[Xenomai-core] My work on Xenomai integration to LTTng / LTTV

[Xenomai-core] My work on Xenomai integration to LTTng / LTTV

[Jean-Olivier Villemure]

[-- Attachment #1: Type: text/plain, Size: 1126 bytes --]

Hello,

I'm currently working on the integration of Xenomai tracing in LTTng
(Linux Trace Toolkit Next Generation) and the event viewing in LTTV
(Linux Trace Toolkit Viewer).

At this moment, I 'm tracing events about task handling, period and
soon, mutex and semaphore synchronization.

Inserting traqcing points in Xenomai nucleus is not very difficult, my
main work is focus on the controlflow viewer module of LTTV. At this
moment, we can open a trace and view:
- Task state : init, started, running, suspended, overrunning (period
miss)
- Period timer tick for periodic task
- Textual information about the task (name, period, priority, birth)
- Mutex/semaphore owned by specific task
- Soon task waiting on mutex/semaphore

Look at the screenshot to have a better idea.

The next step will be to generate some statistics, for this step I will
need your help. As Xenomai users, which kind of stats would you want to
compute?

After that, I'll study the possibilities of simulating an execution
trace from an existing trace by modifying some parameters (period
length, ...)

Any ideas are welcome!

Jean-Olivier Villemure

[-- Attachment #2: Type: image/png, Size: 121459 bytes --]

[-- Attachment #3: Type: image/png, Size: 1209 bytes --]

[-- Attachment #4: Type: image/png, Size: 66360 bytes --]

Re: [Xenomai-core] My work on Xenomai integration to LTTng / LTTV

[Jan Kiszka]

[-- Attachment #1: Type: text/plain, Size: 2196 bytes --]

Jean-Olivier Villemure wrote:
> Hello,
> 
> I'm currently working on the integration of Xenomai tracing in LTTng
> (Linux Trace Toolkit Next Generation) and the event viewing in LTTV
> (Linux Trace Toolkit Viewer).
> 
> At this moment, I 'm tracing events about task handling, period and
> soon, mutex and semaphore synchronization.

Wow, great news!

> 
> Inserting traqcing points in Xenomai nucleus is not very difficult, my
> main work is focus on the controlflow viewer module of LTTV. At this
> moment, we can open a trace and view:
> - Task state : init, started, running, suspended, overrunning (period
> miss)
> - Period timer tick for periodic task
> - Textual information about the task (name, period, priority, birth)
> - Mutex/semaphore owned by specific task
> - Soon task waiting on mutex/semaphore
> 
> Look at the screenshot to have a better idea.
> 
> The next step will be to generate some statistics, for this step I will
> need your help. As Xenomai users, which kind of stats would you want to
> compute?

 - CPU usage (see /proc/xenomai/stat for average numbers, but we should
   now be able to precisely calculate them for a specific period)
 - Waiting times after activation (maybe one can define deadlines later
   and verify them)
 - Number of preemptions per second or whatever while a task is runnable
 - Number of blockades a task faces due to unavailable resources
   (mutexes, semaphores, etc.)

...just to through some first ideas in.

> 
> After that, I'll study the possibilities of simulating an execution
> trace from an existing trace by modifying some parameters (period
> length, ...)
> 
> Any ideas are welcome!

So you have an integrated I-pipe+LTTng patch which is able to record
events under any context and push them to Linux non-RT user-space,
right? How does this work, what impact does it have (also compared to
the I-pipe tracer, which is not light-weight)? And, most important, when
will we be able to see some code (ipipe add-on patch)? Would be good to
assess how this could be integrated on the long term into
I-pipe/Xenomai, at least until required LTTng parts become mainstream.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 249 bytes --]

Re: [Xenomai-core] My work on Xenomai integration to LTTng / LTTV

[Philippe Gerum]

On Tue, 2006-10-31 at 18:34 +0100, Jan Kiszka wrote:
> Jean-Olivier Villemure wrote:
> > 
> > The next step will be to generate some statistics, for this step I will
> > need your help. As Xenomai users, which kind of stats would you want to
> > compute?
> 
>  - CPU usage (see /proc/xenomai/stat for average numbers, but we should
>    now be able to precisely calculate them for a specific period)
>  - Waiting times after activation (maybe one can define deadlines later
>    and verify them)
>  - Number of preemptions per second or whatever while a task is runnable
>  - Number of blockades a task faces due to unavailable resources
>    (mutexes, semaphores, etc.)

- Average number of elements linked to the readyq per second (some kind
of loadavg for us, i.e. answers the question "should the user switch to
the scalable scheduler?")
- Average number of outstanding timers per second (i.e. answers the
question "should the user switch to the binary heap-based timer
management?").
- Total number of timers missing their wakeup date from more than a
given threshold.
- Pressure on the system heap (average and max, probably).
- Transfer rate on message pipes (i.e. nucleus/pipe.c); at some point,
it would be interesting to extend this to skins exposing data channels
(e.g. native API's queues, POSIX mqueues).

-- 
Philippe.