I've posted a patch with some changes here (https://gist.github.com/4416803), it includes the record change here and a small change to interpret-instruction, the benchmark runs > 2x the default as it did for Marshall. The patch also modifies the main loop to use a thread pool instead of agents and allows you to set the number of threads, this might help diagnosing the parallel performance issue.
On the modified benchmark I'm seeing ~4x speedup with the parallel version on an 8 core machine and the profiler reports that the parallel version is using twice as much cpu time. I also had another look at the native calls issue & modified the clojure runtime to avoid most to the calls the profiler said were taking significantly more time in the parallel version, it did speed things up but only by ~6%, not the large margin the profiling results had led me to believe were possible, it looks like the profiler overstates these methods times. The modified clojure 1.5 is available here https://github.com/cdorrat/clojure/commit/dfb5f99eb5d0a45165978e079284bab1f25bd79f if anyone's interested YourKit is reporting that a number of clojure.core functions are taking longer in the parallel version than the serial and they all seem to be methods that have one or more instanceof or instance? calls but given the results above I'm not sure how much weight to give this. It's seems the elephant is still in the room and responsible to ~50% of the cpu time :) Cameron. On Saturday, December 22, 2012 10:57:28 AM UTC+11, Marshall Bockrath-Vandegrift wrote: > > Lee Spector <lspe...@hampshire.edu <javascript:>> writes: > > > FWIW I used records for push-states at one point but did not observe a > > speedup and it required much messier code, so I reverted to > > struct-maps. But maybe I wasn't doing the right timings. I'm curious > > about how you changed to records without the messiness. I'll include > > below my sig the way that I had to do it... maybe you can show me what > > you did instead. > > I just double-checked, and I definitely see a >2x speedup on Josiah’s > benchmark. That may still be synthetic, of course. Here’s what I did: > > (eval `(defrecord ~'PushState [~'trace ~@(map (comp symbol name) > push-types)])) > > (let [empty-state (map->PushState {})] > (defn make-push-state > "Returns an empty push state." > [] empty-state)) > > > Still, I guess the gorilla in the room, which is eating the multicore > > performance, hasn't yet been found. > > No, not yet... I’ve become obsessed with figuring it out though, so > still slogging at it. > > -Marshall > > -- 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 Note that posts from new members are moderated - please be patient with your first post. 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
