RFR: 8241828: JFR: Some streaming tests require a larger heap size with ZGC

[Stefan Karlsson]

On 2020-03-31 10:43, Erik Gahlin wrote:
> Doesn’t look good.
> 
> One problem is that TestChunkGap has TestFilledChunks in @run main. If 
> you change it to TestChunkGap, that test should be fine.

OK. So we found an old bug lurking here. It's not the first time I see 
this kind of bugs. It would be nice if we could detect those @run 
copy-n-paste bugs.

> 
> When it comes to TestFilledChunk, I can now see how this can happen due 
> to events being sorted by default and the high rate events are emitted, 
> You could try to sett s.setSorted(false), and it should be fine as well.
> 
>   63         try (EventStream s = EventStream.openRepository()) {
>   64             try (Recording r1 = new Recording()) {
> +65 s.setSorted(false);
>   66                 s.setStartTime(Instant.EPOCH);
> 

That seems to be the code in TestChunkGap, that now (after the fix 
above) works OK with ZGC. Do you have a similar fix for TestFilledCunk?

Thanks,
StefanK

> 
> Thanks
> Erik
> 
>> On 31 Mar 2020, at 10:08, Stefan Karlsson <stefan.karlsson at oracle.com 
>> <mailto:stefan.karlsson at oracle.com>> wrote:
>>
>> Did some extra testing on TestChunkGap and G1:
>>
>> With compressed oops:
>> -Xmx300m : OOME
>> -Xmx350m : Barely passes
>>
>> Without compressed oops:
>> -Xmx450m : OOME
>> -Xmx500m : Passes
>>
>> I issued a 'jcmd GC.class_histogram' in the middle of the run and got:
>>
>> num     #instances         #bytes  class name (module)
>> -------------------------------------------------------
>>   1:       2473067      158635784  [Ljava.lang.Object; 
>> (java.base at 15-internal)
>>   2:       2463224      118234752  jdk.jfr.consumer.RecordedEvent 
>> (jdk.jfr at 15-internal)
>>   3:       4894598      117470352  java.lang.Integer 
>> (java.base at 15-internal)
>>   4:           157       39960680  [Ljdk.jfr.consumer.RecordedEvent; 
>> (jdk.jfr at 15-internal)
>>   5:        369731       18447144  [B (java.base at 15-internal)
>>   6:        368815       11802080  java.lang.String 
>> (java.base at 15-internal)
>>   7:           331         584344  [I (java.base at 15-internal)
>> ...
>>
>> Does this look right?
>>
>> StefanK
>>
>> On 2020-03-30 22:03, Erik Gahlin wrote:
>>> Seems strange that the test would require over 512 MB. I would expect 
>>> the TestChunkGap not to use a live set of more than 25 MB. The test 
>>> only emits three events.
>>> TestFilledChunks is easier to understand as it may emit millions of 
>>> events. Still it must be a bug in JFR then, because it should not 
>>> keep them around.
>>> Erik
>>>> On 30 Mar 2020, at 19:24, mikhailo.seledtsov at oracle.com 
>>>> <mailto:mikhailo.seledtsov at oracle.com> 
>>>> <mailto:mikhailo.seledtsov at oracle.com> wrote:
>>>>
>>>> Hi Stefan,
>>>>
>>>> I would recommend adding "@requires os.maxMemory > " to the test, to 
>>>> make sure that test does not execute on a host/node that does not 
>>>> have sufficient memory. E.g. "@requires os.maxMemory > 1G"
>>>>
>>>> Otherwise the change looks good to me.
>>>>
>>>>
>>>> Thanks,
>>>>
>>>> Misha
>>>>
>>>> On 3/30/20 4:07 AM, Stefan Karlsson wrote:
>>>>> Hi all,
>>>>>
>>>>> Please review this small patch to increase the max heap size of a 
>>>>> couple of JFR streaming tests.
>>>>>
>>>>> https://cr.openjdk.java.net/~stefank/8241828/webrev.01/
>>>>> https://bugs.openjdk.java.net/browse/JDK-8241828
>>>>>
>>>>> The default heap size is 512m when these tests are run. G1 uses 
>>>>> almost 500m, but ZGC needs a bit more. I propose that we set the 
>>>>> max heap size to 768m.
>>>>>
>>>>> Thanks,
>>>>> StefanK
>