Re: JavaMemoryModel: a naive proposal

From: Assaf Schuster (assaf@cs.technion.ac.il)
Date: Thu Jul 12 2001 - 17:26:45 EDT


>
> Plus, SC would substantially hurt performance of JVM's on many
> uniprocessors, and cripple it on SMP's. I don't think our JSR could
> ever be approved if it crippled the performance of Java.
>

I recall a paper by Marc Hill proposing SC as a hardware memory model.
He basically points out that there are many optimizations that can help SC
gain performance.

My point in the prev mail was that a release/acquire model will probably
drive people to over-synchronizing, and redundant memory barriers will make
the program
far slower than SC (as every Software DSM person can testify).

My other point was that nothing in this area is known *for sure* in a
quantitative way,
not even performance relations.

------

As for a different point: I mentioned the liveness of SC (or in general
the liveness of non-R/A models). This liveness property helps real-time
performance-bounded
programmers in avoiding memory barriers by including races in the program.
However, liveness can also be incorporated as a property into R/A models
(you are guaranteed that even in the absense of a R/A pair, value will
sometimes be
transfered).

Of course, such liveness can cause lots of strange effects if causality is
not maintained,
(as is allowed in Bill's model) but maintaining causality is known to hurt
performance.

--- Assaf

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



This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:00:33 EDT