http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54406
Bug #: 54406
Summary: random_number() sometimes returns duplicate values
when called from parallel threads
Classification: Unclassified
Product: gcc
Version: 4.7.1
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: fortran
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: bugzilla.tms...@wronghead.com
When called from an OMP parallel section (e.g. $OMP parallel do), the
random_number() intrinsic occasionally returns the same value twice in
succession.
I've attached a simple test program, to be compiled with:
gfortran -fopenmp rnd.f08
Some quick statistics showed that the effect is less when:
- directly using random_number() in the main loop,
- using a static schedule, and/or
- omitting the (useless) inner loop.
The effect vanishes when using 'print*,fct()' directly, i.e. omitting storage
or the return value in variable r.
For BASH users, there's a simple and obvious script to get some statistics on
the number of duplicated values:
while true; do echo "10-`./a.out|uniq|wc -l`" | bc; done
I quote from "Using GNU Fortran":
"Please note, this RNG is thread safe if used within OpenMP directives, i.e.,
its state will be consistent while called from multiple threads. However, the
KISS generator does not create random numbers in parallel from multiple
sources, but in sequence from a single source. If an OpenMP-enabled application
heavily relies on random numbers, one should consider employing a dedicated
parallel random number generator instead."
While I understand that the period is finite, it says that "The overall period
exceeds 2^123", i.e. ca. 1.06E37.
Thus, to my understanding, duplication of values should never happen even when
considering all the caveats mentioned.
