A bunch of separate real concerns here: --------------- On the issue of STM Slowdown:
- Clojure is typically slower than Java by an interesting amount when doing Plain Olde Java Stuff, unless great care is taken (it's possible to take great care). Such a slowdown matters less than the ability to quickly produce functioning code in many cases. - STM'd C/C++ code is *typically* slower than normal code by 3x to 10x (depends on the STM) because of the need to track *all* variables. - Clojure does not need to track all variables, just REFs - so it's "STM slowdown" factor is probably 10x reduced from the typical C/C++ slowdown. This coupled with a typically lower base speed means Clojure XTNs don't slow down Clojure programs just for existing. --------------- On the issue of 99.9% retry: Clojure's STM suffers from being "obstruction free" and not "lock free" (I think; I haven't paid real close attention to the details). This means that conflicting XTN's can obstruct one another when making committing. This problem is typical of STM's, BTW. Basically, I wrote a tiny microbenchmark with endless conflicting transactions. As I add *real CPUs* (not just threads), total throughput *drops*. With a locking implementation, total throughput would drop some, then hit some lower bound and stay there. Note that this is *total* throughput, not just throughput-per-CPU or throughput- per-thread. It's a performance inversion problem: adding more CPUs hurts! Until this is fixed, multi-threaded performance on Clojure will remain fragile. The problem is generic to STMs - so really I'm saying that STMs, in general, have a more fragile performance profile than locks and this is one of the reasons that STMs have had problems getting traction in the Real World. Now, just to be a pessimist I'll point out that the java.util.concurrent Locks are *also* using obstruction-free algorithms and when an Azul box cuts loose with a few hundred threads there's *always* a CPU spare to "obstruct" somebody, so we hit live- lock real quick. Given 32-way and 64-way X86 boxes rapidly approaching I'm hopeful Doug Lea will finally have to do something about this. Cliff --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "Clojure" group. To post to this group, send email to clojure@googlegroups.com To unsubscribe from this group, send email to clojure+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/clojure?hl=en -~----------~----~----~----~------~----~------~--~---
