Robust Futexes status

Robust Futexes status

[david singleton]

[-- Attachment #1: Type: text/plain, Size: 1119 bytes --]


	Inaky Perez-Gonzalez   has a wonderful suite of performance, stress and
functionality tests for the fusyn pthreads mutex package.

	I've been able to use this test suite to measure robust futex 
performance with very little
change.  The posix test suite compiles cleanly against the nptl header 
files and libararies,
which is very good news.   It means we got the header files and 
libraries right.

	The following are three  graphs on the pthread mutex ownership change 
latency test
for all three flavors of kernel,  PREEMPT_NONE,  PREEMPT_DESKTOP and 
PREEMPT_RT,
using the realtime preemption patch-2.6.14-rc5-rt5.   The 
ownership_change_latency
test ran unchanged, modulo changing the clock-id for the real time 
clock in the run script.

	The graphs are just to show relative performance for the different 
flavor kernels.  The
kernel's perform quite closely regardless of the 'flavor' of kernel.  
The kernels have quite a few
debugging options turned so I can look for any problems so performance 
is not optimal.

	It appears the robust futex functionality is healthy in all flavors of 
kernel.

	

David


[-- Attachment #2: mutex-none.png --]
[-- Type: image/png, Size: 17289 bytes --]

[-- Attachment #3: Type: text/plain, Size: 3 bytes --]





[-- Attachment #4: mutex-desktop.png --]
[-- Type: image/png, Size: 5270 bytes --]

[-- Attachment #5: Type: text/plain, Size: 2 bytes --]




[-- Attachment #6: mutex-rt.png --]
[-- Type: image/png, Size: 5336 bytes --]

Re: Robust Futexes status

[Sven-Thorsten Dietrich]

On Sun, 2005-10-23 at 13:52 -0700, david singleton wrote:
> 	Inaky Perez-Gonzalez   has a wonderful suite of performance, stress and
> functionality tests for the fusyn pthreads mutex package.

> 	The graphs are just to show relative performance for the different 
> flavor kernels.  The
> kernel's perform quite closely regardless of the 'flavor' of kernel.  
> The kernels have quite a few
> debugging options turned so I can look for any problems so performance 
> is not optimal.
> 

The Mutex ownership change seems to climb with waiting threads.

Its hard to tell with the log-n X-axis scale, but does this possibly
correlate to the deadlock-detect option?

If Deadlock-detect is enabled we should be seeing a graph proportional
to n-squared on a linear X axis.

If deadlock detect is disabled, the wait time should plateau for very
large N.

Sven


> 	It appears the robust futex functionality is healthy in all flavors of 
> kernel.
> 
> 	
> 
> David
> 
> 
> 
> 
-- 
***********************************
Sven-Thorsten Dietrich
Real-Time Software Architect
MontaVista Software, Inc.
1237 East Arques Ave.
Sunnyvale, CA 94085

Phone: 408.992.4515
Fax: 408.328.9204

http://www.mvista.com
Platform To Innovate
*********************************** 

Re: Robust Futexes status

[david singleton]

On Oct 23, 2005, at 8:48 PM, Sven-Thorsten Dietrich wrote:

> On Sun, 2005-10-23 at 13:52 -0700, david singleton wrote:
>> 	Inaky Perez-Gonzalez   has a wonderful suite of performance, stress 
>> and
>> functionality tests for the fusyn pthreads mutex package.
>
>> 	The graphs are just to show relative performance for the different
>> flavor kernels.  The
>> kernel's perform quite closely regardless of the 'flavor' of kernel.
>> The kernels have quite a few
>> debugging options turned so I can look for any problems so performance
>> is not optimal.
>>
>
> The Mutex ownership change seems to climb with waiting threads.
>
> Its hard to tell with the log-n X-axis scale, but does this possibly
> correlate to the deadlock-detect option?

It could be.  I have been running all the tests mainly as stress and
functionality tests.
>
> If Deadlock-detect is enabled we should be seeing a graph proportional
> to n-squared on a linear X axis.
>
> If deadlock detect is disabled, the wait time should plateau for very
> large N.

The original run script ran up to 7500 waiting threads.  I ran a version
that only went up to 400 threads.

When I get full performance data I'll post it.

David
>
> Sven
>
>
>> 	It appears the robust futex functionality is healthy in all flavors 
>> of
>> kernel.
>>
>> 	
>>
>> David
>>
>>
>>
>>
> -- 
> ***********************************
> Sven-Thorsten Dietrich
> Real-Time Software Architect
> MontaVista Software, Inc.
> 1237 East Arques Ave.
> Sunnyvale, CA 94085
>
> Phone: 408.992.4515
> Fax: 408.328.9204
>
> http://www.mvista.com
> Platform To Innovate
> ***********************************
>