RFR: [XS] 8228658: test GetTotalSafepointTime.java fails on fast Linux machines with Total safepoint time 0 ms
Jean Christophe Beyler
jcbeyler at google.com
Tue Jul 30 23:08:23 UTC 2019
FWIW, I would have done something like what David was suggesting, just
slightly tweaked:
public static long executeThreadDumps() {
long value;
long initial_value = mbean.getTotalSafepointTime();
do {
Thread.getAllStackTraces();
value = mbean.getTotalSafepointTime();
} while (value == initial_value);
return value;
}
This ensures that the value is a new value as opposed to the current value
and if something goes wrong, as David said, it will timeout; which is ok.
But I come back to not really understanding why we are doing this at this
point of relaxing (just get a new value of safepoint time). Because, if we
accept timeouts now as a failure here, then really the whole test becomes:
executeThreadDumps();
executeThreadDumps();
Since the first call will return when value > 0 and the second call will
return when value2 > value (I still wonder why we want to ensure it works
twice...).
So both failures and even testing for it is kind of redundant, once you
have a do/while until a change?
Thanks,
Jc
On Tue, Jul 30, 2019 at 2:35 PM David Holmes <david.holmes at oracle.com>
wrote:
> On 30/07/2019 10:39 pm, Baesken, Matthias wrote:
> > Hi David, "put that whole code (the while loop) in a helper method."
> was JC's idea, and I like the idea .
>
> Regardless I think the way you are using NUM_THREAD_DUMPS is really
> confusing. As an all-caps static you'd expect it to be a constant.
>
> Thanks,
> David
>
> > Let's see what others think .
> >
> >>
> >> Overall tests like this are not very useful, yet very fragile.
> >>
> >
> > I am also fine with putting the test on the exclude list.
> >
> > Best regards, Matthias
> >
> >
> >> -----Original Message-----
> >> From: David Holmes <david.holmes at oracle.com>
> >> Sent: Dienstag, 30. Juli 2019 14:12
> >> To: Baesken, Matthias <matthias.baesken at sap.com>; Jean Christophe
> >> Beyler <jcbeyler at google.com>
> >> Cc: hotspot-dev at openjdk.java.net; serviceability-dev <serviceability-
> >> dev at openjdk.java.net>
> >> Subject: Re: RFR: [XS] 8228658: test GetTotalSafepointTime.java fails
> on fast
> >> Linux machines with Total safepoint time 0 ms
> >>
> >> Hi Matthias,
> >>
> >> On 30/07/2019 9:25 pm, Baesken, Matthias wrote:
> >>> Hello JC / David, here is a second webrev :
> >>>
> >>> http://cr.openjdk.java.net/~mbaesken/webrevs/8228658.1/
> >>>
> >>> It moves the thread dump execution into a method
> >>> executeThreadDumps(long) , and also adds while loops (but with a
> >>> limitation for the number of thread dumps, really don’t
> >>> want to cause timeouts etc.). I removed a check for
> >>> MAX_VALUE_FOR_PASS because we cannot go over Long.MAX_VALUE .
> >>
> >> I don't think executeThreadDumps is worth factoring out like out.
> >>
> >> The handling of NUM_THREAD_DUMPS is a bit confusing. I'd rather it
> >> remains a constant 100, and then you set a simple loop iteration count
> >> limit. Further with the proposed code when you get here:
> >>
> >> 85 NUM_THREAD_DUMPS = NUM_THREAD_DUMPS * 2;
> >>
> >> you don't even know what value you may be starting with.
> >>
> >> But I was thinking of simply:
> >>
> >> long value = 0;
> >> do {
> >> Thread.getAllStackTraces();
> >> value = mbean.getTotalSafepointTime();
> >> } while (value == 0);
> >>
> >> We'd only hit a timeout if something is completely broken - which is
> fine.
> >>
> >> Overall tests like this are not very useful, yet very fragile.
> >>
> >> Thanks,
> >> David
> >>
> >>> Hope you like this version better.
> >>>
> >>> Best regards, Matthias
> >>>
> >>> *From:*Jean Christophe Beyler <jcbeyler at google.com>
> >>> *Sent:* Dienstag, 30. Juli 2019 05:39
> >>> *To:* David Holmes <david.holmes at oracle.com>
> >>> *Cc:* Baesken, Matthias <matthias.baesken at sap.com>;
> >>> hotspot-dev at openjdk.java.net; serviceability-dev
> >>> <serviceability-dev at openjdk.java.net>
> >>> *Subject:* Re: RFR: [XS] 8228658: test GetTotalSafepointTime.java fails
> >>> on fast Linux machines with Total safepoint time 0 ms
> >>>
> >>> Hi Matthias,
> >>>
> >>> I wonder if you should not do what David is suggesting and then put
> that
> >>> whole code (the while loop) in a helper method. Below you have a
> >>> calculation again using value2 (which I wonder what the added value of
> >>> it is though) but anyway, that value2 could also be 0 at some point,
> no?
> >>>
> >>> So would it not be best to just refactor the getAllStackTraces and
> >>> calculate safepoint time in a helper method for both value / value2
> >>> variables?
> >>>
> >>> Thanks,
> >>>
> >>> Jc
> >>>
> >>> On Mon, Jul 29, 2019 at 7:50 PM David Holmes <david.holmes at oracle.com
> >>> <mailto:david.holmes at oracle.com>> wrote:
> >>>
> >>> Hi Matthias,
> >>>
> >>> On 29/07/2019 8:20 pm, Baesken, Matthias wrote:
> >>> > Hello , please review this small test fix .
> >>> >
> >>> > The test
> >>>
> >> test/jdk/sun/management/HotspotRuntimeMBean/GetTotalSafepointTime.
> >> java
> >>> fails sometimes on fast Linux machines with this error message :
> >>> >
> >>> > java.lang.RuntimeException: Total safepoint time illegal
> value: 0
> >>> ms (MIN = 1; MAX = 9223372036854775807)
> >>> >
> >>> > looks like the total safepoint time is too low currently on
> these
> >>> machines, it is < 1 ms.
> >>> >
> >>> > There might be several ways to handle this :
> >>> >
> >>> > * Change the test in a way that it might generate nigher
> >>> safepoint times
> >>> > * Allow safepoint time == 0 ms
> >>> > * Offer an additional interface that gives safepoint
> times
> >>> with finer granularity ( currently the HS has safepoint time
> values
> >>> in ns , see jdk/src/hotspot/share/runtime/safepoint.cpp
> >>> SafepointTracing::end
> >>> >
> >>> > But it is converted on ms in this code
> >>> >
> >>> > 114jlong RuntimeService::safepoint_time_ms() {
> >>> > 115 return UsePerfData ?
> >>> > 116
> >>> Management::ticks_to_ms(_safepoint_time_ticks->get_value()) : -1;
> >>> > 117}
> >>> >
> >>> > 064jlong Management::ticks_to_ms(jlong ticks) {
> >>> > 2065 assert(os::elapsed_frequency() > 0, "Must be non-zero");
> >>> > 2066 return (jlong)(((double)ticks /
> >>> (double)os::elapsed_frequency())
> >>> > 2067 * (double)1000.0);
> >>> > 2068}
> >>> >
> >>> >
> >>> >
> >>> > Currently I go for the first attempt (and try to generate
> >>> higher safepoint times in my patch) .
> >>>
> >>> Yes that's probably best. Coarse-grained timing on very fast
> machines
> >>> was bound to eventually lead to problems.
> >>>
> >>> But perhaps a more future-proof approach is to just add a
> do-while loop
> >>> around the stack dumps and only exit when we have a non-zero
> >> safepoint
> >>> time?
> >>>
> >>> Thanks,
> >>> David
> >>> -----
> >>>
> >>> > Bug/webrev :
> >>> >
> >>> > https://bugs.openjdk.java.net/browse/JDK-8228658
> >>> >
> >>> > http://cr.openjdk.java.net/~mbaesken/webrevs/8228658.0/
> >>> >
> >>> >
> >>> > Thanks, Matthias
> >>> >
> >>>
> >>>
> >>> --
> >>>
> >>> Thanks,
> >>>
> >>> Jc
> >>>
>
--
Thanks,
Jc
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.java.net/pipermail/serviceability-dev/attachments/20190730/63420f78/attachment-0001.html>
More information about the serviceability-dev
mailing list