pio_server.c 37.5 KB
Newer Older
Deike Kleberg's avatar
Deike Kleberg committed
1
2
/** @file ioServer.c
*/
3
4
5
6
7
8
#ifdef HAVE_CONFIG_H
#  include "config.h"
#endif

#ifdef USE_MPI

Deike Kleberg's avatar
Deike Kleberg committed
9
10
11
#include "pio_server.h"


12
#include <limits.h>
Deike Kleberg's avatar
Deike Kleberg committed
13
14
#include <stdlib.h>
#include <stdio.h>
15
16
17
18
19
20

#ifdef HAVE_PARALLEL_NC4
#include <core/ppm_combinatorics.h>
#include <core/ppm_rectilinear.h>
#include <ppm/ppm_uniform_partition.h>
#endif
21
#include <yaxt.h>
22

Deike Kleberg's avatar
Deike Kleberg committed
23
#include "cdi.h"
24
#include "namespace.h"
25
#include "taxis.h"
Deike Kleberg's avatar
Deike Kleberg committed
26
#include "pio.h"
Deike Kleberg's avatar
Deike Kleberg committed
27
#include "pio_comm.h"
28
#include "pio_interface.h"
Deike Kleberg's avatar
Deike Kleberg committed
29
#include "pio_rpc.h"
Deike Kleberg's avatar
Deike Kleberg committed
30
#include "pio_util.h"
31
#include "cdi_int.h"
32
33
34
#ifndef HAVE_NETCDF_PAR_H
#define MPI_INCLUDED
#endif
35
#include "pio_cdf_int.h"
36
#include "resource_handle.h"
37
#include "resource_unpack.h"
Thomas Jahns's avatar
Thomas Jahns committed
38
#include "stream_cdf.h"
Deike Kleberg's avatar
Deike Kleberg committed
39
#include "vlist_var.h"
40

41

42
extern resOps streamOps;
43
extern void arrayDestroy ( void );
Deike Kleberg's avatar
Deike Kleberg committed
44

45
46
47
static struct
{
  size_t size;
Thomas Jahns's avatar
Thomas Jahns committed
48
  unsigned char *buffer;
49
  int dictSize;
50
51
} *rxWin = NULL;

Thomas Jahns's avatar
Thomas Jahns committed
52
static MPI_Win getWin = MPI_WIN_NULL;
Thomas Jahns's avatar
Thomas Jahns committed
53
static MPI_Group groupModel = MPI_GROUP_NULL;
Deike Kleberg's avatar
Deike Kleberg committed
54

55
56
57
58
59
#ifdef HAVE_PARALLEL_NC4
/* prime factorization of number of pio collectors */
static uint32_t *pioPrimes;
static int numPioPrimes;
#endif
Deike Kleberg's avatar
Deike Kleberg committed
60

Deike Kleberg's avatar
Deike Kleberg committed
61
62
/************************************************************************/

63
static
Deike Kleberg's avatar
Deike Kleberg committed
64
65
void serverWinCleanup ()
{
66
67
  if (getWin != MPI_WIN_NULL)
    xmpi(MPI_Win_free(&getWin));
68
69
  if (rxWin)
    {
70
      free(rxWin[0].buffer);
71
      free(rxWin);
Deike Kleberg's avatar
Deike Kleberg committed
72
    }
73

74
  xdebug("%s", "cleaned up mpi_win");
Deike Kleberg's avatar
Deike Kleberg committed
75
}
76

Deike Kleberg's avatar
Deike Kleberg committed
77
 /************************************************************************/
78

79
80
static size_t
collDefBufferSizes()
Deike Kleberg's avatar
Deike Kleberg committed
81
{
82
  int nstreams, * streamIndexList, streamNo, vlistID, nvars, varID, iorank;
83
84
  int modelID;
  size_t sumGetBufferSizes = 0;
85
  int rankGlob = commInqRankGlob ();
Deike Kleberg's avatar
Deike Kleberg committed
86
  int nProcsModel = commInqNProcsModel ();
87
  int root = commInqRootGlob ();
Deike Kleberg's avatar
Deike Kleberg committed
88

89
  xassert(rxWin != NULL);
Deike Kleberg's avatar
Deike Kleberg committed
90

Deike Kleberg's avatar
Deike Kleberg committed
91
  nstreams = reshCountType ( &streamOps );
92
93
  streamIndexList = xmalloc ( nstreams * sizeof ( streamIndexList[0] ));
  reshGetResHListOfType ( nstreams, streamIndexList, &streamOps );
Deike Kleberg's avatar
Deike Kleberg committed
94
95
  for ( streamNo = 0; streamNo < nstreams; streamNo++ )
    {
96
      // space required for data
97
      vlistID = streamInqVlist ( streamIndexList[streamNo] );
Deike Kleberg's avatar
Deike Kleberg committed
98
99
100
101
      nvars = vlistNvars ( vlistID );
      for ( varID = 0; varID < nvars; varID++ )
        {
          iorank = vlistInqVarIOrank ( vlistID, varID );
Deike Kleberg's avatar
Deike Kleberg committed
102
          xassert ( iorank != CDI_UNDEFID );
Deike Kleberg's avatar
Deike Kleberg committed
103
104
          if ( iorank == rankGlob )
            {
Deike Kleberg's avatar
Deike Kleberg committed
105
              for ( modelID = 0; modelID < nProcsModel; modelID++ )
106
                {
107
108
109
110
111
112
113
114
                  int decoChunk;
                  {
                    int varSize = vlistInqVarSize(vlistID, varID);
                    int nProcsModel = commInqNProcsModel();
                    decoChunk =
                      (int)ceilf(cdiPIOpartInflate_
                                 * (varSize + nProcsModel - 1)/nProcsModel);
                  }
Deike Kleberg's avatar
Deike Kleberg committed
115
                  xassert ( decoChunk > 0 );
116
                  rxWin[modelID].size += decoChunk * sizeof (double)
117
118
119
120
121
122
123
124
                    /* re-align chunks to multiple of double size */
                    + sizeof (double) - 1
                    /* one header for data record, one for
                     * corresponding part descriptor*/
                    + 2 * sizeof (union winHeaderEntry)
                    /* FIXME: heuristic for size of packed Xt_idxlist */
                    + sizeof (Xt_int) * decoChunk * 3;
                  rxWin[modelID].dictSize += 2;
125
                }
Deike Kleberg's avatar
Deike Kleberg committed
126
            }
127
        }
Deike Kleberg's avatar
Deike Kleberg committed
128
129
      // space required for the 3 function calls streamOpen, streamDefVlist, streamClose 
      // once per stream and timestep for all collprocs only on the modelproc root
130
      rxWin[root].size += numRPCFuncs * sizeof (union winHeaderEntry)
131
132
133
134
        /* serialized filename */
        + MAXDATAFILENAME
        /* data part of streamDefTimestep */
        + (2 * CDI_MAX_NAME + sizeof (taxis_t));
135
      rxWin[root].dictSize += numRPCFuncs;
Deike Kleberg's avatar
Deike Kleberg committed
136
    }
137
  free ( streamIndexList );
Deike Kleberg's avatar
Deike Kleberg committed
138
139

  for ( modelID = 0; modelID < nProcsModel; modelID++ )
140
    {
141
      /* account for size header */
142
      rxWin[modelID].dictSize += 1;
143
      rxWin[modelID].size += sizeof (union winHeaderEntry);
144
145
146
      rxWin[modelID].size = roundUpToMultiple(rxWin[modelID].size,
                                              PIO_WIN_ALIGN);
      sumGetBufferSizes += (size_t)rxWin[modelID].size;
147
    }
Deike Kleberg's avatar
Deike Kleberg committed
148
  xassert ( sumGetBufferSizes <= MAXWINBUFFERSIZE );
149
  return sumGetBufferSizes;
Deike Kleberg's avatar
Deike Kleberg committed
150
}
151

Deike Kleberg's avatar
Deike Kleberg committed
152
 /************************************************************************/
153
154

static 
Deike Kleberg's avatar
Deike Kleberg committed
155
156
 void serverWinCreate ()
{ 
Deike Kleberg's avatar
Deike Kleberg committed
157
  int ranks[1], modelID;
158
  MPI_Comm commCalc = commInqCommCalc ();
Deike Kleberg's avatar
Deike Kleberg committed
159
  MPI_Group groupCalc;
160
  int nProcsModel = commInqNProcsModel ();
Deike Kleberg's avatar
Deike Kleberg committed
161

Deike Kleberg's avatar
Deike Kleberg committed
162
  xmpi ( MPI_Win_create ( MPI_BOTTOM, 0, 1, MPI_INFO_NULL,
163
                          commCalc, &getWin ));
Deike Kleberg's avatar
Deike Kleberg committed
164
165

  /* target group */
166
167
  ranks[0] = nProcsModel;
  xmpi ( MPI_Comm_group ( commCalc, &groupCalc ));
Deike Kleberg's avatar
Deike Kleberg committed
168
169
  xmpi ( MPI_Group_excl ( groupCalc, 1, ranks, &groupModel ));

170
  rxWin = xcalloc(nProcsModel, sizeof (rxWin[0]));
171
172
173
174
175
176
177
178
  size_t totalBufferSize = collDefBufferSizes();
  rxWin[0].buffer = xmalloc(totalBufferSize);
  size_t ofs = 0;
  for ( modelID = 1; modelID < nProcsModel; modelID++ )
    {
      ofs += rxWin[modelID - 1].size;
      rxWin[modelID].buffer = rxWin[0].buffer + ofs;
    }
Deike Kleberg's avatar
Deike Kleberg committed
179

180
  xdebug("%s", "created mpi_win, allocated getBuffer");
Deike Kleberg's avatar
Deike Kleberg committed
181
182
}

Deike Kleberg's avatar
Deike Kleberg committed
183
184
/************************************************************************/

185
186
static void
readFuncCall(struct funcCallDesc *header)
Deike Kleberg's avatar
Deike Kleberg committed
187
188
{
  int root = commInqRootGlob ();
189
  int funcID = header->funcID;
Deike Kleberg's avatar
Deike Kleberg committed
190

191
  xassert(funcID >= MINFUNCID && funcID <= MAXFUNCID);
Deike Kleberg's avatar
Deike Kleberg committed
192
193
  switch ( funcID )
    {
194
195
    case STREAMCLOSE:
      {
196
197
        int streamID
          = namespaceAdaptKey2(header->funcArgs.streamChange.streamID);
198
199
200
201
        streamClose(streamID);
        xdebug("READ FUNCTION CALL FROM WIN:  %s, streamID=%d,"
               " closed stream",
               funcMap[(-1 - funcID)], streamID);
202
203
      }
      break;
Deike Kleberg's avatar
Deike Kleberg committed
204
    case STREAMOPEN:
205
      {
206
        size_t filenamesz = header->funcArgs.newFile.fnamelen;
Deike Kleberg's avatar
Deike Kleberg committed
207
        xassert ( filenamesz > 0 && filenamesz < MAXDATAFILENAME );
208
209
210
211
212
213
        const char *filename
          = (const char *)(rxWin[root].buffer
                           + header->funcArgs.newFile.offset);
        xassert(filename[filenamesz] == '\0');
        int filetype = header->funcArgs.newFile.filetype;
        int streamID = streamOpenWrite(filename, filetype);
214
        xassert(streamID != CDI_ELIBNAVAIL);
215
216
        xdebug("READ FUNCTION CALL FROM WIN:  %s, filenamesz=%zu,"
               " filename=%s, filetype=%d, OPENED STREAM %d",
217
               funcMap[(-1 - funcID)], filenamesz, filename,
218
               filetype, streamID);
219
      }
220
      break;
221
222
    case STREAMDEFVLIST:
      {
223
224
225
        int streamID
          = namespaceAdaptKey2(header->funcArgs.streamChange.streamID);
        int vlistID = namespaceAdaptKey2(header->funcArgs.streamChange.vlistID);
226
227
228
229
        streamDefVlist(streamID, vlistID);
        xdebug("READ FUNCTION CALL FROM WIN:  %s, streamID=%d,"
               " vlistID=%d, called streamDefVlist ().",
               funcMap[(-1 - funcID)], streamID, vlistID);
230
231
      }
      break;
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
    case STREAMDEFTIMESTEP:
      {
        MPI_Comm commCalc = commInqCommCalc ();
        int streamID = header->funcArgs.streamNewTimestep.streamID;
        int nspTarget = namespaceResHDecode(streamID).nsp;
        streamID = namespaceAdaptKey2(streamID);
        int tsID
          = header->funcArgs.streamNewTimestep.tsID;
        int oldTaxisID
          = vlistInqTaxis(streamInqVlist(streamID));
        int position = header->funcArgs.streamNewTimestep.offset;
        int changedTaxisID
          = taxisUnpack((char *)rxWin[root].buffer, (int)rxWin[root].size,
                        &position, nspTarget, &commCalc, 0);
        taxis_t *oldTaxisPtr = taxisPtr(oldTaxisID);
        taxis_t *changedTaxisPtr = taxisPtr(changedTaxisID);
        ptaxisCopy(oldTaxisPtr, changedTaxisPtr);
        taxisDestroy(changedTaxisID);
        streamDefTimestep(streamID, tsID);
      }
      break;
Deike Kleberg's avatar
Deike Kleberg committed
253
    default:
254
      xabort ( "REMOTE FUNCTIONCALL NOT IMPLEMENTED!" );
Deike Kleberg's avatar
Deike Kleberg committed
255
256
257
258
259
    }
}

/************************************************************************/

260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
static void
resizeVarGatherBuf(int vlistID, int varID, double **buf, int *bufSize)
{
  int size = vlistInqVarSize(vlistID, varID);
  if (size <= *bufSize) ; else
    *buf = xrealloc(*buf, (*bufSize = size) * sizeof (buf[0][0]));
}

static void
gatherArray(int root, int nProcsModel, int headerIdx,
            int vlistID,
            double *gatherBuf, int *nmiss)
{
  union winHeaderEntry *winDict
    = (union winHeaderEntry *)rxWin[root].buffer;
  int streamID = winDict[headerIdx].dataRecord.streamID;
  int varID = winDict[headerIdx].dataRecord.varID;
277
  int varShape[3] = { 0, 0, 0 };
278
  cdiPioQueryVarDims(varShape, vlistID, varID);
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
  Xt_int varShapeXt[3];
  static const Xt_int origin[3] = { 0, 0, 0 };
  for (unsigned i = 0; i < 3; ++i)
    varShapeXt[i] = varShape[i];
  int varSize = varShape[0] * varShape[1] * varShape[2];
  int *partOfs = xmalloc(2 * varSize * sizeof (partOfs[0])),
    *gatherOfs = partOfs + varSize;
  Xt_idxlist *part = xmalloc(nProcsModel * sizeof (part[0]));
  MPI_Comm commCalc = commInqCommCalc();
  {
    int nmiss_ = 0, partOfsOfs = 0;
    for (int modelID = 0; modelID < nProcsModel; modelID++)
      {
        struct dataRecord *dataHeader
          = &((union winHeaderEntry *)
              rxWin[modelID].buffer)[headerIdx].dataRecord;
        struct partDescRecord *partHeader
          = &((union winHeaderEntry *)
              rxWin[modelID].buffer)[headerIdx + 1].partDesc;
        int position = partHeader->offset;
299
        xassert(namespaceAdaptKey2(dataHeader->streamID) == streamID
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
                && dataHeader->varID == varID
                && partHeader->partDescMarker == PARTDESCMARKER
                && position > 0
                && ((size_t)position
                    >= sizeof (union winHeaderEntry) * rxWin[modelID].dictSize)
                && ((size_t)position < rxWin[modelID].size));
        part[modelID] = xt_idxlist_unpack(rxWin[modelID].buffer,
                                          (int)rxWin[modelID].size,
                                          &position, commCalc);
        Xt_int partSize = xt_idxlist_get_num_indices(part[modelID]);
        size_t charOfs = (rxWin[modelID].buffer + dataHeader->offset)
          - rxWin[0].buffer;
        xassert(charOfs % sizeof (double) == 0
                && charOfs / sizeof (double) + partSize <= INT_MAX);
        int elemOfs = charOfs / sizeof (double);
        for (int i = 0; i < (int)partSize; ++i)
          partOfs[partOfsOfs + i] = elemOfs + i;
        partOfsOfs += partSize;
        nmiss_ += dataHeader->nmiss;
      }
    *nmiss = nmiss_;
  }
  Xt_idxlist srcList = xt_idxlist_collection_new(part, nProcsModel);
323
  for (int modelID = 0; modelID < nProcsModel; modelID++)
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
    xt_idxlist_delete(part[modelID]);
  free(part);
  Xt_xmap gatherXmap;
  {
    Xt_idxlist dstList
      = xt_idxsection_new(0, 3, varShapeXt, varShapeXt, origin);
    struct Xt_com_list full = { .list = dstList, .rank = 0 };
    gatherXmap = xt_xmap_intersection_new(1, &full, 1, &full, srcList, dstList,
                                        MPI_COMM_SELF);
    xt_idxlist_delete(dstList);
  }
  xt_idxlist_delete(srcList);
  for (int i = 0; i < varSize; ++i)
    gatherOfs[i] = i;

  Xt_redist gatherRedist
    = xt_redist_p2p_off_new(gatherXmap, partOfs, gatherOfs, MPI_DOUBLE);
  xt_xmap_delete(gatherXmap);
342
  xt_redist_s_exchange1(gatherRedist, rxWin[0].buffer, gatherBuf);
343
344
  free(partOfs);
  xt_redist_delete(gatherRedist);
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
}

struct xyzDims
{
  int sizes[3];
};

static inline int
xyzGridSize(struct xyzDims dims)
{
  return dims.sizes[0] * dims.sizes[1] * dims.sizes[2];
}

#ifdef HAVE_PARALLEL_NC4
static void
360
queryVarBounds(struct PPM_extent varShape[3], int vlistID, int varID)
361
{
362
363
  varShape[0].first = 0;
  varShape[1].first = 0;
364
  varShape[2].first = 0;
365
  int sizes[3];
366
  cdiPioQueryVarDims(sizes, vlistID, varID);
367
368
  for (unsigned i = 0; i < 3; ++i)
    varShape[i].size = sizes[i];
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
}

/* compute distribution of collectors such that number of collectors
 * <= number of variable grid cells in each dimension */
static struct xyzDims
varDimsCollGridMatch(const struct PPM_extent varDims[3])
{
  xassert(PPM_extents_size(3, varDims) >= commInqSizeColl());
  struct xyzDims collGrid = { { 1, 1, 1 } };
  /* because of storage order, dividing dimension 3 first is preferred */
  for (int i = 0; i < numPioPrimes; ++i)
    {
      for (int dim = 2; dim >=0; --dim)
        if (collGrid.sizes[dim] * pioPrimes[i] <= varDims[dim].size)
          {
            collGrid.sizes[dim] *= pioPrimes[i];
            goto nextPrime;
          }
      /* no position found, retrack */
      xabort("Not yet implemented back-tracking needed.");
      nextPrime:
      ;
    }
  return collGrid;
}

static void
myVarPart(struct PPM_extent varShape[3], struct xyzDims collGrid,
          struct PPM_extent myPart[3])
{
  int32_t myCollGridCoord[3];
  {
    struct PPM_extent collGridShape[3];
    for (int i = 0; i < 3; ++i)
      {
        collGridShape[i].first = 0;
        collGridShape[i].size = collGrid.sizes[i];
      }
    PPM_lidx2rlcoord_e(3, collGridShape, commInqRankColl(), myCollGridCoord);
    xdebug("my coord: (%d, %d, %d)", myCollGridCoord[0], myCollGridCoord[1],
           myCollGridCoord[2]);
  }
  PPM_uniform_partition_nd(3, varShape, collGrid.sizes,
                           myCollGridCoord, myPart);
}
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
#elif defined (HAVE_LIBNETCDF)
/* needed for writing when some files are only written to by a single process */
/* cdiOpenFileMap(fileID) gives the writer process */
int cdiPioSerialOpenFileMap(int streamID)
{
  return stream_to_pointer(streamID)->ownerRank;
}
/* for load-balancing purposes, count number of files per process */
/* cdiOpenFileCounts[rank] gives number of open files rank has to himself */
static int *cdiSerialOpenFileCount = NULL;
int cdiPioNextOpenRank()
{
  xassert(cdiSerialOpenFileCount != NULL);
  int commCollSize = commInqSizeColl();
  int minRank = 0, minOpenCount = cdiSerialOpenFileCount[0];
  for (int i = 1; i < commCollSize; ++i)
    if (cdiSerialOpenFileCount[i] < minOpenCount)
      {
        minOpenCount = cdiSerialOpenFileCount[i];
        minRank = i;
      }
  return minRank;
}

void cdiPioOpenFileOnRank(int rank)
{
  xassert(cdiSerialOpenFileCount != NULL
          && rank >= 0 && rank < commInqSizeColl());
  ++(cdiSerialOpenFileCount[rank]);
}


void cdiPioCloseFileOnRank(int rank)
{
  xassert(cdiSerialOpenFileCount != NULL
          && rank >= 0 && rank < commInqSizeColl());
  xassert(cdiSerialOpenFileCount[rank] > 0);
  --(cdiSerialOpenFileCount[rank]);
}

454
455
456
457
458
459
460
461
462
463
static void
cdiPioServerCdfDefVars(stream_t *streamptr)
{
  int rank, rankOpen;
  if (commInqIOMode() == PIO_NONE
      || ((rank = commInqRankColl())
          == (rankOpen = cdiPioSerialOpenFileMap(streamptr->self))))
    cdfDefVars(streamptr);
}

464
465
#endif

466
467
468
469
470
471
struct streamMapping {
  int streamID, filetype;
  int firstHeaderIdx, lastHeaderIdx;
  int numVars, *varMap;
};

472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
struct streamMap
{
  struct streamMapping *entries;
  int numEntries;
};

static struct streamMap
buildStreamMap(union winHeaderEntry *winDict)
{
  int streamIDOld = CDI_UNDEFID;
  int oldStreamIdx = CDI_UNDEFID;
  int filetype = CDI_UNDEFID;
  int sizeStreamMap = 16;
  struct streamMapping *streamMap = xmalloc(sizeStreamMap * sizeof (streamMap[0]));
  int numDataEntries = winDict[0].headerSize.numDataEntries;
  int numStreamIDs = 0;
  for (int headerIdx = 1; headerIdx < numDataEntries; headerIdx += 2)
    {
      int streamID
        = winDict[headerIdx].dataRecord.streamID
        = namespaceAdaptKey2(winDict[headerIdx].dataRecord.streamID);
      xassert(streamID > 0);
      if (streamID != streamIDOld)
        {
          for (int i = numStreamIDs - 1; i >= 0; --i)
            if ((streamIDOld = streamMap[i].streamID) == streamID)
              {
                filetype = streamMap[i].filetype;
                oldStreamIdx = i;
                goto streamIDInventorized;
              }
          if (numStreamIDs < sizeStreamMap) ; else
            streamMap = xrealloc(streamMap,
                                 (sizeStreamMap *= 2)
                                 * sizeof (streamMap[0]));
          streamMap[numStreamIDs].streamID = streamID;
          streamMap[numStreamIDs].firstHeaderIdx = headerIdx;
          streamMap[numStreamIDs].numVars = -1;
          oldStreamIdx = numStreamIDs;
          streamIDOld = streamID;
          filetype = streamInqFiletype(streamID);
          streamMap[numStreamIDs].filetype = filetype;
          if (filetype == FILETYPE_NC || filetype == FILETYPE_NC2
              || filetype == FILETYPE_NC4)
            {
              int vlistID = streamInqVlist(streamID);
              int nvars = vlistNvars(vlistID);
              streamMap[numStreamIDs].numVars = nvars;
              streamMap[numStreamIDs].varMap
                = xmalloc(sizeof (streamMap[numStreamIDs].varMap[0])
                          * nvars);
              for (int i = 0; i < nvars; ++i)
                streamMap[numStreamIDs].varMap[i] = -1;
            }
          ++numStreamIDs;
        }
      streamIDInventorized:
      streamMap[oldStreamIdx].lastHeaderIdx = headerIdx;
      if (filetype == FILETYPE_NC || filetype == FILETYPE_NC2
          || filetype == FILETYPE_NC4)
        {
          int varID = winDict[headerIdx].dataRecord.varID;
          streamMap[oldStreamIdx].varMap[varID] = headerIdx;
        }
    }
  return (struct streamMap){ .entries = streamMap, .numEntries = numStreamIDs };
}



542
static void readGetBuffers()
Deike Kleberg's avatar
Deike Kleberg committed
543
{
544
  int nProcsModel = commInqNProcsModel ();
Deike Kleberg's avatar
Deike Kleberg committed
545
  int root        = commInqRootGlob ();
546
#ifdef HAVE_NETCDF4
547
  int myCollRank = commInqRankColl();
548
  MPI_Comm collComm = commInqCommColl();
549
#endif
550
  xdebug("%s", "START");
551

552
553
554
  union winHeaderEntry *winDict
    = (union winHeaderEntry *)rxWin[root].buffer;
  xassert(winDict[0].headerSize.sizeID == HEADERSIZEMARKER);
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
  {
    int dictSize = rxWin[root].dictSize,
      firstNonRPCEntry = dictSize - winDict[0].headerSize.numRPCEntries - 1,
      headerIdx,
      numFuncCalls = 0;
    for (headerIdx = dictSize - 1;
         headerIdx > firstNonRPCEntry;
         --headerIdx)
      {
        struct funcCallDesc *header
          = &(winDict[headerIdx].funcCall);
        xassert(header->funcID >= MINFUNCID
                && header->funcID <= MAXFUNCID);
        ++numFuncCalls;
        readFuncCall(header);
      }
    xassert(numFuncCalls == winDict[0].headerSize.numRPCEntries);
  }
Thomas Jahns's avatar
Thomas Jahns committed
573
  /* build list of streams, data was transferred for */
574
  {
575
    struct streamMap map = buildStreamMap(winDict);
576
577
578
579
    double *data = NULL;
#if defined (HAVE_PARALLEL_NC4)
    double *writeBuf = NULL;
#endif
Thomas Jahns's avatar
Thomas Jahns committed
580
    int currentDataBufSize = 0;
581
    for (int streamIdx = 0; streamIdx < map.numEntries; ++streamIdx)
Thomas Jahns's avatar
Thomas Jahns committed
582
      {
583
        int streamID = map.entries[streamIdx].streamID;
Thomas Jahns's avatar
Thomas Jahns committed
584
        int vlistID = streamInqVlist(streamID);
585
        int filetype = map.entries[streamIdx].filetype;
Thomas Jahns's avatar
Thomas Jahns committed
586

587
        switch (filetype)
588
589
590
          {
          case FILETYPE_GRB:
          case FILETYPE_GRB2:
Thomas Jahns's avatar
Thomas Jahns committed
591
            {
592
593
              int headerIdx, lastHeaderIdx = map.entries[streamIdx].lastHeaderIdx;
              for (headerIdx = map.entries[streamIdx].firstHeaderIdx;
594
                   headerIdx <= lastHeaderIdx;
595
                   headerIdx += 2)
596
597
598
599
                if (streamID == winDict[headerIdx].dataRecord.streamID)
                  {
                    int varID = winDict[headerIdx].dataRecord.varID;
                    int size = vlistInqVarSize(vlistID, varID);
600
601
602
603
604
                    int nmiss;
                    resizeVarGatherBuf(vlistID, varID, &data,
                                       &currentDataBufSize);
                    gatherArray(root, nProcsModel, headerIdx,
                                vlistID, data, &nmiss);
605
606
607
608
609
610
611
612
613
                    streamWriteVar(streamID, varID, data, nmiss);
                    if ( ddebug > 2 )
                      {
                        char text[1024];
                        sprintf(text, "streamID=%d, var[%d], size=%d",
                                streamID, varID, size);
                        xprintArray(text, data, size, DATATYPE_FLT);
                      }
                  }
Thomas Jahns's avatar
Thomas Jahns committed
614
            }
615
            break;
616
617
618
619
620
621
622
#ifdef HAVE_NETCDF4
          case FILETYPE_NC:
          case FILETYPE_NC2:
          case FILETYPE_NC4:
#ifdef HAVE_PARALLEL_NC4
            /* HAVE_PARALLE_NC4 implies having ScalES-PPM and yaxt */
            {
623
624
              int nvars = map.entries[streamIdx].numVars;
              int *varMap = map.entries[streamIdx].varMap;
625
626
627
628
629
630
631
632
633
634
              int *varIsWritten = xmalloc(sizeof (varIsWritten[0]) * nvars);
              for (int varID = 0; varID < nvars; ++varID)
                varIsWritten[varID] = ((varMap[varID] != -1)
                                       ?myCollRank+1 : 0);
              xmpi(MPI_Allreduce(MPI_IN_PLACE, varIsWritten, nvars,
                                 MPI_INT, MPI_BOR, collComm));
              for (int varID = 0; varID < nvars; ++varID)
                if (varIsWritten[varID])
                  {
                    struct PPM_extent varShape[3];
635
                    queryVarBounds(varShape, vlistID, varID);
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
                    struct xyzDims collGrid = varDimsCollGridMatch(varShape);
                    xdebug("writing varID %d with dimensions: "
                           "x=%d, y=%d, z=%d,\n"
                           "found distribution with dimensions:"
                           " x=%d, y=%d, z=%d.", varID,
                           varShape[0].size, varShape[1].size, varShape[2].size,
                           collGrid.sizes[0], collGrid.sizes[1],
                           collGrid.sizes[2]);
                    struct PPM_extent varChunk[3];
                    myVarPart(varShape, collGrid, varChunk);
                    int myChunk[3][2];
                    for (int i = 0; i < 3; ++i)
                      {
                        myChunk[i][0] = PPM_extent_start(varChunk[i]);
                        myChunk[i][1] = PPM_extent_end(varChunk[i]);
                      }
                    xdebug("Writing chunk { { %d, %d }, { %d, %d },"
                           " { %d, %d } }", myChunk[0][0], myChunk[0][1],
                           myChunk[1][0], myChunk[1][1], myChunk[2][0],
                           myChunk[2][1]);
                    Xt_int varSize[3];
                    for (int i = 0; i < 3; ++i)
                      varSize[2 - i] = varShape[i].size;
                    Xt_idxlist preRedistChunk, preWriteChunk;
                    /* prepare yaxt descriptor for current data
                       distribution after collect */
                    int nmiss;
                    if (varMap[varID] == -1)
                      {
                        preRedistChunk = xt_idxempty_new();
                        xdebug("%s", "I got none\n");
                      }
                    else
                      {
                        Xt_int preRedistStart[3] = { 0, 0, 0 };
                        preRedistChunk
                          = xt_idxsection_new(0, 3, varSize, varSize,
                                              preRedistStart);
                        resizeVarGatherBuf(vlistID, varID, &data,
                                           &currentDataBufSize);
                        int headerIdx = varMap[varID];
                        gatherArray(root, nProcsModel, headerIdx,
                                    vlistID, data, &nmiss);
                        xdebug("%s", "I got all\n");
                      }
                    MPI_Bcast(&nmiss, 1, MPI_INT, varIsWritten[varID] - 1,
                              collComm);
                    /* prepare yaxt descriptor for write chunk */
                    {
                      Xt_int preWriteChunkStart[3], preWriteChunkSize[3];
                      for (int i = 0; i < 3; ++i)
                        {
                          preWriteChunkStart[2 - i] = varChunk[i].first;
                          preWriteChunkSize[2 - i] = varChunk[i].size;
                        }
                      preWriteChunk = xt_idxsection_new(0, 3, varSize,
                                                        preWriteChunkSize,
                                                        preWriteChunkStart);
                    }
                    /* prepare redistribution */
                    {
                      Xt_xmap xmap = xt_xmap_all2all_new(preRedistChunk,
                                                         preWriteChunk,
                                                         collComm);
                      Xt_redist redist = xt_redist_p2p_new(xmap, MPI_DOUBLE);
                      xt_idxlist_delete(preRedistChunk);
                      xt_idxlist_delete(preWriteChunk);
                      xt_xmap_delete(xmap);
                      writeBuf = xrealloc(writeBuf,
                                          sizeof (double)
                                          * PPM_extents_size(3, varChunk));
707
                      xt_redist_s_exchange1(redist, data, writeBuf);
708
709
710
711
712
713
714
715
716
717
                      xt_redist_delete(redist);
                    }
                    /* write chunk */
                    streamWriteVarChunk(streamID, varID,
                                        (const int (*)[2])myChunk, writeBuf,
                                        nmiss);
                  }
            }
#else
            /* determine process which has stream open (writer) and
718
719
720
             * which has data for which variable (var owner)
             * three cases need to be distinguished */
            {
721
722
              int nvars = map.entries[streamIdx].numVars;
              int *varMap = map.entries[streamIdx].varMap;
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
              int *varIsWritten = xmalloc(sizeof (varIsWritten[0]) * nvars);
              for (int varID = 0; varID < nvars; ++varID)
                varIsWritten[varID] = ((varMap[varID] != -1)
                                       ?myCollRank+1 : 0);
              xmpi(MPI_Allreduce(MPI_IN_PLACE, varIsWritten, nvars,
                                 MPI_INT, MPI_BOR, collComm));
              int writerRank;
              if ((writerRank = cdiPioSerialOpenFileMap(streamID))
                  == myCollRank)
                {
                  for (int varID = 0; varID < nvars; ++varID)
                    if (varIsWritten[varID])
                      {
                        int nmiss;
                        int size = vlistInqVarSize(vlistID, varID);
                        resizeVarGatherBuf(vlistID, varID, &data,
                                           &currentDataBufSize);
                        int headerIdx = varMap[varID];
                        if (varIsWritten[varID] == myCollRank + 1)
                          {
                            /* this process has the full array and will
                             * write it */
                            xdebug("gathering varID=%d for direct writing",
                                   varID);
                            gatherArray(root, nProcsModel, headerIdx,
                                        vlistID, data, &nmiss);
                          }
                        else
                          {
                            /* another process has the array and will
                             * send it over */
                            MPI_Status stat;
                            xdebug("receiving varID=%d for writing from"
                                   " process %d",
                                   varID, varIsWritten[varID] - 1);
                            xmpiStat(MPI_Recv(&nmiss, 1, MPI_INT,
                                              varIsWritten[varID] - 1,
                                              COLLBUFNMISS,
                                              collComm, &stat), &stat);
                            xmpiStat(MPI_Recv(data, size, MPI_DOUBLE,
                                              varIsWritten[varID] - 1,
                                              COLLBUFTX,
                                              collComm, &stat), &stat);
                          }
                        streamWriteVar(streamID, varID, data, nmiss);
                      }
                }
              else
                for (int varID = 0; varID < nvars; ++varID)
                  if (varIsWritten[varID] == myCollRank + 1)
                    {
                      /* this process has the full array and another
                       * will write it */
                      int nmiss;
                      int size = vlistInqVarSize(vlistID, varID);
                      resizeVarGatherBuf(vlistID, varID, &data,
                                         &currentDataBufSize);
                      int headerIdx = varMap[varID];
                      gatherArray(root, nProcsModel, headerIdx,
                                  vlistID, data, &nmiss);
                      MPI_Request req;
                      MPI_Status stat;
                      xdebug("sending varID=%d for writing to"
                             " process %d",
                             varID, writerRank);
                      xmpi(MPI_Isend(&nmiss, 1, MPI_INT,
                                     writerRank, COLLBUFNMISS,
                                     collComm, &req));
                      xmpi(MPI_Send(data, size, MPI_DOUBLE,
                                    writerRank, COLLBUFTX,
                                    collComm));
                      xmpiStat(MPI_Wait(&req, &stat), &stat);
                    }
            }
797
798
799
#endif
            break;
#endif
800
801
802
          default:
            xabort("unhandled filetype in parallel I/O.");
          }
803
      }
804
    free(map.entries);
Thomas Jahns's avatar
Thomas Jahns committed
805
    free(data);
806
  }
807
  xdebug("%s", "RETURN");
808
809
810
811
} 

/************************************************************************/

Deike Kleberg's avatar
Deike Kleberg committed
812

Thomas Jahns's avatar
Thomas Jahns committed
813
814
static
void clearModelWinBuffer(int modelID)
Deike Kleberg's avatar
Deike Kleberg committed
815
816
817
{
  int nProcsModel = commInqNProcsModel ();

Deike Kleberg's avatar
Deike Kleberg committed
818
819
  xassert ( modelID                >= 0           &&
            modelID                 < nProcsModel &&
820
            rxWin != NULL && rxWin[modelID].buffer != NULL &&
821
822
            rxWin[modelID].size > 0 &&
            rxWin[modelID].size <= MAXWINBUFFERSIZE );
823
  memset(rxWin[modelID].buffer, 0, rxWin[modelID].size);
Deike Kleberg's avatar
Deike Kleberg committed
824
825
826
827
828
829
}


/************************************************************************/


830
static
831
void getTimeStepData()
Deike Kleberg's avatar
Deike Kleberg committed
832
{
833
  int modelID;
834
  char text[1024];
835
  int nProcsModel = commInqNProcsModel ();
Thomas Jahns's avatar
Thomas Jahns committed
836
837
  void *getWinBaseAddr;
  int attrFound;
838

839
  xdebug("%s", "START");
Deike Kleberg's avatar
Deike Kleberg committed
840
841

  // todo put in correct lbs and ubs
842
  xmpi(MPI_Win_start(groupModel, 0, getWin));
843
844
  xmpi(MPI_Win_get_attr(getWin, MPI_WIN_BASE, &getWinBaseAddr, &attrFound));
  xassert(attrFound);
Deike Kleberg's avatar
Deike Kleberg committed
845
846
  for ( modelID = 0; modelID < nProcsModel; modelID++ )
    {
Thomas Jahns's avatar
Thomas Jahns committed
847
      clearModelWinBuffer(modelID);
848
      xdebug("modelID=%d, nProcsModel=%d, rxWin[%d].size=%zu,"
Thomas Jahns's avatar
Thomas Jahns committed
849
             " getWin=%p, sizeof(int)=%u",
850
             modelID, nProcsModel, modelID, rxWin[modelID].size,
Thomas Jahns's avatar
Thomas Jahns committed
851
             getWinBaseAddr, (unsigned)sizeof(int));
852
      /* FIXME: this needs to use MPI_PACK for portability */
853
854
855
      xmpi(MPI_Get(rxWin[modelID].buffer, rxWin[modelID].size,
                   MPI_UNSIGNED_CHAR, modelID, 0,
                   rxWin[modelID].size, MPI_UNSIGNED_CHAR, getWin));
Deike Kleberg's avatar
Deike Kleberg committed
856
    }
857
  xmpi ( MPI_Win_complete ( getWin ));
Deike Kleberg's avatar
Deike Kleberg committed
858

859
  if ( ddebug > 2 )
Deike Kleberg's avatar
Deike Kleberg committed
860
    for ( modelID = 0; modelID < nProcsModel; modelID++ )
861
      {
862
        sprintf(text, "rxWin[%d].size=%zu from PE%d rxWin[%d].buffer",
863
                modelID, rxWin[modelID].size, modelID, modelID);
864
        xprintArray(text, rxWin[modelID].buffer,
865
866
                    rxWin[modelID].size / sizeof (double),
                    DATATYPE_FLT);
867
      }
868
869
  readGetBuffers();

870
  xdebug("%s", "RETURN");
Deike Kleberg's avatar
Deike Kleberg committed
871
}
Deike Kleberg's avatar
Deike Kleberg committed
872
873
874

/************************************************************************/

875
876
877
878
879
880
881
882
883
884
885
886
#if defined (HAVE_LIBNETCDF) && ! defined (HAVE_PARALLEL_NC4)
static int
cdiPioStreamCDFOpenWrap(const char *filename, const char *filemode,
                        int filetype, stream_t *streamptr,
                        int recordBufIsToBeCreated)
{
  switch (filetype)
    {
    case FILETYPE_NC4:
    case FILETYPE_NC4C:
      {
        int rank, fileID;
Thomas Jahns's avatar
Thomas Jahns committed
887
888
        int ioMode = commInqIOMode();
        if (ioMode == PIO_NONE
889
890
891
892
            || commInqRankColl() == (rank = cdiPioNextOpenRank()))
          fileID = cdiStreamOpenDefaultDelegate(filename, filemode, filetype,
                                                streamptr,
                                                recordBufIsToBeCreated);
Thomas Jahns's avatar
Thomas Jahns committed
893
        if (ioMode != PIO_NONE)
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
          xmpi(MPI_Bcast(&fileID, 1, MPI_INT, rank, commInqCommColl()));
        streamptr->ownerRank = rank;
        return fileID;
      }
    default:
      return cdiStreamOpenDefaultDelegate(filename, filemode, filetype,
                                          streamptr, recordBufIsToBeCreated);
    }
}

static void
cdiPioStreamCDFCloseWrap(stream_t *streamptr, int recordBufIsToBeDeleted)
{
  int fileID   = streamptr->fileID;
  int filetype = streamptr->filetype;
  if ( fileID == CDI_UNDEFID )
    Warning("File %s not open!", streamptr->filename);
  else
    switch (filetype)
      {
      case FILETYPE_NC:
      case FILETYPE_NC2:
      case FILETYPE_NC4:
      case FILETYPE_NC4C:
        {
          int rank, rankOpen;
          if (commInqIOMode() == PIO_NONE
              || ((rank = commInqRankColl())
                  == (rankOpen = cdiPioSerialOpenFileMap(streamptr->self))))
            cdiStreamCloseDefaultDelegate(streamptr, recordBufIsToBeDeleted);
          break;
        }
      default:
        cdiStreamCloseDefaultDelegate(streamptr, recordBufIsToBeDeleted);
      }
}
Thomas Jahns's avatar
Thomas Jahns committed
930
931
932
933
934
935
936
937
938
939
940

static void
cdiPioCdfDefTimestep(stream_t *streamptr, int tsID)
{
  int rank, rankOpen, streamID = streamptr->self;
  if (commInqIOMode() == PIO_NONE
      || ((rank = commInqRankColl())
          == (rankOpen = cdiPioSerialOpenFileMap(streamID))))
    cdfDefTimestep(streamptr, tsID);
}

941
942
#endif

Deike Kleberg's avatar
Deike Kleberg committed
943
944
945
946
947
948
949
950
/**
  @brief is encapsulated in CDI library and run on I/O PEs.

  @param

  @return
*/

951
void IOServer ()
Deike Kleberg's avatar
Deike Kleberg committed
952
{
953
  int source, tag, size, nProcsModel=commInqNProcsModel();
954
  static int nfinished = 0;
Deike Kleberg's avatar
Deike Kleberg committed
955
  char * buffer;
956
957
  MPI_Comm commCalc;
  MPI_Status status;
Deike Kleberg's avatar
Deike Kleberg committed
958

959
  xdebug("%s", "START");
Deike Kleberg's avatar
Deike Kleberg committed
960

961
  backendInit ();
962
963
  if ( commInqRankNode () == commInqSpecialRankNode ()) 
    backendFinalize ();
964
  commCalc = commInqCommCalc ();
965
#ifdef HAVE_PARALLEL_NC4
966
  cdiPioEnableNetCDFParAccess();
967
968
  numPioPrimes = PPM_prime_factorization_32((uint32_t)commInqSizeColl(),
                                            &pioPrimes);
969
970
971
#elif defined (HAVE_LIBNETCDF)
  cdiSerialOpenFileCount = xcalloc(sizeof (cdiSerialOpenFileCount[0]),
                                   commInqSizeColl());
972
973
974
975
976
977
978
979
  namespaceSwitchSet(NSSWITCH_STREAM_OPEN_BACKEND,
                     NSSW_FUNC(cdiPioStreamCDFOpenWrap));
  namespaceSwitchSet(NSSWITCH_STREAM_CLOSE_BACKEND,
                     NSSW_FUNC(cdiPioStreamCDFCloseWrap));
  namespaceSwitchSet(NSSWITCH_CDF_DEF_TIMESTEP,
                     NSSW_FUNC(cdiPioCdfDefTimestep));
  namespaceSwitchSet(NSSWITCH_CDF_STREAM_SETUP,
                     NSSW_FUNC(cdiPioServerCdfDefVars));
980
#endif
981
982
983
  namespaceSwitchSet(NSSWITCH_FILE_WRITE,
                     NSSW_FUNC(cdiPioFileWrite));

Deike Kleberg's avatar
Deike Kleberg committed
984
  for ( ;; )
985
    {
Deike Kleberg's avatar
Deike Kleberg committed
986
      xmpi ( MPI_Probe ( MPI_ANY_SOURCE, MPI_ANY_TAG, commCalc, &status ));
987
      
Deike Kleberg's avatar
Deike Kleberg committed
988
989
      source = status.MPI_SOURCE;
      tag    = status.MPI_TAG;
990
      
Deike Kleberg's avatar
Deike Kleberg committed
991
      switch ( tag )
992
        {
993
994
995
996
997
998
999
        case FINALIZE:
          {
            int i;
            xdebugMsg(tag, source, nfinished);
            xmpi(MPI_Recv(&i, 1, MPI_INTEGER, source,
                          tag, commCalc, &status));
          }
1000
          xdebug("%s", "RECEIVED MESSAGE WITH TAG \"FINALIZE\"");
For faster browsing, not all history is shown. View entire blame