lstagner
commented
4 years ago From what I have read support for random_init was introduced in gfortran 9. Because of this, at this time, we cannot implement this feature since most fusion clusters are hopelessly out of date. Your suggestion is a good one though and when gfortran 9 becomes common enough we will change it.
There has been a push for the use of Singularity containers in the future so this fix may happen sooner rather than later.
Hello, I'd propose to (conditionally) replace direct use of /dev/urandom with RANDOM_INIT/RANDOM_NUMBER builtins at least in case of GNU fortran, because it is more standard and portable method and internally employs /dev/urandom if available, with some fallbacks and alternative OS-specific mechanisms of obtaining random values. Please note that libgfortran has been using /dev/urandom internally since SVN r239356 (2016-08-11), whereas RANDOM_INIT was introduced in 2018, therefore the patch proposed below has no way to silently pick up some ancient dumb impementation (It would then fail to compile instead).
--- src/utilities.f90.orig Mon Jan 13 21:57:14 2020 +++ src/utilities.f90 Sun Jul 12 12:48:08 2020 @@ -214,9 +214,16 @@ integer(Int32) :: seed !+ Seed value +#ifdef __GFORTRAN__ + real(8) :: temp_r + call random_init(.false., .true.) + call random_number(temp_r) + seed = FLOOR(2147483647*temp_r) +#else open(89, file="/dev/urandom", access="stream", form="unformatted", action="read", status="old") read(89) seed close(89) +#endif seed = abs(seed) end function rng_seed