[concurrency-interest] Relativity of guarantees provided by volatile

Zhong Yu zhong.j.yu at gmail.com
Sat Aug 18 16:38:07 EDT 2012


On Fri, Aug 17, 2012 at 8:21 PM, Boehm, Hans <hans.boehm at hp.com> wrote:
> My reading is that “subsequent” in chapter 17 of the spec refers to
> synchronization order, except in one case in which it refers to a
> hypothetical sequentially consistent memory model.
>
>
>
> I agree that the originally described behavior is allowed.   There’s another
> complication here that can explain this behavior (as can delayed stores).
> AFAIK, the spec does not clearly require that  “wall clock time” is
> perfectly synchronized across threads, and implementations may, for good
> reason, fail to ensure that, particularly in this case.  We had this
> discussion in connection with C++11, and that spec does require that a
> specific clock is consistent with happens-before.  But even that spec, which
> is strict enough to raise some controversy, isn’t sufficient here, since
> there is no happens-before relation between the threads.

It seems imperative that, on any physical model of JMM, among
synchronization actions, if a happens-before b, then there's a causal
relationship from a to b, and it's necessary that
a.startTime<b.endTime (on any observer's clock). In that sense, the
term happens-before is literal after all?

In OP's example, there are so(r_i, w_j), but there's no hb(r_i, w_j),
therefore there's no temporal constraints, w_j can occur before r_i.
If there were something that established hb(r_i, w_j), then the
example is impossible under JMM.

Interestingly, JMM also defines r_i and w_j to be in a data race, even
though they are all volatile.

Could it be true that, if an execution is free of data race, it is
temporally consistent, i.e. its synchronization order is consistent
with temporal order.


>
>
>
> Hans
>
>
>
> From: concurrency-interest-bounces at cs.oswego.edu
> [mailto:concurrency-interest-bounces at cs.oswego.edu] On Behalf Of David
> Holmes
> Sent: Friday, August 17, 2012 5:03 PM
> To: Yuval Shavit; Vitaly Davidovich
> Cc: concurrency-interest at cs.oswego.edu
>
>
> Subject: Re: [concurrency-interest] Relativity of guarantees provided by
> volatile
>
>
>
> Ah I knew there was something about "immediate" in there somewhere :) So if
> instructions were immediate (ie they completed as soon as they started) then
> wall clock observations would be consistent with actual executions. But on
> real systems instructions are not immediate so you have to take the
> completion of the instruction as the point at which to make a wall clock
> observation.
>
>
>
> Aside: as well as being atomic and immediate you also need to preclude two
> instructions from executing simultaneously :)
>
>
>
> David (now departing to enjoy the weekend :) )
>
> -----Original Message-----
> From: Yuval Shavit [mailto:yshavit at akiban.com]
> Sent: Saturday, 18 August 2012 9:49 AM
> To: Vitaly Davidovich
> Cc: dholmes at ieee.org; concurrency-interest at cs.oswego.edu
> Subject: Re: [concurrency-interest] Relativity of guarantees provided by
> volatile
>
> (Resending -- I'd accidentally sent this only to Nathan)
>
> So, Nathan, you're essentially saying that "subsequent" is defined in terms
> of clock time. If we take that as the definition, then I agree that it works
> as we'd all expect. Furthermore, I agree that that's the intuitive
> definition, as well as the de-facto one. But it doesn't seem to be the one
> explicitly defined by the JLS. (And if it is the one implicitly defined by
> the JLS, I there's a serious bug waiting for any distributed-node JVM whose
> nodes are traveling away from each other at relativistic speeds! ... though
> I'm no physics expert.)
>
>
>
> That said, 17.4.3 does imply that the reads will be viewable in a
> wall-clock-sequential way, albeit informally
>
>
>
>     Sequential consistency is a very strong guarantee that is made about
> visibility and ordering in an execution of a program. Within a sequentially
> consistent execution, there is a total order over all individual actions
> (such as reads and writes) which is consistent with the order of the
> program, and each individual action is atomic and is immediately visible to
> every thread.
>
>
>
> (emphasis on "is immediately visible")
>
>
>
> On Fri, Aug 17, 2012 at 7:41 PM, Vitaly Davidovich <vitalyd at gmail.com>
> wrote:
>
> David and you may be right in the theoretical aspect.  In practice, I can't
> fathom how a JVM can do this type of analysis.  That's an issue that I have
> with JMM's wording of happens-before -- it doesn't translate to reality, it
> seems like.
>
> Sent from my phone
>
> On Aug 17, 2012 7:30 PM, "Yuval Shavit" <yshavit at akiban.com> wrote:
>
> Sure, but it could decide that the execution order is [w1, w2, w3, r]. In
> fact, as far as we know, the thread may have been scheduled such that that
> was the clock-time ordering, too.
>
> On Fri, Aug 17, 2012 at 7:21 PM, Vitaly Davidovich <vitalyd at gmail.com>
> wrote:
>
> Two volatile writes emit a store-store barrier in between, which to me means
> they cannot be collapsed and must be made visible in that order (on non-TSO
> this would require a h/w fence).  In other words, I don't think compiler can
> remove the redundant stores as if this was a non-volatile field, where it's
> a perfectly valid (and good) optimization.
>
> Sent from my phone
>
> On Aug 17, 2012 7:18 PM, "David Holmes" <davidcholmes at aapt.net.au> wrote:
>
> How does it violate the JMM? There is nothing to establish that any read has
> to have occurred prior to w=3. An external observer may say "hey if we'd
> actually written w=1 at this point then the read would see 1" but that is
> irrelevant. The program can not tell the other writes did not occur.
>
>
>
> David
>
> -----Original Message-----
> From: Vitaly Davidovich [mailto:vitalyd at gmail.com]
> Sent: Saturday, 18 August 2012 9:12 AM
> To: dholmes at ieee.org
> Cc: Marko Topolnik; concurrency-interest at cs.oswego.edu
> Subject: Re: [concurrency-interest] Relativity of guarantees provided by
> volatile
>
> I don't think the writes to w can be reduced to just the last one as it
> would violate the JMM.  R may only see the last one due to interleaving
> though. Not sure if that's what you meant.
>
> Sent from my phone
>
> On Aug 17, 2012 7:03 PM, "David Holmes" <davidcholmes at aapt.net.au> wrote:
>
> Hi Marko,
>
> I think the "surprise" is only in the way you formulated this. Said another
> way a write takes a finite amount of time from when the instruction starts
> to execute to when the store is actually available for a read to see.
> (Similarly a read takes a finite amount of time.) So depending on those two
> times a read and write that happen "around the same time" may appear to have
> occurred in either order. But when you program with threads you never know
> the relative interleavings (or should never assume) so it makes no
> difference how the program perceives the order compared to how some external
> observer might perceive it.
>
> As for your optimization to "chunk" volatile writes, I don't see a problem
> here if you are basically asking if given:
>
> w = 1;  // w is volatile
> w = 2;
> w = 3;
>
> that this could be reduced to the last write alone? I see no reason why not.
> Without some additional coordination between a reader thread and the writer
> thread, reading w==3 is a legitimate outcome. If you are thinking about how
> the hardware might chunk things then that is a different matter. We have to
> use the hardware in a way that complies with the memory model - if the
> hardware can't comply then you can't run Java on it.
>
> David Holmes
> ------------
>
>> -----Original Message-----
>> From: concurrency-interest-bounces at cs.oswego.edu
>> [mailto:concurrency-interest-bounces at cs.oswego.edu]On Behalf Of Marko
>> Topolnik
>> Sent: Saturday, 18 August 2012 7:24 AM
>> To: concurrency-interest at cs.oswego.edu
>> Subject: [concurrency-interest] Relativity of guarantees provided by
>> volatile
>>
>>
>> Consider the following synchronization order of a program
>> execution involving a total of two threads, R and W:
>>
>> - thread R begins;
>>
>> - thread R reads a volatile int sharedVar several times. Each
>> time it reads the value 0;
>>
>> - thread R completes;
>>
>> - thread W begins;
>>
>> - thread W writes the sharedVar several times. Each time it
>> writes the value 1;
>>
>> - thread W completes.
>>
>> Now consider the wall-clock timing of the events:
>>
>> - thread R reads 0 at t = {1, 4, 7, 10};
>> - thread W writes 1 at t = {0, 3, 6, 9}.
>>
>> As far as the Java Memory Model is concerned, there is no
>> contradiction between the synchronization order and the
>> wall-clock times, as the JMM is wall-clock agnostic. However, I
>> have yet to meet a single Java professional who wouldn't at least
>> be very surprised to hear that the specification allows this.
>>
>> I understand that the SMP architecture that dominates the world
>> of computing today practically never takes these liberties and
>> makes the volatile writes visible almost instantaneously. This
>> may change at any time, however, especially with the advent of
>> massively parrallel architectures that seem to be the future. For
>> example, an optimization technique may choose to chunk many
>> volatile writes together and make them visible in a single bulk
>> operation. This can be safely done as long as there are no
>> intervening read-y actions (targets of the synchronizes-with
>> edges as defined by JLS/JSE7 17.4.4).
>>
>> Now, my questions are:
>>
>> 1. Is there a loophole in my reasoning?
>>
>> 2. If there is no loophole, is there anything to worry about,
>> given that practically 100% developers out there consider as
>> guaranteed something that isn't?
>>
>>
>> -Marko
>>
>>
>>
>>
>> _______________________________________________
>> Concurrency-interest mailing list
>> Concurrency-interest at cs.oswego.edu
>> http://cs.oswego.edu/mailman/listinfo/concurrency-interest
>>
>
> _______________________________________________
> Concurrency-interest mailing list
> Concurrency-interest at cs.oswego.edu
> http://cs.oswego.edu/mailman/listinfo/concurrency-interest
>
>
> _______________________________________________
> Concurrency-interest mailing list
> Concurrency-interest at cs.oswego.edu
> http://cs.oswego.edu/mailman/listinfo/concurrency-interest
>
>
>
>
>
>
> _______________________________________________
> Concurrency-interest mailing list
> Concurrency-interest at cs.oswego.edu
> http://cs.oswego.edu/mailman/listinfo/concurrency-interest
>



More information about the Concurrency-interest mailing list