| 1 |
C $Header: /u/gcmpack/models/MITgcmUV/eesupp/src/read_field.F,v 1.6 2001/02/04 14:38:44 cnh Exp $ |
| 2 |
C $Name: $ |
| 3 |
|
| 4 |
#include "CPP_EEOPTIONS.h" |
| 5 |
|
| 6 |
SUBROUTINE READ_FIELD_XYZR8( |
| 7 |
O fld, |
| 8 |
I filNam, filFmt, myThid ) |
| 9 |
C /==========================================================\ |
| 10 |
C | SUBROUTINE READ_FIELD_XYZR8 | |
| 11 |
C | o Reads a file into three-dimensional model array | |
| 12 |
C |==========================================================| |
| 13 |
C | Routine that controls the reading of external datasets | |
| 14 |
C | into the model. In a multi-threaded and/or MPI world | |
| 15 |
C | this can be a non-trivial exercise. Here we use the | |
| 16 |
C | following approach: | |
| 17 |
C | Thread 1. reads data for the complete domain i.e. all | |
| 18 |
C | processes and all threads into a buffer. Each individual | |
| 19 |
C | thread then reads its portion of data into the actual | |
| 20 |
C | model array. This is clean because there is just one | |
| 21 |
C | input file with a single format irrespective of the | |
| 22 |
C | of processes or threads in use. However, it has several | |
| 23 |
C | potential difficulties. | |
| 24 |
C | 1. Very large problems may have individual fields of | |
| 25 |
C | several GB. For example 1/20th degree global and | |
| 26 |
C | fifty levels is 10GB per field at 8 byte precision. | |
| 27 |
C | 2. MPI 1.nn is vague about I/O support - not all | |
| 28 |
C | processes have to support I/O. | |
| 29 |
C | MPI 2. includes a standard API for distributed data, | |
| 30 |
C | parallel I/O. If applications funnel all their field | |
| 31 |
C | I/O through this routine then adopting this or some | |
| 32 |
C | alternative should be fairly straight-forward. | |
| 33 |
C | In the light of problem 1. the following strategy | |
| 34 |
C | is adopted. Files are read one layer at a time. After | |
| 35 |
C | each layer has been read there is a barrier and then | |
| 36 |
C | the threads all copy data from the buffer to the arrays.| |
| 37 |
C | This creates a lower-performance I/O code but reduces | |
| 38 |
C | the degree to which a single large array is required for| |
| 39 |
C | the master thread. To be consistent with this binary | |
| 40 |
C | input files must be written by code of the form | |
| 41 |
C | WRITE(N) ((array(I,J,K),I=1,Nx),J=1,Ny) | |
| 42 |
C | rather than of the form | |
| 43 |
C | WRITE(N) array | |
| 44 |
C | The approach taken here also avoids one other ugly | |
| 45 |
C | behaviour. On several systems even Fortran internal | |
| 46 |
C | reads and writes are not thread-safe. This means that | |
| 47 |
C | the portion of the code that builds file names has to | |
| 48 |
C | be a critical section. However, if only the master | |
| 49 |
C | thread is interested in the value of the file name then | |
| 50 |
C | only the master need set its value. | |
| 51 |
C | Finally the IO performed here is for the whole XY | |
| 52 |
C | domain - even under MPI. The input files can stay the | |
| 53 |
C | same no matter what processor count is being used. | |
| 54 |
C | This is not a scalable approach to IO and MPI 2 has much| |
| 55 |
C | better support for this. Output is handled differently. | |
| 56 |
C | By default output files are written split and have to be| |
| 57 |
C | merged in a post-processing stage - YUK! | |
| 58 |
C \==========================================================/ |
| 59 |
IMPLICIT NONE |
| 60 |
|
| 61 |
C == GLobal variables == |
| 62 |
#include "SIZE.h" |
| 63 |
#include "EEPARAMS.h" |
| 64 |
#include "EESUPPORT.h" |
| 65 |
#include "EEIO.h" |
| 66 |
|
| 67 |
C == Routine arguments == |
| 68 |
C fld - Array into which data will be written. |
| 69 |
C filNam - Name of file to read. |
| 70 |
C filFmt - Format to use to read the file. |
| 71 |
C myNz - No. vertical layers for array fld. |
| 72 |
C myThid - Thread number for this instance of the routine. |
| 73 |
_RL fld(1-OLx:sNx+OLx,1-OLy:sNy+OLy,1:Nr,nSx, nSy ) |
| 74 |
CHARACTER*(*) filNam |
| 75 |
CHARACTER*(*) filFmt |
| 76 |
INTEGER myThid |
| 77 |
|
| 78 |
#ifdef USE_EEIO |
| 79 |
|
| 80 |
C == Local variables == |
| 81 |
C msgBuf - Variable for writing error messages |
| 82 |
C I,J,K, bi,bj - Loop counters |
| 83 |
C dUnit - Unit number for file I/O |
| 84 |
C ioStatus - I/O error code |
| 85 |
CHARACTER*(MAX_LEN_MBUF) msgBuf |
| 86 |
INTEGER I |
| 87 |
INTEGER J |
| 88 |
INTEGER K |
| 89 |
INTEGER bi |
| 90 |
INTEGER bj |
| 91 |
INTEGER iG, jG |
| 92 |
INTEGER dUnit |
| 93 |
INTEGER ioStatus |
| 94 |
C |
| 95 |
dUnit = 42 |
| 96 |
|
| 97 |
C-- Open the file |
| 98 |
C Note: The error trapping here is inelegant. There is no |
| 99 |
C easy way to tell other threads and/or MPI processes that |
| 100 |
C there was an error. Here we simply STOP if there is an error. |
| 101 |
C Under a multi-threaded mode this will halt all the threads. |
| 102 |
C Under MPI the other processes may die or they may just hang! |
| 103 |
_BEGIN_MASTER(myThid) |
| 104 |
OPEN(dUnit,FILE=filNam,FORM='unformatted',STATUS='old', |
| 105 |
& IOSTAT=ioStatus) |
| 106 |
IF ( ioStatus .GT. 0 ) THEN |
| 107 |
WRITE(msgBuf,'(A)') |
| 108 |
& 'S/R READ_FIELD_XYZR8' |
| 109 |
CALL PRINT_ERROR( msgBuf , myThid) |
| 110 |
WRITE(msgBuf,'(A)') |
| 111 |
& 'Open for read failed for' |
| 112 |
CALL PRINT_ERROR( msgBuf , myThid) |
| 113 |
WRITE(msgBuf,'(A,A50)') |
| 114 |
& 'file ',filNam |
| 115 |
CALL PRINT_ERROR( msgBuf , myThid) |
| 116 |
STOP 'ABNORMAL END: S/R READ_FIELD_XYZR8' |
| 117 |
ENDIF |
| 118 |
_END_MASTER(myThid) |
| 119 |
|
| 120 |
DO K = 1, Nr |
| 121 |
C-- Read data from file one XY layer at a time |
| 122 |
_BEGIN_MASTER(myThid) |
| 123 |
C READ ... |
| 124 |
DO J=1,Ny |
| 125 |
DO I=1,Nx |
| 126 |
IF ( filNam(1:1) .EQ. 'u' ) THEN |
| 127 |
IO_tmpXY_R8(I,J) = 0.0 _d 0 |
| 128 |
IF ( J .GT. 15 .AND. J .LT. 24 ) |
| 129 |
& IO_tmpXY_R8(I,J) = 0.1 _d 0 |
| 130 |
ELSEIF ( filNam(1:1) .EQ. 'v' ) THEN |
| 131 |
IO_tmpXY_R8(I,J) = 0.0 _d 0 |
| 132 |
ELSE |
| 133 |
IO_tmpXY_R8(I,J) = 0.0 _d 0 |
| 134 |
ENDIF |
| 135 |
ENDDO |
| 136 |
ENDDO |
| 137 |
_END_MASTER(myThid) |
| 138 |
_BARRIER |
| 139 |
C-- Copy data into per thread data structures |
| 140 |
DO bj=myByLo(myThid),myByHi(myThid) |
| 141 |
DO bi=myBxLo(myThid),myBxHi(myThid) |
| 142 |
DO j=1,sNy |
| 143 |
DO i=1,sNx |
| 144 |
iG = myXGlobalLo+(bi-1)*sNx+I-1 |
| 145 |
jG = myYGlobalLo+(bj-1)*sNy+J-1 |
| 146 |
fld(i,j,k,bi,bj) = IO_tmpXY_R8(iG,jG) |
| 147 |
ENDDO |
| 148 |
ENDDO |
| 149 |
ENDDO |
| 150 |
ENDDO |
| 151 |
_BARRIER |
| 152 |
ENDDO |
| 153 |
C |
| 154 |
_EXCH_XYZ_R8(fld, myThid ) |
| 155 |
C |
| 156 |
#endif |
| 157 |
|
| 158 |
RETURN |
| 159 |
END |
Parent Directory
| 
Revision Log
| 
Revision Graph