On Fri, 15 Feb 2002, Etienne M. Gagnon wrote: > Have you heard of "thin locks" and other VM level optimizations that reduces > "significantly" the cost of locks?
Even the ideal lock still carries a large cost. The CPU will probably stall, flush its store buffer, reload values from memory, etc. There's no escaping that. (However I'm not sure if this is true when reentering a lock already held by the current thread.) > The "synchronize" will force the "current processor" to update its data > cache. Most modern multiprocessors use hardware cache coherency (though the Java memory model doesn't require this, and relying on it isn't a good idea). More importantly, synchronization forces ordering of loads and stores. My impression is that hardware cache coherency is a bad idea because it gives programmers a false sense of security and may even limit the scalability of SMP hardware due to the extra bus traffic required. (I wonder if anyone's tried Java on a NUMA architecture?) > Mutiprocessors + threads => difficult programming. Most definitely. Jeff _______________________________________________ Classpath mailing list [EMAIL PROTECTED] http://mail.gnu.org/mailman/listinfo/classpath
