Running the tests with diagnostic options enabled, I see the following:
Running
org.apache.commons.math4.analysis.integration.gauss.HermiteParametricTest
Compiled method (c2) 25704 794
org.apache.commons.math4.util.FastMath::pow (635 bytes)
total in heap [0xf37f3ac8,0xf37f4bc0] = 4344
reloc
I wonder if in this case the problem is related to branch prediction or
instruction re-ordering.
The FastMath.pow() method has a lot of ifs to handle special cases, but
they are all independent and and handled in sequence. No results are
calculated or stored in them, so the JVM might wrongly assum
Might be worth running a version of the test with JIT disabled.
This will make the test much longer so I suggest first seeing if the
tests fail on their own, and then disable JIT.
You could experiment with the additional Jenkins Math job I set up for
the H10 tests:
https://builds.apache.org/job/Co
On 2/24/15 12:56 PM, Thomas Neidhart wrote:
Interesting. I could not get this to happen on
OSX:
java version "1.7.0_71"
Java(TM) SE Runtime Environment (build 1.7.0_71-b14)
Java HotSpot(TM) 64-Bit Server VM (build 24.71-b01, mixed mode)
java version "1.8.0_11"
Java(TM) SE Runtime Environment (b
Hi,
after switching to Java 1.7 there seems to be another problem with
FastMath, see the two test failures below when running on jenkins.
Can somebody reproduce it on a 64 bit machine (I just have a 32 bit
laptop at home)?
Test setup: Java 7 latest (1.7.0_75?), maven 3.1.1.
Tests run: 65, Failu