g77 compiler and mpif77 compiler in SCYLD

Edwards Arthur edwards at icantbelieveimdoingthis.com
Thu Feb 21 14:43:01 PST 2002


I'm using the older (-7) version of the free SCYLD and I have found that 
the mpif77 compiler gives about half the speed of g77 at the same level 
of optimization. Also, it behaves in a very unstable way when one uses 
individual optimization commands (-fcse-follow-jumps, for example). The 
calculation can simply blow up. At the same time, it seems to share some 
instability with the g77 compiler. We are using a Sandia DFT code, QUEST, 
and -O2 and -O3 lead to substantial differences in computed results. I 
have found that the optimization call -fcaller-saves is the culprit in 
QUEST and that its absence leads to no degradation in performance on a 
standard non-SCYLD linux box. Also using -O2 -fcaller-saves doesn't lead 
to really bizarre behavior. It even gives completely correct results, 
even though it is half as fast as a normal serial Linux box. 

So, I have the newer free version (from Cheap-bytes) and I'm thinking of 
upgrading. Is the new mpif77 compiler actually optimized better? Is the 
newer (-8) SCYLD substantially better in other ways? Are there any 
problems with the upgrade?

Art Edwards





More information about the Beowulf mailing list