(j3.2006) (SC22WG5.4112) [ukfortran] Standard intrinsics and coarrays
Thu Oct 22 08:35:46 EDT 2009
On Oct 22 2009, Bill Long wrote:
>> Systems with suitable compare-and-swap hardware can implement
>> parallel use of a shared RANDOM_NUMBER safely without locking. Adding
>> RANDOM_SEED to the mix makes that much trickier, and it's not something
>> that any sane programmer would do, anyway.
>As a (my) performance rule of thumb, if you can generate 1.e9 64-bit
>values per second you are OK. In some cases, 1.e8 might be considered.
>Even with hardware compare-and-swap, a shared generator on an even
>modest sized system will have trouble with this. The scheme has no
>chance of scaling to large numbers of images. And it still fails the
Oh, yes. It's a crazy approach for scalable HPC systems - no dissention
there. But I don't think that it should be forbidden, as it's a very
plausible approach for single-chip, non-HPC implementations. In those
cases, 1.0e7 values a second would be adequate, or even 1.0e6.
More information about the J3