[concurrency-interest] Enforcing total sync order on modern hardware

Marko Topolnik marko at hazelcast.com
Tue Mar 24 15:30:41 EDT 2015


On Tue, Mar 24, 2015 at 7:15 PM, Alexander Terekhov <TEREKHOV at de.ibm.com>
wrote:

> again, IRIW is about write atomicity... minimal scenario for that needs
> only three threads:
>
> T1: X = 1;
> T2: if (X) Y = 1;
> T3: A = Y; B = X;
>
> A == 1 && B == 0 is possible without write atomicity.
>

But this dispenses with the essential feature of IRIW, which is the
independency of two writes so that any other thread can receive them in
arbitrary order without breaching causality. In your example you have
second write dependent on the first, introducing a clear happens-before
edge between X = 1 and Y = 1. Put another way, a system may pass this test
and still fail IRIW. Therefore they cannot be considered equivalent.

---
Marko


> Marko Topolnik <marko at hazelcast.com> on 24.03.2015 18:44:16
>
> To:     Oleksandr Otenko <oleksandr.otenko at oracle.com>
> cc:     Alexander Terekhov/Germany/IBM at IBMDE, concurrency-interest
>        <Concurrency-interest at cs.oswego.edu>
> Subject:        Re: [concurrency-interest] Enforcing total sync order on
> modern
>        hardware
>
>
> On Tue, Mar 24, 2015 at 3:32 PM, Oleksandr Otenko <
> oleksandr.otenko at oracle.com> wrote:
>       The outcome r1==0, r3==0 is disallowed through the same reasoning in
>       Dekker's idiom as in IRIW: since r1==0, y=1 must not happen before
>       r1=y; then r3=x, being after y=1 in program order, must also not
>       precede r1=y, and transitively cannot precede x=1. You need this step
>       that from "Y must not precede X" follows that "X happens-before Y" -
>       but that's what makes the order total.
>
>       In IRIW you have the same chain of reasoning. Even if you get rid of
>       synchronization order, but keep synchronizes-with and transitive
>       closure, then from observing r0==1 and r2==1 you still have the same
>       edges. The key here still is the conclusion that since r1==0, y=1
>       must not happen before r1=y, which is the only part that requires
>       total ordering - cannot be partially ordered. But it is exactly the
>       same bit that cannot be partially ordered in Dekker's idiom: r1=y
>       happens-before y=1 because otherwise r1=y should observe 1. If you
>       don't have mutual exclusion in ordering r1=y and y=1 in Dekker's
>       idiom, why would the outcome r1==0, r3==0 be forbidden?
>
> This the IRIW (1,0,1,0) result in a happens-before diagram:
>
> T4: x=0 --> y=0 ----> r2=y --> r3=x
>                    /--^
> T3: x=0 --> y=0 --/-----> r0=x --> r1=y
>                  /    /---^
> T2: y=1 --------/    /
>                     /
> T1: x=1 -----------/
>
> (note that I made the setting of initial values in T3 and T4 explicit)
>
> As you can witness, there are no violations of the union of
> synchronizes-with and program order ("transitive closure", even though that
> term actually applies to something else). Therefore IRIW only violates
> sequential consistency and does not violate what I assume under the term
> "happens-before consistency".
>
> BTW as for your earlier observation that IRIW is not minimal because Dekker
> is smaller, note that the "issue of interest" for IRIW is the sequentially
> inconsistent observation of mutually independent stores, as observed by
> mutually independent loads. Here "mutually independent" means "done by
> separate, independent threads". Clearly, IRIW is exactly the minimal
> scenario for that.
>
> ---
> Marko
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs.oswego.edu/pipermail/concurrency-interest/attachments/20150324/dd874f2d/attachment.html>


More information about the Concurrency-interest mailing list