[E-Lang] Reflection invoke() timing measurement
Wed, 18 Jul 2001 22:25:36 -0700 (PDT)
>Once we have an E-to-JVM-bytecode compiler and have picked various
>performance low hanging fruit, we may very well find reflective invoke() to
>be our bottleneck. Reflective invoke() is fundamental to mapping E
>computation onto the JVM. Fortunately, Sun says their JDK 1.4
>implementation does reflective invoke() much faster, but I haven't seen a
>quantification of "much faster". If someone would care to try Michael's
>code on a JDK 1.3 and 1.4 on the same hardware and report the ratio, I'd
>sure appreciate it.
Your wish is my command:
Processor JDK 1.3.0_02 JDK 1.4.0-beta-b65
--------- ------------ ------------------
350Mhz SPARC (Solaris 5.7) 19.7 10.3
600Mhz Pentium III (Red Hat 7.0) 7.0 3.5
Times in microseconds averaged over about 20 runs of each.
Looks to me like a significant improvement alright.
(Since I've been performance tuning some compute-intensive Java code
lately, including some somewhat invoke-intensive stuff, I was pretty
motivated to try this.)