Issues running jcstress on multi-core systems

[Aleksey Shipilev]

On 08/10/2016 07:41 PM, Johnson, Andy wrote:
> I have recently downloaded, built and run jcstress (custom suite
> only, for now) on a 24-core Linux server-class system.  I am finding
> that several of the "interrupt" tests are failing.  However, I have
> discovered that the number of failures seems to be related to the
> number of processor cores being used during the test.

"interrupt" tests are inherently time-based: they usually wait for
timeout. There is a chance that timeout is not enough on busy machines
-- while the test tries to assert that the code quits, it runs out of
time before quitting, and infra fails it.

Now, I realized we have a hard-coded timeout in the tests. Oops. Fixed,
now it is derived from the iteration time:
  http://hg.openjdk.java.net/code-tools/jcstress/rev/b0c5e26af669

So, with an updated jcstress, you may want to try different "-time
<msec>" settings and see if it helps. If it does help, we may need to
throttle other tests while interrupt tests are executing.

> I am using "taskset" to control the number of cores used by java.

jcstress harness accepts "-sc <numCPUs>" to artificially limit the
number of hardware threads to use.

Thanks,
-Aleksey