A
Adam Preble
I've been dealing with some microbenchmarks off the Internet for Java
profiling at work. I've concluded that all we've seen so far are
obsolete with today's JVMs. The biggest problem is incosideration of
dynamic optimization. Some JVM implements are starting to recompile
their methods as they get taxed, and using the additional, realtime
profiling information for optimizations.
The two big ones I've been studying are Java Grande and Shootout. Note
that these are suites of much smaller tests. They're preferable because
their code generation can be easily dumped and studied. I understand
that to really get a thorough assessment, some more rigorous stuff like
SpecJBB need to be tried.
It's tempting for me to publish my own suggestions, but I wanted to make
sure there's a need for it yet. As it stands, tests from the 2001 time
frame are becoming obsolete. They don't "train" the JVM and in some
cases the code is so trivial that it gets removed entirely.
profiling at work. I've concluded that all we've seen so far are
obsolete with today's JVMs. The biggest problem is incosideration of
dynamic optimization. Some JVM implements are starting to recompile
their methods as they get taxed, and using the additional, realtime
profiling information for optimizations.
The two big ones I've been studying are Java Grande and Shootout. Note
that these are suites of much smaller tests. They're preferable because
their code generation can be easily dumped and studied. I understand
that to really get a thorough assessment, some more rigorous stuff like
SpecJBB need to be tried.
It's tempting for me to publish my own suggestions, but I wanted to make
sure there's a need for it yet. As it stands, tests from the 2001 time
frame are becoming obsolete. They don't "train" the JVM and in some
cases the code is so trivial that it gets removed entirely.