Hi,
I have been successfully testing JamVM + GNU Classpath 0.98 on ARM a couple of weeks now. However, it seems that JamVM suffers significant performance problems in my test board.
For example running Scimark2 gives the following results:
Composite Score: 0.754515649534562
FFT (1024): 0.43564193417098535
SOR (100x100): 1.3841795476602183
Monte Carlo : 0.2546323439119081
Sparse matmult (N=1000, nz=5000): 0.7180925550940853
LU (100x100): 0.9800318668356128
In contrast to proprietary VMs such as Sun CDC HotSpot and HotSpot client VM, the performance is significantly slower. HotSpot client VM had composite score of 20 with JIT and 2,46 when running interpreter mode.
I have used both direct-threaded interpreter with stack-caching and inline-threaded interpreter with stack-caching. Currently, I don't have a cross compiler with hard-float support so I'm only compiling JamVM with soft-float. Of course hard-float could improve the performance in numerical computing, but still, the performance is also quite bad when running for example a simple OSGi core implementation such as equinox. Startup takes approximately 1 minute.
The board has a ARM1136JF-S core.
Is this just an ARM issue with JamVM or am I having some other problems for example with JamVM features?
Very best regards,
Ronem
