Re: rt20 scheduling latency testcase and failure data

["Sébastien Dugué" <sebastien.dugue@bull.net>]

On Fri, 2006-05-12 at 19:24 -0700, Darren Hart wrote:
> I have been noticing unexpected intermittant large latencies.  I wrote the 
> attached test case to try and capture some information on them.  The librt.h 
> file contains convenience functions I use for writing other tests as well, so 
> much of it is irrelevant, but the test case itself is pretty clear I believe.
> 
> The test case emulates a periodic thread that wakes up on time%PERIOD=0, so 
> rather than sleeping the same amount of time each round, it checks now 
> against the start of its next period and sleeps for that length of time.  
> Every so often it will miss it's period, I've captured that data and included 
> a few of the interesting bits below.  The results are from a run with a 
> period of 5ms, although I have seen them with periods as high as 17ms.  The 
> system was under heavy network load for some of the time, but not all.
> 

  Hi Darren,

  FWIW:

  I've been running you test program on my box under a stress-kernel
load and did not observe any failure as you describe, not even a max
latency overshooting the 100 us limit (max latencies in the 60~70 us).

  I even went to decrease PERIOD to 1 ms and still no failure. 

  I'm running rt20 with the futex priority based wakeup patch on
a dual 2.8 GHz HT Xeon box. All hardirq and softirq threads are at their
default priority.

  How do you generate the network load you mention? Maybe I could try at
least with the same load you're using.

  Sébastien.