lat worst
> RTD|     10.428|     14.131|     19.923|       0|     0|       
> 6.943|     27.980
> Load script terminated, terminating checked scripts
> ---| 
> -----------| 
> -----------|-----------|--------|------|-------------------------
> RTS|      6.943|     14.483|     27.980|       0|     0|     
> 00:15:04/00:15:04
> 
> So does it makes sense to try and further optimise the cache access  
> for bringing down the latency or is it as good as it can get?

The I-pipe for 3.14 disables the L2 cache "write-allocate" policy,
which, on my omap4430 at least brings down latencies under 25us at 10
kHz. As Philippe said, the figures at 10kHz are usually smaller than
those at 1kHz.

I will run long tests this week to confirm the I-pipe for 3.14 figures.


-- 
                                                                Gilles.

Re: [Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq XC702 Board with Xenomai 2.6.3

[Andreas Glatz]

>
> Message: 5
> Date: Thu, 24 Jul 2014 11:16:23 +0200
> From: Philippe Gerum <rpm@xenomai.org>
> To: "Botev, Roumen" <roumen.botev@brooks.com>,  "xenomai@xenomai.org"
> 	<xenomai@xenomai.org>
> Subject: Re: [Xenomai] Latency Test Results (xeno-test) on Xilinx Zynq
> 	XC702 Board with Xenomai 2.6.3
> Message-ID: <53D0CEE7.6010208@xenomai.org>
> Content-Type: text/plain; charset=ISO-8859-1
>
> On 07/21/2014 08:50 PM, Botev, Roumen wrote:
>> Hello, All,
>>
>> I am running Xenomai's out of the box battery of tests (xeno-test)  
>> on a Xilinx Zynq XC702 evaluation board.
>>
>> I am new to Xenomai and I am wondering if these results look good  
>> to people having more experience in running Xenomai on a Cortex-A9  
>> hardware. In particular, I had higher expectations of the average  
>> and worst case latency numbers - I was expecting < 10 usec worst  
>> latency for the highest priority user mode task. Here is the  
>> summary of the latency test (the full output of the test is  
>> attached to this message):
>>
>> RTT|  00:15:04  (periodic user-mode task, 1000 us period, priority  
>> 99)
>> RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat  
>> best|--lat worst
>> . . . . . . . . . .
>> ---| 
>> -----------| 
>> -----------|-----------|--------|------|-------------------------
>> RTS|      3.675|     14.049|     35.913|       0|     0|     
>> 00:15:04/00:15:04
>>
>
> Figures look ok particularly in dual core mode, regarding the  
> scheduling
> latency of a userland thread over a 1Khz period. A large part of such
> latency is induced by cache effects, including those entailed by
> switching memory contexts.
>


I tried the test on an omap4460 (A9; 2x1GHz) with Linux 3.14.1 and  
Xenomai 2.6.3 and got similar results:

RTT|  00:15:04  (periodic user-mode task, 1000 us period, priority 99)
RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat best|-- 
lat worst
RTD|     10.428|     14.131|     19.923|       0|     0|       
6.943|     27.980
Load script terminated, terminating checked scripts
---| 
-----------| 
-----------|-----------|--------|------|-------------------------
RTS|      6.943|     14.483|     27.980|       0|     0|     
00:15:04/00:15:04

So does it makes sense to try and further optimise the cache access  
for bringing down the latency or is it as good as it can get?

Andreas