JRockit as fast or faster than Sun

Question

I've been very impressed by my tests with JRockit. For the code that I've tried
so far, I've found it to generally be at least as fast as Sun's latest 1.4.1 JVM
and in some cases substantially faster. Great work.
From profiling the execution under VTune, I have a few observations that perhaps
somebody at Bea might like to comment on. It appears that JRockit does very little
inlining compared to Sun's JVM. I found simple routines like javax.vecmath.Tuple3d.set(Tuple3d)
were not being inlined. (This routine simply copies three doubles from one object
to another.) Also access to direct buffers (from ByteBuffer.allocDirect()) seems
slow compared to Sun. In my case I was reading and writing doubles using a DoubleBuffer
(with native byte ordering). And lastly, it sometimes seems to unnecessarily
copy arguments from one place on the stack to another instead of using them directly
in place. This is made worse by copying doubles using two move dword instructions
which can cause store-forwarding stalls.
But I'm not complaining. The performance is state of the art for Java. And I
fully expect it to continue to get faster.
Bruce Walter
Research Associate, Cornell University

Answer 1

Bruce,
Always good to hear that JRockit is faster. And you are right - we will
continue to improve performance.
JRockit of course has other heuristics for determining inlining candidates
than other JVMs. In some cases we will inline more aggressively and in other
cases less so. I am wondering a little about how you came to your
conclusions. Since you were running VTune I assume that you were using JVMPI
to inspect the Java code. With JVMPI turned on JRockit will behave somewhat
different than it normally will due to the overhead required to support
JVMPI, thus information gathered this way may not always be correct.
Regards,
/Staffan Larsen, BEA
"Bruce Walter" <[email protected]> skrev i meddelandet
news:[email protected]..
>
I've been very impressed by my tests with JRockit. For the code that I'vetried
so far, I've found it to generally be at least as fast as Sun's latest1.4.1 JVM
and in some cases substantially faster. Great work.
From profiling the execution under VTune, I have a few observations thatperhaps
somebody at Bea might like to comment on. It appears that JRockit doesvery little
inlining compared to Sun's JVM. I found simple routines likejavax.vecmath.Tuple3d.set(Tuple3d)
were not being inlined. (This routine simply copies three doubles fromone object
to another.) Also access to direct buffers (fromByteBuffer.allocDirect()) seems
slow compared to Sun. In my case I was reading and writing doubles usinga DoubleBuffer
(with native byte ordering). And lastly, it sometimes seems tounnecessarily
copy arguments from one place on the stack to another instead of usingthem directly
in place. This is made worse by copying doubles using two move dwordinstructions
which can cause store-forwarding stalls.
But I'm not complaining. The performance is state of the art for Java.And I
fully expect it to continue to get faster.
Bruce Walter
Research Associate, Cornell University