Directory /afs/cern.ch/eng/sl/lintrack/Beta-Beat.src/DRIVE_src.Harry contains a parallelised version of Drive_God_lin using the gcc and Intel Fortran IA32 32-bit Openmp compiler options. A good Openmp tutorial is at: https://computing.llnl.gov/tutorials/openMP/ To use the Intel compiler on CERN lxplus needs to execute:
source /afs/cern.ch/sw/IntelSoftware/linux/all-setup.csh ia32
For the number of cores to use enter: setenv OMP_NUM_THREADS n where n is from 1 (use for validation runs) to the number available. Tests show the optimum is about 12 - probably due to i/o issues.
This version requires the number of turns to be processed to be given in the DrivingTerms input file. If this is more than actually in the BPM data then extra values of zero will be filled and used.
Files in this directory are:
README these instructions Drive_God_lin.c the parallelised C-code with #pragma omp constructs. sussix4drivexxNoO.f the parallelised Fortran with !$OMP constructs. Makefile the make file for building the executable (needs cleaning!) Drive_God_lin the executable created by running make in this directory Drive_God_lin.o the compiled C-code sussix4drivexxNoO.o the compiled Fortran TEST a directory containing the TEST data from E.Macleod The TEST directory contains required input files and also ALLBPMs_linx.orig and ALLBPMs_liny.orig to compare with the TEST output from the original non-parallel version.
The Makefile uses the static linking option and does not need the Intel compiler environment on lxplus but may do on other machines.
To run the test case in single and multicore modes:
cd TEST setenv OMP_NUM_THREADS 1 time ../Drive_God_lin . > ! testout which on lxplus436 gives: 43.715u 0.220s 0:44.58 98.5% then setenv OMP_NUM_THREADS 12 time ../Drive_God_lin . > ! testout 62.255u 0.456s 0:06.26 1001.5% so real time from 44 to 6 seconds and effectively about 10 cores being used. Note that the stdout in the testout file is in a random order in the multicore case while the ALLBPMs files are sorted into the original order. There is often a small numerical difference in the 6th or 7th significant figures of the Q2RMS values in line 2 of the ALLBPMs linx and liny files between the single and multicore executions. This is put down to the different order in which the tune values are summed during parallel execution to which the RMS values are sensitive and is not considered important.
A detail is that a minor bug whereby the last bpm data was processed twice wasting a little time but not affecting the results has been fixed at line 581 in Drive_God_lin.c. Another minor problem is that the sussix_v4.inp file created by Drive_God_lin.c defines NLINE= 0 which, using the Intel Fortran compiler, turns the following read of L, M, K into a zero-trip loop and all values after are taken out of sequence. This does not affect execution so has been left as it is but if NLINE is given a value then ISME= 1 would create a smear file so this has been changed to ISME= 0 at line 1337 in Drive_God_lin.c.
A change was made on 29.03.2012 to fix a bug when executing with more than one thread. The global constant WINDOW values were mistakenly included in the Drive_God_lin.c threadprivate pragma meaning they were undefined (zeroes) in secondary threads. This resulted in the noise1, co and co2 values being computed as zero in all secondary threads.