JavaMemoryModel: The range of memory models now in play

From: Bill Pugh (pugh@cs.umd.edu)
Date: Thu Aug 08 2002 - 16:11:16 EDT


Despite all the recent talk, I think are actually moving together on
many fronts.

For the moment, lets ignore the issue of a total order over volatile
memory actions, and just assume that volatile memory is sequentially
consistent.

Consider two programming models, A and B.

In both models, you check the validity of a totally ordered execution trace E.

For each thread T, the actions of T in E must be consistent with the
single-thread semantics of the program (ignoring the values returned
by reads, which are determined by the memory model).

A happens-before relation is defined on actions in E, defined by
order within a thread, and by matching unlock/lock and volatile
write/read pairs.

In both models, the values read in E are consistent with hb; i.e.,
if R is a read that returns the value of a write W, then we cannot
have R hb
W and we cannot have a W' to the same location where W hb W' hb R in E.

In model B only, a read must see the value of a previous write.

I think all of the models now being discussed fall somewhere between A and B.

A is too relaxed, because it allows causal loops and allows for
correctly synchronized programs to exhibit non-sequentially
consistent behavior.

B is too strict, because it does allow reordering of a read and a
following write.

Jason and Sarita: do you agree that your models fit between A and B?
Jason, I know that your model isn't specified this way, but I'm
pretty sure the semantics of your model fall between A and B.

If I'm correct, this will help us focus on the remaining differences
between our models.

        Bill
-------------------------------
JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel



This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:00:41 EDT