Review request:7197210: java/lang/invoke/CallSiteTest.java failing on armsflt

[Chris Plummer]

On 10/28/12 12:00 PM, hotspot-dev-request at openjdk.java.net wrote:
> On 27/10/2012 00:52, Jiangli Zhou wrote:
>> >Hi Alan,
>> >
>> >I tried a few experiments with the makefiles. Looks like
>> >jdk/make/java/invoke/Makefile is not the right place to set the
>> >-timeoutFactor option. I'm reluctant to set a large timeoutFactor in
>> >jdk/test/Makefile. Some of the java.lang.invoke jtreg test take long
>> >time to run. For example, the test.java.lang.invoke.MethodHandlesTest
>> >takes about 1.5 hour to run on certain devices. Since there are only
>> >3~4 java.lang.invoke tests have the timeout issue and each of them has
>> >very different execution duration, I'm inclined to set each specific
>> >timeout value for different test. Please let me know your opinion.
>> >
>> >Thanks,
>> >Jiangli
>> >
> If you are running the tests via the make file then this should work:
>
> make EXTRA_JTREG_OPTIONS=-timeoutFactor:10 jdk_lang
>
> but I see that the Makefile puts its own defaults,
> including-timeoutFactor:4, after the options set via
> EXTRA_JTREG_OPTIONS. That plus jtreg doesn't allow the -timeoutFactor to
> be specified more than once on the command line. Minimally we should fix
> the make file so that options specified via EXTRA_JTREG_OPTIONS override
> any defaults in the make file.
>
> I think you'll need to check with Christian or John as to whether they
> would object to have very high timeouts on these tests. Personally I
> think it should be possible to specify a timeout scaling factor when
> running the tests rather than having each test specify a /timeout for
> the slowest possible machine that the test might run on.
>
> -Alan.
It seems we need a combination of both. Longer tests need a scaling 
factor, as do slower devices. They should both be combined to make sure 
the longer running tests will complete on the slower devices. I think 
the "timeout" value in the test serves the former, and the jtreg 
-timeoutFactor option serves the later. However, I'm not sure if SQE is 
using -timeoutFactor for the slower devices. If they are not, then we 
should ask that they start using it. It might save the need to add or 
increase the "timeout" for many tests. For CallSiteTest.java, since it 
is a long running test, it should have a timeout added to it.

I don't see we would ever want to specify -timeoutFactor in a makefile 
when it will be equally applied to all tests and all devices. This 
requires it be set in such a way that the longest running test will 
complete on the slowest device. This will cause valid timeouts to take 
too long to happen on faster devices and short running tests.

Chris