Parent Directory
| 
Revision Log
| 
Revision Graph
Changes to restore/introduce true shared memory (i.e. pure SMP model) support
for lat-lon, cartesian and cube.
o Threads support restored for global_sum, global_max.
Needed COMMON blocks in place of stack variables
for work space.
o COMMON added for temps that are exchanged
shap_filt, solve_for_pressure
o exch...template edits to allow multithreaded calling
of cube exchanges. Cube exchanges are done single threaded,
sequential at the moment and are hard coded for six-tile,
one tile per face configuration.
o Threading support only tested on Sun SMP (6500) system.
o Default is SYSTEM() exists (check_threads.F), and barrier and memsync
are used for threaded exchanges (ini_communications_patterns).
o Killed K-level plot (ini_masks_etc.F). Overlaps not set in
multi-threads (local array), so fails in plot_field().
o Fixed t-ave of phihyd ( DYNAMICS() ). Can only do it for
first tile.
o Added commented out nTx settings to hs94.cs-32x32x5 eedata.
#nTx=2
#nTx=3
#nTx=6
Need to activate for setenv PARALLEL
o Multi-threaded pickup is 100% validated.
Testing notes
o Code passes "testscript" tests.
o To get "hs94.cs-32x32x5" to run OK need to patch
dynamics.F. Patch is in hs94.cs-32x32x5/code/dynamics.patch.
o Multithreading only tested on Sun 6500 SMP for hs94.cs and
hs94.128. hs94.128 was OK at c37 but has a problem under multi-threading
at pre38 at the moment.
o Under Solaris need one of
setenv PARALLEL 2
setenv PARALLEL 3
setenv PARALLEL 6
setenv PARALLEL 8 ( hs94.128 only )
- and -
setenv STACKSIZE 16384
o Under Solaris f77 need to remove -u from Makefile FOPTS for compiling
fftpack.o. Need to declare everything in fftpack or have
custom compile options for fftpack.F to fix this.
| 1 | C $Header: /u/gcmpack/models/MITgcmUV/eesupp/inc/BARRIER.h,v 1.3 2001/02/04 14:38:41 cnh Exp $ |
| 2 | C $Name: checkpoint37 $ |
| 3 | C |
| 4 | C /==========================================================\ |
| 5 | C | GLOBAL_MAX.h | |
| 6 | C | o Globals used by Fortran global max routine. | |
| 7 | C |==========================================================| |
| 8 | C | The global max shared memory scheme uses global heap data| |
| 9 | C | structures (.i.e COMMON blocks ). Each thread writes to | |
| 10 | C | an its own element of the shared memory array and then | |
| 11 | C | one thread reads all the entries and maxs them. The max | |
| 12 | C | result is then read by all threads. | |
| 13 | C | Remember - you are working with regions of memory that | |
| 14 | C | are being updated concurrently by different threads. | |
| 15 | C | What happens, when it happens and who gets to see what | |
| 16 | C | happens at what stage depends on the computer systems | |
| 17 | C | memory model. Every computer has a different memory model| |
| 18 | C | and they are never simple. In all current platforms it is| |
| 19 | C | possible for one thread to see events happening in a | |
| 20 | C | different order from the order they are written in the | |
| 21 | C | code. | |
| 22 | C | Unless you understand this it is not a good idea to | |
| 23 | C | make modifications te way these header files are setup or| |
| 24 | C | the way the global sum routines work. | |
| 25 | C \==========================================================/ |
| 26 | COMMON / GMAX_COMMON_R8 / phiGMRL |
| 27 | Real*8 phiGMRL(lShare8, MAX_NO_THREADS ) |
| 28 | |
| 29 | COMMON / GMAX_COMMON_R4 / phiGMRS |
| 30 | Real*4 phiGMRS(lShare4, MAX_NO_THREADS ) |
| 31 | |
| 32 | COMMON / GMAX_COMMON_I / phiGMI |
| 33 | INTEGER phiGMI (lShare4, MAX_NO_THREADS ) |
| ViewVC Help | |
| Powered by ViewVC 1.1.22 |