green_threads

[Stefan Seifert]

On Thursday 20 October 2011 18:14:31 Matthew Boyle wrote:
> On 19/10/11 17:33, Matthew Boyle wrote:
> > it currently fails to build on Linux (Fedora 14, x64) when configured
> 
> > using { --cc=g++ --ld=g++ --link=g++ }:
> I'm also seeing intermittent failures in <t/pmc/alarm.t>.  usually just
> tests 3 and 4, occasionally 1 and 2, but i've seen the first four tests
> all fail at once.
> 
> i've been able to consistently reproduce it thus:
> 
> COUNT=1; while prove t/pmc/alarm.t; do echo $((COUNT++)); done;
> 
> the failure seems to be extremely load-dependent (kicking off a couple
> of { cat /dev/zero | sha384sum } pipelines in the background can make it
> fail within a couple of iterations).

Thank you. That's an excellent test procedure.
I can reproduce it and it's pretty obvious, that the test itself is in error.

The test assumes that the alarms will fire in the order in which they are set 
and that they execute in the same order. But on a highly loaded system and 
with active preemption it's no longer possible to tell which task will run in 
which order. So the only meaningful tests are the "All alarms executed", 
"Alarms actually waited" and "Alarm/sleep interaction" tests.

And even of these these, the "Alarms actually waited" test could wrongly 
succeed and the "Alarm/sleep interaction" wrongly fail on an extremely loaded 
system. But those are probably inherent with timer related tests.

Stefan