RFR: 2178143: VM crashes if the number of bound CPUs changed during runtime

David Holmes david.holmes at oracle.com
Wed Mar 20 18:35:04 PDT 2013


Correction: we have two API's concerning the processor count: one static 
and one dynamic. We actually use the dynamic version in a number of 
places including Runtime.availableProcessors() - so libraries and apps 
can act dynamically (but most won't eg default ForkJoinPool will be 
sized based on active processor count at VM initialization time).

That makes the "NumberOfProcessors" version even more 
complex/problematic. Does it mean the static count or the dynamic count? 
Might different use cases want different answers to that?

I think there is no clear answer for how a general "NumberOfProcessors" 
variable might be used, and I don't expect to have one in the short to 
mid term for JDK 8.

To that end I suggest simply changing this to AssumeMP and use it to 
only affect is_MP. That is not incompatible with adding 
NumberOfProcessors later on, if general desirable semantics can be 
worked out.

But there is still a question, as raised in the CR as to what the 
default for this should be. It is argued in the CR that in the pervasive 
multi-core world, we should AssumeMP, and allow it to be disabled by 
apps that are truly run on uni-processors and which would get a 
performance hit otherwise. Though we might want to only do that with SE 
rather than Embedded.

David

On 21/03/2013 11:02 AM, David Holmes wrote:
> On 21/03/2013 7:27 AM, Yumin Qi wrote:
>> Hi, can I have your code review of a small change?
>
> Not really small conceptually. :)
>
> I don't think this form of the fix addresses the underlying issue as
> discussed in the bug report. If the variable was renamed
> MinimumNumberOfProcessors, or MinimumAssumedProcessors, then simply
> using it to turn on is_MP would be okay. Such a flag would suit the
> initial problem perfectly. Of course any value >2 would be semantically
> indistinct, so this really acts as a boolean flag - AssumeMP.
>
> The more general NumberOfProcessors approach, which I'm still unsure of,
> should to me control what is reported for available-processors. That way
> it would affect everything in the VM, libraries and application code
> that configures itself based on the number of available processors. The
> main usecase for that, in my opinion, would be for apps running on large
> systems but you want to constrain it to using a subset of the physical
> CPUs (without having to configure processor sets). That is a different
> kind of problem and a different kind of flag. Using NumberOfProcessors
> but not having it control anything except is_MP just seems wrong - and
> using it to replace available_processors is not a complex change.
>
> The VM is not designed for dynamic adaptation of threads/pools so if the
> number of processors does change dynamically neither of the above
> options are going to provide solutions to the potential performance
> problems that will be encountered (too many or too few threads). Any
> apps that starts on single core (as reported by the OS) is going to be
> under-provisioned.
>
> David
> -----
>
>> 2178143:  VM crashes if the number of bound CPUs changed during runtime.
>>
>> Situation: Customer first configure only one CPU online and turn others
>> offline to run java application, after java program started, bring more
>> CPUs back online. Since VM started on a single CPU, os::is_MP() will
>> return false, but after more CPUs available, OS will schedule the app
>> run on multiple CPUs, this caused SEGV in various places where data
>> consistency was broken. The solution is supply a flag to assume it is
>> running on MP, so lock is forced to be called.
>>
>> http://cr.openjdk.java.net/~minqi/2178143/
>>
>> Thanks
>> Yumin


More information about the hotspot-gc-dev mailing list