RFR : JDK-8154166 - java/lang/management/MemoryMXBean/ResetPeakMemoryUsage.java fails with RuntimeException

[Jaroslav Bachorik]

On Mon, Apr 25, 2016 at 9:27 AM, Harsha Wardhana B <
harsha.wardhana.b at oracle.com> wrote:

> Hi,
>
> Please review below patch to disable concurrent GC option.
> http://cr.openjdk.java.net/~hb/8154166/webrev.01/
>

I'm sorry to be a PITA, but why it is not possible to use the @require tag?


>
>
> Jaroslav,
>
> According to Javadoc of Runtime.gc(),
>
> <https://docs.oracle.com/javase/8/docs/api/java/lang/Runtime.html#gc-->
> https://docs.oracle.com/javase/8/docs/api/java/lang/Runtime.html#gc--
>
> The call will only make it's best effort to do a GC and provides no
> guarantee that a given object can be collected even if GC runs.
> It does not say that Runtime.gc() call will block till entire GC cycle is
> finished and hence we cannot be making that assumption.
>

I know, I had the same discussion a while ago when fixing some other tests
failing when run with allowed concurrent explicit GC and I was pointed to
the fact that all the known implementation actually do wait until the
complete GC cycle is over before returning. Otherwise all those tests
relying on some memory having been reclaimed or some counters having been
increased would have to be considered random.


>
> Hence it is required that we encapsulate the target object in
> WeakReference and repeatedly call GC till weakRef returns null.
> Granted that we will have a small window when weakRef returns null and the
> target object is not removed from memory. But I see no way how to fix that
> problem.
>

Exactly. The only guarantee for all the GC related metrics having been
updated before proceeding with the test is being able to run the explicit
GC in blocking manner. Otherwise the tests are not really deterministic and
can intermittently fail.

-JB-


>
> -Harsha
>
>
> On Sunday 24 April 2016 03:17 PM, Jaroslav Bachorik wrote:
>
> The reproducer would be very time sensitive as with the provided
> 'ExplicitGCInvokesConcurrent' it will run GC concurrently with the invoker.
> Otherwise, in the current implementation, calling Runtime.gc() would
> guarantee the GC cycle has finished before that method returns.
>
> The WeakReference javadoc (
> https://docs.oracle.com/javase/7/docs/api/java/lang/ref/WeakReference.html)
> is only stating that the referenced object will be made finalizable at the
> same time as the reference is cleared. As a consequence a cleared reference
> might not always mean that the heap usage has been changed (unless a
> particular GC implementation makes some additional guarantees).
>
> I know we were stabilizing a bunch of related tests relying on GC doing
> its work before checking for some post-conditions and the only way to make
> the tests reliable was to forbid running those tests with
> '-XX:+ExplicitGCInvokesConcurrent'.
>
> -JB-
>
> On Sat, Apr 23, 2016 at 12:15 PM, Harsha Wardhana B <
> <harsha.wardhana.b at oracle.com>harsha.wardhana.b at oracle.com> wrote:
>
>> Hello,
>>
>> The issue was not reproducible with or without,
>>
>> "-XX:+ExplicitGCInvokesConcurrent"
>>
>> Flag. The patch ensures that GC happens before we start measuring memory.
>> Without the patch, GC might or might not happen.
>>
>> -Harsha
>>
>>
>> On Friday 22 April 2016 07:58 PM, Jaroslav Bachorik wrote:
>>
>> Hi,
>>
>> On Fri, Apr 22, 2016 at 9:10 AM, Harsha Wardhana B <
>> <harsha.wardhana.b at oracle.com>harsha.wardhana.b at oracle.com> wrote:
>>
>>> Hi,
>>>
>>> Please review the below simple fix for issue,
>>>
>>> issue : https://bugs.openjdk.java.net/browse/JDK-8154166
>>> webrev : http://cr.openjdk.java.net/~hb/8154166/webrev.00/
>>>
>>
>> Shouldn't this test rather declare the conditions when it is supposed to
>> work? According to the issue this was caused by introducing the "-XX:+ExplicitGCInvokesConcurrent"
>> which makes it very tricky to make any assumptions about the GC process.
>>
>> See eg. jdk/tests/java/lang/management/MemoryMXBean/LowMemoryTest.java
>> for enabling the test only for allowed configurations.
>>
>> Cheers,
>>
>> -JB-
>>
>>
>>>
>>>
>>> -Harsha
>>>
>>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/serviceability-dev/attachments/20160426/b8c68cec/attachment-0001.html>