[wildfly-dev] issues with IBM JVM on Linux

[Alexey Loubyansky]

Thanks. I tried -Xnojit (the tests did take significantly more time to 
complete) but it didn't affect the results. The testsuite failed 3 times 
out of 5 runs with the same errors.

Alexey

On 09/25/2015 08:14 PM, Radoslaw Rodak wrote:
> Hi
>
> Few months ago I had pretty similar issue on IBM JDK , but version 1.7.0 .
>
> Doing 10 time same xslt transformation randomly it was failing.
>
> Problem was JIT Compiler Bug in IBM JDK…. issue was fix in newer Version of IBM JDK.
> Best ist to get newest IBM JDK Version.
>
> Work Arround for me on IBM JDK was to use  TransformerFactory , which is used as default on Oracle JDK on Linux and Windows, which is not the case on IBM JDK...:
>
> -Djavax.xml.transform.TransformerFactory=org.apache.xalan.processor.TransformerFactoryImpl
>
> To see if the error is really because of IBM JIT ( begin of this year IBM JDK had a lot of JIT bugfixes… ) you can also just turn off JIT ( -Xnojit ) and see if the problem still happened ( side effect test will run very slow )
>
> P.S:
>
> Interesting side Effect was, transformation was 50% faster with Xalan Implementation compared to default IBM TransformerFactory …
>
> Cheers
> Radek
>
> Am 25.09.2015 um 18:39 schrieb Alexey Loubyansky <alexey.loubyansky at redhat.com>:
>
>> On 09/25/2015 03:18 PM, Scott Marlow wrote:
>>> Are you using the latest JVM?  From [1], it looks like the last update
>>> was on July 17th 2015.
>>
>> No, I've been using the one we have installed in brontes to be able to
>> reproduce.
>>
>> But I've tried the latest and it still fails occasionally although less
>> often: once or twice in 10 runs. So it appears to be better than the one
>> we've been using. Tomaz is going to upgrade it.
>>
>>> Does it make a difference whether the classes were generated by the IBM
>>> versus Oracle versus OpenJDK compiler?
>>
>> Haven't tried that yet.
>>
>> Thanks,
>> Alexey
>>
>>> Scott
>>>
>>> [1] http://www.ibm.com/developerworks/java/jdk/linux/download.html
>>>
>>> On 09/25/2015 09:08 AM, Alexey Loubyansky wrote:
>>>> I've been trying to narrow down sporadic failures in the wfcore patching
>>>> tests on the IBM JVM on Linux (version info below) for the last few
>>>> days. I wanted to share this because what I see doesn't make sense to
>>>> me. And I am not sure whether other modules have similar issues.
>>>>
>>>> So, in patching tests the (sporadic) problem appears when parsing XML
>>>> attributes. E.g. if there is an element with attributes 'name' and
>>>> 'add-on', sometimes the value of name will end up containing the value
>>>> of add-on attribute. Here is the actual code for this (Attribute is an
>>>> enum) from org.jboss.as.patching.metadata.PatchXmlUtils:
>>>>
>>>> String name = null;
>>>> boolean isAddOn = false;
>>>> final int count = reader.getAttributeCount();
>>>> for (int i = 0; i < count; i++) {
>>>>        final String value = reader.getAttributeValue(i);
>>>>        final Attribute attribute =
>>>> Attribute.forName(reader.getAttributeLocalName(i));
>>>>        if(Attribute.NAME == attribute) {
>>>>            name = value;
>>>>        } else if (Attribute.ADD_ON == attribute) {
>>>>            isAddOn = Boolean.valueOf(value);
>>>>        } else {
>>>>            throw unexpectedAttribute(reader, i);
>>>>        }
>>>> }
>>>>
>>>> So, sometimes the tests fail because the value of name is 'true',
>>>> apparently. This is just one example. It affects parsing attributes in
>>>> general in this class.
>>>>
>>>> I noticed that if I do anything with variable 'value' in the loop (or
>>>> pass it as an argument to some method and actually do something with it
>>>> in that method), e.g. System.err.println("value=" + value); the problem
>>>> goes away.
>>>>
>>>> Another workaround I've found is to inline the variable, i.e. use
>>>> reader.getAttributeValue(i) directly instead of using a variable.
>>>>
>>>> By the "problem goes away" I mean that I haven't seen these kind of
>>>> failures anymore with the changes I mentioned. To give some statistics,
>>>> with the original code the tests would fail from 2 to 5 times out of 10
>>>> runs.
>>>> With the inlined variable value they pass 100%. I've be running them for
>>>> 3 days now.
>>>>
>>>> Here is the JVM info.
>>>>
>>>> java version "1.8.0"
>>>> Java(TM) SE Runtime Environment (build pxi3280-20150129_02)
>>>> IBM J9 VM (build 2.8, JRE 1.8.0 Linux x86-32 20150116_231420 (JIT
>>>> enabled, AOT enabled)
>>>> J9VM - R28_Java8_GA_20150116_2030_B231420
>>>> JIT  - tr.r14.java_20150109_82886.02
>>>> GC   - R28_Java8_GA_20150116_2030_B231420
>>>> J9CL - 20150116_231420)
>>>> JCL - 20150123_01 based on Oracle jdk8u31-b12
>>>>
>>>>
>>>> Any ideas or suggestions?
>>>>
>>>>
>>>> Thanks,
>>>> Alexey
>>>> _______________________________________________
>>>> wildfly-dev mailing list
>>>> wildfly-dev at lists.jboss.org
>>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>>>
>>> _______________________________________________
>>> wildfly-dev mailing list
>>> wildfly-dev at lists.jboss.org
>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>>
>> _______________________________________________
>> wildfly-dev mailing list
>> wildfly-dev at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>