John, I am forwarding this to the list too. Perhaps there is someone who might be able to help. John Maddock wrote:

Roland Schwarz wrote:

...

John Maddock wrote:

...

Can you try running the .exe on it's own and see if it passes?

I tried to run the very same executable built during the regression test (msvc-7.1 stlport) from the command line:

On my machine this needed about 9 minutes. I observed an interesting fact though: Both cpu units (this is a hyperthreading cpu) were saturated for some time then suddenly falling to 0% load and back again to 100%. This switching between 0% and 100% was almost during the whole test period.

Finally the test passed. On the latest status pages I had 20 minutes kill off time, but the test failed.

So at the moment it seems as if the test only is failing when run from regression.py.

:-(

...

Btw.: Do you have an explanation for the observed behaviour (0%/100%)? An additional note: The load was almost user space only.

No. The test spawns 10 threads that each then run through the whole test suite in parrellel, the aim is to really saturate the test and have a good chance of getting failures if there are any threading issues that have been missed.

Could the observed behavior be an indication that there indeed is a real problem? One that shows up under certain conditions only, as e.g. the one we have?

Could the 0% loads occur when waiting on a lock and getting switched out to another process? It stays at 100% on my single CPU machine, and that's what I would expect to see, there certainly aren't any deliberate waits going on.

I guess I could reduce the thread load and see if it clears the problem, although it'll reduce the usefulness of the test as well unfortunately.

Many thanks for trying this out,

John.

Roland