Parallel performance of the XL Fortran random{_}number intrinsic function on Seaborg

Gerber, Richard A.

Lawrence Berkeley National Laboratory

关键词: Configuration;    Performance;    Fortran;    99 General And Miscellaneous//Mathematics, Computing, And Information Science;   

DOI  :  10.2172/817243
RP-ID  :  LBNL--53531
RP-ID  :  AC03-76SF00098
RP-ID  :  817243

美国|英语

来源: UNT Digital Library

【 摘 要 】

The Fortran intrinsic function random{_}number is shown to perform very poorly when simultaneously called from 16 tasks per node on NERSC's IBM SP Seaborg in its default runtime configuration. Setting the runtime option intrinthds=16 improves runtime performance significantly and gives good results for all possible numbers of tasks per node. It is speculated that the cause of the problem is the creation of an excessive number of threads in the default configuration. It is noted that these threads appear to be created by default, without specifying a ''thread-safe'' compiler or other user interaction.