plugwash wrote:I've recently uploaded 7u75-2.5.4-3+rpi1 which has the new jit to.raspbian jessie. I expect 2.5.5 to come through to both wheezy and jessie in-time but i'm currently watching what happens in Debian.
Anyone know how the new jit compares to oracle's propietary one?
There is something missing in the 7u75-2.5.4-3+rpi1 package it looks like the new jit is included in the sourcetree but disabled at runtime. I will have to examine the debian package to see if there is a missing build flag.
I tried to benchmark the new jit on the raspberry pi 1 and compared it with edward nevills reference build where the new jit work
http://openjdk.linaro.org/arm32jit/
I used SciMark 2.0a
http://math.nist.gov/scimark2/download_java.html
7u75-2.5.4-3+rpi raspbian
Composite Score:
2.773446704209826
FFT (1024): 1.6724202743674736
SOR (100x100): 6.285735331895986
Monte Carlo : 0.6632359390308941
Sparse matmult (N=1000, nz=5000): 2.4615385518271573
LU (100x100): 2.78430342392762
Edwards arm32 microJIT
Composite Score:
9.7088881566708
FFT (1024): 6.936563765379863
SOR (100x100): 17.71098303263207
Monte Carlo : 3.3608204434787985
Sparse matmult (N=1000, nz=5000): 9.579045785957213
LU (100x100): 10.957027755906072
oracle jdk8 b132
Composite Score:
23.92186747096536
FFT (1024): 21.518957822414773
SOR (100x100): 48.46466057515884
Monte Carlo : 8.494792808844737
Sparse matmult (N=1000, nz=5000): 20.277227914236352
LU (100x100): 20.8536982341721
The new OpenJDK JIT is 3 times faster compared to the old jit on a raspberry Pi 1
Oracles properitary JIT version is still about 2 times faster compared to the new jit in OpenJDK.