[concurrency-interest] padding in Exchanger

Nathan Reynolds nathan.reynolds at oracle.com
Tue Jan 17 14:12:54 EST 2012


 > How would the processor even tell the JVM about this? Read some 
register? How does the cpu report this? As a percentage?

Intel's x86 processors have hardware performance counters.  I can't 
remember which counter to use.  I think the counter captures this 
information: last level cache miss when the cache line was modified by 
another core.  The counter tells us the data address of the miss and 
some other key information... which I can't remember at the moment.

 > How does it notify you about multiple memory addresses being contended?

The JVM would have to poll the counter periodically and see what is 
there.  Since the counter is changing often, the JVM will see multiple 
addresses show up.

 > Would this register be per CPU?

If the correct performance counter is the one I am thinking of, then 
this performance counter is per processor socket.

 > When does the VM poll this?

I am not sure.  I would guess just on a timer.  The JVM could poll it 
right after a profiled method executes.

 > Etc as you say this is rare so I think the cost/benefit of this vs 
developer annotation isn't in favor of it.

True and false sharing happens a lot.  Its just that it happens on very 
few objects.  The JVM has to fix only those few objects and the whole 
program benefits greatly.

Nathan Reynolds 
<http://psr.us.oracle.com/wiki/index.php/User:Nathan_Reynolds> | 
Consulting Member of Technical Staff | 602.333.9091
Oracle PSR Engineering <http://psr.us.oracle.com/> | Server Technology

On 1/17/2012 12:02 PM, Vitaly Davidovich wrote:
>
> Yes the expensive op would be field contention requiring re-layout and 
> probably object movement.  How would the processor even tell the JVM 
> about this? Read some register? How does the cpu report this? As a 
> percentage? How does it notify you about multiple memory addresses 
> being contended? How do you know which field in the cache line is 
> actually the culprit? Would this register be per CPU? When does the VM 
> poll this? Etc as you say this is rare so I think the cost/benefit of 
> this vs developer annotation isn't in favor of it.  But it's not for 
> me to implement :).
>
> Sent from my phone
>
> On Jan 17, 2012 1:33 PM, "Nathan Reynolds" <nathan.reynolds at oracle.com 
> <mailto:nathan.reynolds at oracle.com>> wrote:
>
>     If an application is contending on a cache line, then the
>     processors will look 100% busy but very little work will get
>     done.  Fixing the contention will almost always get huge gains.
>
>     Let's say the contention is due to 2 objects being next to each
>     other.  The solution is to separate them by moving 1 object to
>     another place in the heap.  Moving an object in the heap is fairly
>     cheap.  GC does it a lot!
>
>     Let's say the contention is due to 2 fields being next to each
>     other.  The solution is to separate them by rearranging the
>     fields.  This is probably expensive.  Once that is done, the
>     contention will go away but the memory will increase.  Very few
>     locks actually contend.  Even fewer fields contend on cache
>     lines.  So, the memory impact should be fairly small and probably
>     insignificant.
>
>     Nathan Reynolds
>     <http://psr.us.oracle.com/wiki/index.php/User:Nathan_Reynolds> |
>     Consulting Member of Technical Staff | 602.333.9091 <tel:602.333.9091>
>     Oracle PSR Engineering <http://psr.us.oracle.com/> | Server Technology
>
>     On 1/17/2012 11:03 AM, Vitaly Davidovich wrote:
>>
>>     Doing dynamic re-layout and contention adjustment seems nice in
>>     thought but how practical is that? I can't see how that would be
>>     cheap enough where it's worth the cost. What if the app goes
>>     through phases of contention? Initially high but then no sharing
>>     - would the bloated object layout be worth it at that point?
>>     Seems like this is a place where explicit developer instructions
>>     is better than heuristics with potentially expensive consequences.
>>
>>     Sent from my phone
>>
>>     On Jan 17, 2012 12:44 PM, "Nathan Reynolds"
>>     <nathan.reynolds at oracle.com <mailto:nathan.reynolds at oracle.com>>
>>     wrote:
>>
>>         Assuming that the JVM can optimize for true and false sharing
>>         (and that is a big assumption at the moment), then you can
>>         focus your time on writing useful code.  Furthermore,
>>         optimizing for true and false sharing will never be able to
>>         fix actual data contention.  We still need clever ways of
>>         sharing data without bottlenecks.
>>
>>         Nathan Reynolds
>>         <http://psr.us.oracle.com/wiki/index.php/User:Nathan_Reynolds> |
>>         Consulting Member of Technical Staff | 602.333.9091
>>         <tel:602.333.9091>
>>         Oracle PSR Engineering <http://psr.us.oracle.com/> | Server
>>         Technology
>>
>>         On 1/17/2012 10:29 AM, Ruslan Cheremin wrote:
>>>         You made my day. Few months ago I was dreaming (in my blog) about
>>>         complexity of false sharing prevention with padding. And come to two
>>>         options, one better another. First one was @PreventFalseSharing
>>>         annotation, next was atomatically contention detection and relocation
>>>         of contended objects by JIT. Readers of my blog soon pointed me to
>>>         @Contended annotation. And now you telling the second -- the best --
>>>         option is also being explored!
>>>
>>>         Just want to ask -- if all good things will be done by JIT -- what I
>>>         will be paid for?
>>>
>>>         2012/1/17 Nathan Reynolds<nathan.reynolds at oracle.com>  <mailto:nathan.reynolds at oracle.com>:
>>>>         It would be nice if the processor could effectively tell the JVM that false
>>>>         sharing is happening.  It would be nice if the JVM could respond by moving
>>>>         objects within the heap or fields with the class to avoid false sharing.
>>>>         Thus, we don't have to pad or worry about placing @Contended or other
>>>>         attributes into the class.
>>>>
>>>>         Intel was looking into to optimizing for true and false sharing.  They had a
>>>>         prototype that worked but required restarting the JVM.  Oracle was looking
>>>>         into dynamically relayout fields in objects.  I haven't heard anything from
>>>>         either group for a while...  I haven't asked either.  *IF* a solution
>>>>         becomes available, then it will be a while.  This is a very difficult thing
>>>>         to do.
>>>>
>>>>         Nathan Reynolds | Consulting Member of Technical Staff |602.333.9091  <tel:602.333.9091>
>>>>         Oracle PSR Engineering | Server Technology
>>>>
>>>>         On 1/17/2012 9:35 AM, Vitaly Davidovich wrote:
>>>>
>>>>         OK I see what you mean now.  I imagine @Contended will be used with fields
>>>>         rather than classes so when the JVM lays out an instance of the class I
>>>>         assume it will do two-sided padding on the contended field if required or if
>>>>         natural layout is such that prior fields already fill up a cache line then
>>>>         only one sided is needed.
>>>>
>>>>         Sent from my phone
>>>>
>>>>         On Jan 17, 2012 11:27 AM, "Ruslan Cheremin"<cheremin at gmail.com>  <mailto:cheremin at gmail.com>  wrote:
>>>>>         2012/1/17 Vitaly Davidovich<vitalyd at gmail.com>  <mailto:vitalyd at gmail.com>:
>>>>>>         I think it's semantics - if you sometimes allocate with 64/128 byte
>>>>>>         alignment then if your object is smaller than 64/128 the rest of the
>>>>>>         space
>>>>>>         is effectively padding.
>>>>>         Agree. But in case of alignment you lose sense of "one-side" or "two
>>>>>         side" padding -- you do not need "two side padding", you just make
>>>>>         sure object align on cache line boundary.
>>>>>
>>>>>         Actually I was asked is my understanding of how @Contended supposed to
>>>>>         be implemented is right?
>>>>>
>>>>>>         Or are you saying you want an @Alignment annotation
>>>>>>         instead so it's more general? What other uses of custom alignment do you
>>>>>>         envision? Java is too high-level  and the underlying hardware/platform
>>>>>>         too
>>>>>>         abstracted away for a general purpose custom alignment hint, IMHO.
>>>>>         No, I do not want such ugly thing to happen with java! It's enough C
>>>>>         for such things...
>>>>>
>>>>>
>>>>>>         Sent from my phone
>>>>>>
>>>>>>         On Jan 17, 2012 10:56 AM, "Ruslan Cheremin"<cheremin at gmail.com>  <mailto:cheremin at gmail.com>  wrote:
>>>>>>>>         Yes. As a practical matter though, until an @Contended attribute
>>>>>>>>         or something like it is supported across JVMS (see list archives for
>>>>>>>>         discussion), you cannot arrange reliable two-sided padding
>>>>>>>>         for objects with mixed field types (ints, longs, refs that may be
>>>>>>>>         either 32 or 64 bits, etc), so one-sided is the best you can do.
>>>>>>>         By the way -- I was not thinking about @Contended as "make padding for
>>>>>>>         me". It seems for me like padding is only dirty hack, since nothing
>>>>>>>         better available. If I would control memory allocation (like JVM does)
>>>>>>>         I just can allocate @Contended objects on 64 (128... etc) bytes
>>>>>>>         boundary. I do not have to "pad" them -- nor both, nor one side. And I
>>>>>>>         suppose @Contended implementation to do exactly this -- "use special
>>>>>>>         allocator for objects of that type, which allocate them on cache line
>>>>>>>         boundary"
>>>>>>>
>>>>>>>         Am I wrong here?
>>>>>>>
>>>>>>>
>>>>>>>>         -Doug
>>>>>>>>         _______________________________________________
>>>>>>>>         Concurrency-interest mailing list
>>>>>>>>         Concurrency-interest at cs.oswego.edu  <mailto:Concurrency-interest at cs.oswego.edu>
>>>>>>>>         http://cs.oswego.edu/mailman/listinfo/concurrency-interest
>>>>>>>         _______________________________________________
>>>>>>>         Concurrency-interest mailing list
>>>>>>>         Concurrency-interest at cs.oswego.edu  <mailto:Concurrency-interest at cs.oswego.edu>
>>>>>>>         http://cs.oswego.edu/mailman/listinfo/concurrency-interest
>>>>         _______________________________________________
>>>>         Concurrency-interest mailing list
>>>>         Concurrency-interest at cs.oswego.edu  <mailto:Concurrency-interest at cs.oswego.edu>
>>>>         http://cs.oswego.edu/mailman/listinfo/concurrency-interest
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs.oswego.edu/pipermail/concurrency-interest/attachments/20120117/4bc17896/attachment-0001.html>


More information about the Concurrency-interest mailing list