Ben: I am sure your observations are certainly correct about differences between OSes and what JITTing does, but we have noticed that universally once a method gets big enough perf sucks. Even if it is a single platform which exhibits this behavior (since perhaps only one OS generates enough code to make it too big), then making it smaller will for sure solve have a positive affect on the platform in question, but also never seems to negatively affect other platforms very much. I am guessing that Java is just much better at mitigating perf problems with lots of small methods, but is stuck once methods get too big.
Charlie: I have always wanted to hack Jay to replace the switch with a class with method dispatch (static dispatch) to eliminate the switch body altogether, but the source code to Jay is ancient C code optimized for a PDP-11 :) -Tom On Sat, Jan 16, 2010 at 10:49 AM, Ben Evans <benjamin.john.ev...@googlemail.com> wrote: > On Sat, Jan 16, 2010 at 1:03 AM, Charles Oliver Nutter <head...@headius.com> > wrote: >> >> I have found an interesting clue to poor startup! >> >> It appears that DefaultRubyParser.yyparse (the main parser method) is >> *still* never being jitted by the JVM, in either client or server >> modes (at least on OS X Java, which should be pretty normal). > > Charlie - Beware. > > Different OSes can have *very* different profiles for both the set of > methods which are compiled and the amount of instructions each produce. > > My research in this area has unfortunately not produced any good rules of > thumb yet, except "It's complicated", even between Unixish OSes. > > For example, Solaris and Linux may have different sets of methods which are > compiled (and it is not simply the case that Linux compiles more than > Solaris does - there are methods which Solaris will go for which Linux won't > touch). > > To make matters worse, the amount of instructions generated for a method may > vary wildly. I have seen Solaris and Linux generating 5x # instructions for > a method (and in each direction). > > So, to do this kind of performance tuning cross-platform, I haven't found a > good way except to repeat runs across all OSes I care about. > > Having said all that, if anyone else on-list has done work in this area, and > has discovered any useful rules-of-thumb or other techniques for > cross-platform optimisation, please share - would love to hear about any > results here. > > Thanks, > > Ben > -- blog: http://blog.enebo.com twitter: tom_enebo mail: tom.en...@gmail.com --------------------------------------------------------------------- To unsubscribe from this list, please visit: http://xircles.codehaus.org/manage_email
