Grant Edwards wrote:
> On 2008-11-20, Grant Edwards <[EMAIL PROTECTED]> wrote:
>
>   
>> I've been doing some performance tests comparing eCos and
>> uClinux on a NIOS2 platform.  TCP throughput is comparable
>> (uClinux does a bit better with only 1 connection&thread, and
>> eCos does a bit better at 32 connections&threads). But, latency
>> on uClinux is much worse.  The table below shows the time in
>> milliseconds required for a user-space thread on an otherwise
>> idle system to receive a small message (two bytes) via TCP and
>> echo it back (10 trials):
>>     
>
> I've repeated testing with three different kernel-premption
> settings and the results are shown below:
>
>                     ----------uClinux---------
>              eCos    PREEMPT  VOLUNTARY   NONE
>             -----    -------  ---------  -----
>             1.864     8.151    3.661     2.934
>             2.079     4.509    7.507     5.805
>             1.972     7.699    7.047     2.583
>             1.871     7.189    7.046     4.702
>             1.728     5.152    4.954     3.537
>             1.842     6.803    6.212     3.416
>             1.733     7.041    7.504     6.048
>             1.847     3.758    3.637     2.847
>             1.850     7.703    6.385     4.939
>             1.842     1.785    3.630     2.924
>             
>        min  1.728     1.785    3.630     2.583
>        avg  1.863     5.979    5.758     3.974
>        max  2.079     8.151    7.507     6.048
>       range 0.351     6.366    3.877     3.465
>
> Changing the kernel pre-emption setting doesn't seem to make a
> noticable difference.  It appears that the "server" setting
> performs a little better, but I don't think the difference is
> statistically significant given the high variability and small
> sample size.
>   
Darn, that would have been too easy.
> Does the tick size matter?  It's set to 10ms right now, so I
> guess I'll try changing it to 1ms to see what happens.
>   
Its worth a try, but be careful because reducing tick size can have the
adverse side effect of increasing context switch overhead.   Reducing
the tick size too much could cause more work for the processor in task
switching rather than doing actual process work and thus degrading your
performance. 


_______________________________________________
uClinux-dev mailing list
[email protected]
http://mailman.uclinux.org/mailman/listinfo/uclinux-dev
This message was resent by [email protected]
To unsubscribe see:
http://mailman.uclinux.org/mailman/options/uclinux-dev

Reply via email to