PSTSWM AlphaSC-667 Serial Performance

Performance Studies using

PSTSWM

Date/Person:	November 15, 2000 / P. Worley
Platform:	Compaq AlphaServer SC at Oak Ridge National Laboratory (colt.ccs.ornl.gov):
	16 ES40 4-way SMP nodes (667 MHz Alpha 21264a with 8MB L2 cache)
Environment:	Digital UNIX V5.0; RMS 2.36
	Compaq Fortran V5.3-1120
Code Version:	6.7.2
Make Options:	MACH=alpha-sc COMM=mpi PRECISION=8 PERF=n WORKSPACE=22000000
Compilation Options:	f90 -O4 -assume accuracy_sensitive -math_library accurate -align dcommons -align records -arch host -tune host
or	f90 -O4 -assume noaccuracy_sensitive -math_library fast -align dcommons -align records -arch host -tune host
or	f90 -O5 -assume noaccuracy_sensitive -math_library fast -align dcommons -align records -arch host -tune host
Number of steps:	T5, T10, T21, T42: 241 or 481
	T85: 49 or 97
	T170: 49 or 97
Notes:	using PSTSWM Fortran routines for Fourier transforms and BLAS

As with most compilers, the Digital Fortran compiler has a very large number of optimization options. While we have not experimented with these extensively, the initial indications are that the standard "safe" optimizations (labelled "-O4 accurate" here) do as well as the more agressive optimizations, at least for this code. For the full atmospheric model CCM/MP-2D, the use of the fast libraries has more impact, due to a heavy use of math intrinsics.
The performance is sensitive to the number of vertical levels (which is the second index). As the number of vertical levels increases, the performance is more dependent on bandwidth and latency to main memory. Particular numbers of vertical levels, for example 16 at horizontal resolution T42, also have poorer performance than those higher or lower, probably due to array alignment and cache-line collision issues. See the graphical summary for more details.
These results indicate a DECREASE in performance since the data collected in April and May, 2000, using an earlier version of the compiler.