On 11/ 6/11 08:05 PM, David Holmes wrote:
Hi Gary,
On 4/11/2011 11:36 PM, Gary Adams wrote:
I've started to look at timing related bugs that have been open
for a while, but have not had sufficient priority to make it to the
top of the list of bugs to be fixed. Thought I'd start with some
low hanging fruit with simple bug fixes.
Sometimes apparently low-hanging fruit is attached to the tree with steel wire
;-) (and sometimes barbed wire at that)
Had a feeling when I pulled on the fruit the trap door
would open in the floor ...
6731620: TEST_BUG: java/util/Timer/Args.java is too optimistic about the
execution time of System.out.printf
This seems like a simply problem to avoid two calls to get the current time
and to eliminated the time to process the print statement
from the evaluation of the test elapsed time.
Replacing this sequence ;
System.out.printf("elapsed=%d%n", System.currentTimeMillis() - start);
check(System.currentTimeMillis() - start < 500);
with
elapsed = System.currentTimeMillis() - start;
System.out.printf("elapsed=%d%n", elapsed);
check(elapsed < 500);
This seems reasonable on the face of it. But a couple of other observations
regarding the test code:
First the 500 is somewhat arbitrary - the test is preparing three timertasks
with initial expiry times in the past and expecting them to all execute
"immediately" (one multiple times). This immediacy is translated into
"completes within 500ms". These type of timing constants should be eradicated
if possible, or preferably made configurable if not. It is a difficult problem
to address if you want to try and detect something taking an unexpectedly long
time, but have no idea what your execution environment will be.
I think I'm signing up for making this one test reliable and
not taking on all of the bad timeout assumptions. The 500 millisecond
constraint seems reasonable in this case for the work expected to be
completed.
Second, the current code won't detect a true failure that results in a hang -
the testcase will hang and presumably get timed-out by the test harness. It
might be better to apply a timeout to the await() calls instead and fail only
if they do timeout.
I thought the hung test was the responsibility of the test harness.
It makes for simpler individual tests if that responsibility is
delegated to the harness.
Third, in this case it would also be prudent to (re-)read the start time after
the setup has been done ie after creating the countDownLatches. If this is run
in samevm mode or agentvm mode those constructions may trigger a full GC and
the ensuing delay could cause the test to again fail.
Sounds reasonable I'll add that change to the fix.
That all said, I'd see combining your suggested fix with moving the point at
which start is measured as definite improvements in the test.
Agreed.
I plan to test the fix on a 300MHz linux/arm device.
I'll provide a proper webrev as soon as I have author rights
confirmed. I'm looking for reviewer and a committer,
once I get the fix tested locally.
I presume by "committer" you mean somebody to actually do the push for you - I
think we generally refer to that as a "sponsor" (even though not part of
OpenJDK terminology). In any case that should be someone from TL group and
I'll let them respond further on that and provide any additional review comments.
I'm happy to act as another Reviewer.
Thanks.
Cheers,
David
Thanks
Gary Adams
