ORAC PERFORMANCE TESTS ON IBM SP6

VECTORIZED AND NON-VECTORIZED CODE; VECTORIZATION COMPILER OPTIONS

1. IN ALL TESTS:

ORAC 5.1.1
CPU is the execution time (s) written by the program
machines:
- cluster linux is clxx.sm.chim.unifi.it, xx=11..18
- SP6 CINECA is sp.sp6.cineca.it
default compiler and options:
- Linux: ifort -O3
- SP6 : mpxlf90 -O3 -qnosave -qrecur -qstrict -qfixed -qarch=pwr6 -qtune=pwr6 -qcache=auto

MACHINE	MOD	CPU	date(2011)	vector code in	compiler options
Cluster Linux		153	13/9	-	ifort -O3
SP6 CINECA	ST	180	<9/9	-
SP6 CINECA		150	13/9	-	-qnostrict -qhot=vector
SP6 CINECA		128	13/9	mts_forpp.f
SP6 CINECA		130	13/9	mts_forpp.f	-qnostrict -qhot=vector

MACHINE	MOD	CPU	date(2011)	vector code in	compiler options	NOTES
SP6 CINECA	SMT	153	21/9
SP6 CINECA	SMT	191	21/9	mts_forpp.f

MACHINE	MOD	CPU	date(2011)	vector code in	compiler options	NOTES
Cluster Linux		131
SP6 CINECA	SMT	215	<9/9
SP6 CINECA	ST	165	<9/9
SP6 CINECA	ST	153	21/9
SP6 CINECA	ST	192	21/9	mts_forpp.f
SP6 CINECA	ST	194	21/9	mts_forpp.f	-qstrict -qhot=vector
SP6 CINECA	ST	110	21/9	mts_forpp.f	-qnostrict -qhot=vector
SP6 CINECA	ST	186	21/9	mts_forpp.f		no exchange (STEP=500000)
SP6 CINECA	ST	146	22/9	-	-qnostrict -qhot=vector

MACHINE	MOD	CPU	date(2011)	vector code in	compiler options	NOTES
SP6 CINECA	SMT	173	21/9	mts_forpp.f

MACHINE	MOD	CPU	date(2011)	compiler options	NOTES
Cluster Linux	50%	1290			24 cores
SP6 CINECA	SMT	986	9/9
SP6 CINECA	SMT	998	9/9	-O4 -qstrict
SP6 CINECA	ST	886	9/9
SP6 CINECA	ST	813	9/9	-O4 -qstrict

MACHINE	MOD	CPU	date(2011)	vector code in	compiler options
SP6 CINECA	ST	1202	21/9	mts_forpp_rem_vector.f
SP6 CINECA	ST	742	21/9	mts_forpp_rem_vector.f	-qnostrict
SP6 CINECA	ST	767	21/9
SP6 CINECA	ST	571	21/9		-qnostrict (all)
SP6 CINECA	ST	610	21/9		-qnostrict (mts* only)
SP6 CINECA	ST	722	21/9	mts_forpp_rem_vector.f	-O4 [-qnostrict]
SP6 CINECA	ST	570	21/9		-O4 [-qnostrict]
SP6 CINECA	SMT	373	22/9		-qnostrict (all)

MACHINE	MOD	CPU	date(2011)	vector code in	compiler options	NOTES
SP6 CINECA	ST	765	22/9
SP6 CINECA	ST	390	22/9		-qnostrict (all)

Results depend on input data: vector code is effective only when there is no solvent (or, solvent is input as "solute")
In serial tests:
- basic_tests/1.in (atoms: solute(includ.solv)=1400) benefits from vectorialized code and, to a lesser extent, from compiler optimization
- REM_tests/1.in (atoms: solute=22, solvent=1500) vectorialized code is slower
In parallel tests:
- REM_tests/1.in:
  - vector code is slower than non-vector with no optimization, but faster with optimization.
  - optimization of non-vector is nearly uneffective
  - communication between processes has a limited effect on time consumption
- Protein+ligand (atoms: 2 solutes=1700, solvent=10000)
  - vector code is much slower than (non-optimized) non-vector code with no optimization, equally fast with optimization.
  - non-vector code is faster with optimization
Explanation: mts_forpp routines are designed for Protein-Protein interactions: protein-solvent interactions are not vectorized; when these are dominant, vectorized code is not efficient. In these cases, a better job is done by the compiler optimization.