Neil Davies
You may not be asking the right question here. Your final system's performance is going to be influenced far more by your algorithm for updating than by STM (or any other concurrency primitive's) performance.
I may not be asking the right question, but I am happy about the answer, including yours :-) I think STM is semantically the right tool for the (my) job, but for it to be useful, the implementation must not be the limiting factor. I.e running on n+1 CPUs should be measurably faster than running on n, at least up to n=8, and preferably more. With the assuming I can get enough parallelism and avoiding too many rollbacks, of course.
Others have already mentioned the granularity of locking - but that one of the performance design decisions that you need to quantify.
Yes. Fine grained - I'm thinking a large Array of TVars. (If you only have a single TVar, it might as well be an MVar, no?) -k -- If I haven't seen further, it is by standing in the footprints of giants