harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Alexey Varlamov" <alexey.v.varla...@gmail.com>
Subject Re: [DRLVM] General stability
Date Wed, 08 Nov 2006 16:08:32 GMT
2006/11/7, Vladimir Ivanov <ivavladimir@gmail.com>:
> On 11/7/06, Alexey Varlamov <alexey.v.varlamov@gmail.com> wrote:
> >
> >
> > But do we have needed scripts/tools readily available to run and
> > analyze such stability testing? I'm also pretty sure existing c-unit
> > and smoke tests would help to reveal certain problems if run
> > repeatedly - just need to add this stuff to CC and run it nightly.
> > Anybody volunteer?
> > And yet there are a lot of excluded tests in smoke suite...
> Actually, we have one. The task 'ant test' from the drlvm module is running
> under CC on linux and windows boxes. Every one can easily reproduce it
> (checkout 'buildtest' module and run it, updated version is available in the
> issue 995).
> The problem is: CC will be useful only to track the regression. While we
> have some failed tests it should be fixed asap. At present time some issues
> that prevent successful CC runs wait for integration more than 1 month :(

AFAIU CC only tracks status changes between subsequent runs, right?
This is not really helpful to detect stability issues.
Apparently a few race conditions present in DRLVM threads suspension,
classloading, and/or elsewhere. So one rather need to collect
statistics to spot suspicious areas. Results can be represented as %%
of failures per test and groups of  failures by symptoms, e.g. the
same assert failed for different tests. Further analysis could bred
better tests with higher failure probability...
Alternative approach could be to employ custom "stressing" test
harness, running a test in several concurrent threads, etc.

Although this is just a temp solution until we have thorough stress
tests and more decent coverage of VM code by unit tests, the latter
will hardly happen in foreseeable future. So we should try to derive
maximum benefit from already available tests.

>  thanks, Vladimir

View raw message