RFR: 8222182: ZGC: Use SoftMaxHeapSize to guide GC heuristics

[Thomas Schatzl]

Hi,

On Thu, 2019-04-11 at 12:30 +0200, Thomas Schatzl wrote:
> Hi,
> 
> On Wed, 2019-04-10 at 19:23 +0200, Per Liden wrote:
> > The ZDriector currently uses max heap size (-Xmx) as input when
> > deciding 
> > on when to start a GC cycle. We should start using SoftMaxHeapSize 
> > instead. Of course, this change depends on JDK-8222145 "Add 
> > -XX:SoftMaxHeapSize flag", which is currently out for review.
> > 
> > Bug: https://bugs.openjdk.java.net/browse/JDK-8222182
> > Webrev: http://cr.openjdk.java.net/~pliden/8222182/webrev.0
> 
>   to be able to review this, what is the expected reaction (and
> limitations) of the collector on this, particularly in unusual cases,
> like what should happen if an allocation request goes over the soft
> limit?
> 
> From what I understand, allocation of ZPages will fail as normal, and
> an allocation stall will occur as normal. This does not sound like a
> "soft" max heap size.
> 
> Which also means that there is opportunity for some allocating Java
> threads not having recognized the new value of SoftMaxHeapSize as it
> is read without synchronization - so some allocations will go beyond
> that soft limit, while others may fail (and seem to cause an
> allocation stall).
> 
> Neither the flag CSR nor the CR talks about this.
> 

The code in ZPageAllocator::max_available() is especially fishy: if
current_max_capacity() is smaller than used (that value can be set to
anything between -Xms and -Xmx), then the result of the available
calculation overflows and seems to result in no action until used()
gets down below SoftMaxHeapSize.

I would expect that the collector starts a GC cycle to hopefully get
used() down to this new SoftMaxHeapSize.

Also, after the user setting the new value, how long would it take for
the collector to recognize the new value? Is it possible that it never
acts on it (i.e. no Zpage allocation at all)? It would really be
interesting to document this to set the expectations for the user.
Afaik there is a regularly scheduled gc cycle in ZGC based on mutator
activity, but that may take how long?

Please document expected behavior depending on current state (at least
heap usage) when various values are set by the user. Otherwise there is
no way to determine whether ZGC is working correctly.

Or maybe point out a location on the wiki where this is documented.

That change would also _at least_ be worth a release note imho to get
at least some recognition (if not a JEP ;))

> What testing has been done for this? There is no test.

Thanks,
  Thomas