On Apr 16, 2010, at 10:24 AM, David Dabbs wrote:
Hello. My apologies if these questions are OT for this list.
On the premise that building my own binaries could yield performance
improvements versus the Sun(Oracle)-provided binaries I've decided to
tinker with building from source.
* My first question is "Is this a reasonable premise?"
Possibly. Might be worth investigating.
* If so, is it possible to build using Intel's compiler?
I have no idea. Certainly not without changing makefiles to some degree,
unless it completely 100% mimics a Visual Studio 2003 compiler.
* Now that Visual Studio 2010 has been released will there be an
option to use it to build the JDK under Windows?
Our plans have always been to advance OpenJDK7 to use a newer Visual
Studio
compiler, 2010 would be a good choice, but it will take time to check it
out and get it all working. Stay tuned.
* If I build on a recent Xeon with "gcc -march=native ..." would one
expect
to see performance differences versus the Sun(Oracle)-provided
binary to
justify the effort?
Not sure. This probably would have no impact on the code generated by
the VM
at runtime, but it might improve the speed of GC code or maybe some of
the
JNI libraries. Assuming the optimizations don't break something
somewhere.
So it kind of depends on what the benchmark you would use does.
* I have recently been tinkering with Intel's memory allocator and
TCMalloc,
etc.
Would compiling with these be expected to a) even work and b) yield
perf
improvements worth the effort.
Completely unknown to me.
So often the raw performance comes down to the VM generated code
-kto
Thank you,
David Dabbs
