RFR: 8020875 java/lang/management/ThreadMXBean/ResetPeakThreadCount.java fails intermittently

[David Holmes]

Sorry - I took a closer look at the full test rather than just that 
patch. We already have this code to try and help expose these 
intermittent failures:

  213         // Nightly testing showed some intermittent failure.
  214         // Check here to get diagnostic information if some strange
  215         // behavior occurs.
  216         checkThreadCount(expectedCount, current, 0);

but the sleep loop you added means this check will rarely fail so we 
won't get to see this unexpected behaviour happening. So this block of 
code could be deleted in my view. Though it is preferable to determine 
exactly why we fail!

Also looking at the sleep() used elsewhere you may as well follow the 
same pattern and abort on interrupt as it isn't expected.

Finally with regard to Daniel's comment about the live array he is right 
that the volatile on the array is not sufficient in theory - a thread 
need never see the value of live[i] become false. There are a number of 
reasons why we are unlikely to see that in practice on hotspot. Using 
synchronized will fix that; or an alternative cancellation mechanism 
could be used.

Cheers,
David

On 23/07/2013 7:19 PM, David Holmes wrote:
> On 23/07/2013 6:29 PM, Jaroslav Bachorik wrote:
>> On 07/23/2013 10:19 AM, David Holmes wrote:
>>> Hi Jaroslav,
>>>
>>> On 22/07/2013 9:55 PM, Jaroslav Bachorik wrote:
>>>> The java/lang/management/ThreadMXBean/ResetPeakThreadCount.java test
>>>> seems to be failing intermittently.
>>>>
>>>> The test checks the functionality of the
>>>> j.l.m.ThreadMXBean.resetPeakThreadCount() method. It does so by
>>>> capturing the current value of "getPeakThreadCount()", starting a
>>>> predefined number of the user threads, stopping them and resetting the
>>>> stored peak value and making sure the new peak equals to the number of
>>>> the actually running threads.
>>>>
>>>> The main problem is that it is not possible to prevent JVM to
>>>> start/stop
>>>> arbitrary system threads while executing the test. This might lead to
>>>> small variations of the reported peak (a short-lived system thread is
>>>> started while the batch of the user threads is running) or the expected
>>>> number of running threads (again, a short-lived system thread is
>>>> started
>>>> at the moment the test asks for the number of running threads).
>>>
>>> Do you know what "system threads" these are? I would not expect VM
>>> internal threads to be counted in getPeakThreadCount(), but even if they
>>> are I can't think of any short-lived threads that get created other than
>>> the Signal handling thread.
>>
>> Unfortunatelly I don't. Capturing the thread dump at the moment of
>> discovering the discrepancy seems to to be too late. I tried monitoring
>> the JVM under the test from external tools but it just brings more
>> entropy to the result.
>
> We'd need to instrument the thread creation logic to keep a separate
> record. Dtrace probes could probably do it - but the problem is getting
> the test to fail.
>
>> I am completely relying on the JVM native thread accounting to be
>> correct and accurate - that it reports the thread count peak based on
>> the real data.
>
> The spec isn't clear but I would only expect these counters to apply to
> Java threads not VM internal threads (compiler, gc etc). So I'd really
> like to know what thread is messing up this count.
>
> David
>
>> -JB-
>>
>>>
>>>> The patch does not fix those shortcomings as it is not really possible
>>>> to do given the nature of the JVM threading system. It rather tries to
>>>> relax the conditions while still maintaining the ability to detect
>>>> functional problems - eg. decreasing peak without explicitly resetting
>>>> it and reporting false number of threads.
>>>>
>>>> The webrev is at:
>>>> http://cr.openjdk.java.net/~jbachorik/8020875/webrev.00
>>>
>>> Seems reasonable.
>>>
>>> David
>>> -----
>>>
>>>> Thanks,
>>>>
>>>> -JB-
>>>>
>>