On 10/4/18, 12:13 PM, Daniel D. Daugherty wrote:
On 10/4/18 12:04 PM, Gary Adams wrote:
We currently use time factor 4 or 10 to scale up the jtreg tests.
The change I proposed was already 10x100 = 1000.
Agreed that you are already bumping it up by 10X.
Since this is just a thread scheduling issue to let the debuggee
run til it blocks, it should not require more time to get the
switch to take place.
Are you really trying to say that system load doesn't affect
thread scheduling time?
In this particular test, the main thread and the debuggee thread
are toggling back and forth around several monitors and
synchronized code blocks. Not sure the tests are written
for an arbitrarily high system load.
If this was a process scheduling issue, then more complex
coordination might be worth considering.
System load doesn't just affect process scheduling. It also
affects thread scheduling. If all the cores are busy, then
running another thread or another process is impacted.
In any case, I wasn't clear in my suggestion.
- Thread.sleep(100);
+ Thread.sleep(100 * timeoutFactor);
assuming there is some easy way to get access to the -timeoutFactor
parameter.
Currently, timeoutFactor has not been introduced to the vmTestbase/nsk
test suite.
Dan
On 10/4/18, 11:45 AM, Daniel D. Daugherty wrote:
Sleeps don't scale under load. That said, I agree with Chris that
the code is already in place. One possible addition is to scale
the sleep value by the timeout factor for the test. That will
further reduce the likelihood of intermittent failures.
Dan
On 10/4/18 7:39 AM, Gary Adams wrote:
Oops, wrong comment used in the patch.
Fresh patch attached.
On 10/4/18, 7:11 AM, Gary Adams wrote:
Patch attached.
I think one reviewer is sufficient for a trivial patch.
On 10/3/18, 4:49 PM, Chris Plummer wrote:
Hi Gary,
Although I don't like relying on timer delays for stuff like
this, the code for it is already in place, so I'm ok with making
the delay longer to make sure there is contention on the monitor.
Could you update the comment to read "// pause to provoke
contention on thread.endingMonitor"
thanks,
Chris
On 10/3/18 11:55 AM, Gary Adams wrote:
While running a block of nsk/jvmti/scenarios tests, I noticed an
occasional failure
for cm02t001 in windows debug platform. After enabling the nsk
verbose
diagnostics and adding a few messages in the main test and the
debuggee
thread, it became clear that the missing contention was due to
the main thread
getting ahead of the debugee thread.
The call to letFinish() below let's the deuggee thread wake up
from it's wait
and proceed to the contention for the endingMonitor. If the main
thread
waits a little longer it should reach the debuggee thread
synchronized block.
I reopened an earlier bug that was closed as CNR.
Issue: https://bugs.openjdk.java.net/browse/JDK-8036026
diff --git
a/test/hotspot/jtreg/vmTestbase/nsk/jvmti/scenarios/capability/CM02/cm02t001.java
b/test/hotspot/jtreg/vmTestbase/nsk/jvmti/scenarios/capability/CM02/cm02t001.java
---
a/test/hotspot/jtreg/vmTestbase/nsk/jvmti/scenarios/capability/CM02/cm02t001.java
+++
b/test/hotspot/jtreg/vmTestbase/nsk/jvmti/scenarios/capability/CM02/cm02t001.java
@@ -82,7 +82,7 @@
thread.letFinish();
// pause to provoke contention
- Thread.sleep(100);
+ Thread.sleep(1000);
} catch (InterruptedException e) {
throw new Failure(e);
}
