VCluster.hpp 28.5 KB
Newer Older
incardon's avatar
incardon committed
1
2
3
#ifndef VCLUSTER
#define VCLUSTER

4
#include "config.h"
incardon's avatar
incardon committed
5
#include <mpi.h>
6
#include "MPI_wrapper/MPI_util.hpp"
incardon's avatar
incardon committed
7
8
9
#include "VCluster_object.hpp"
#include "VCluster_object_array.hpp"
#include "Vector/map_vector.hpp"
10
11
12
#include "MPI_wrapper/MPI_IallreduceW.hpp"
#include "MPI_wrapper/MPI_IrecvW.hpp"
#include "MPI_wrapper/MPI_IsendW.hpp"
incardon's avatar
incardon committed
13
#include "MPI_wrapper/MPI_IAllGather.hpp"
incardon's avatar
incardon committed
14
#include <exception>
15
#include "Vector/map_vector.hpp"
16
17
#ifdef DEBUG
#include "util/check_no_pointers.hpp"
18
#include "util/util_debug.hpp"
19
#endif
incardon's avatar
incardon committed
20
#include "util/Vcluster_log.hpp"
incardon's avatar
incardon committed
21
22
23

#define MSG_LENGTH 1024
#define MSG_SEND_RECV 1025
24
#define SEND_SPARSE 4096
incardon's avatar
incardon committed
25
26
27
#define NONE 1
#define NEED_ALL_SIZE 2

28
29
#define SERIVCE_MESSAGE_TAG 16384
#define SEND_RECV_BASE 8192
incardon's avatar
incardon committed
30
#define GATHER_BASE 24576
incardon's avatar
incardon committed
31

Pietro Incardona's avatar
Pietro Incardona committed
32
// number of vcluster instances
incardon's avatar
incardon committed
33
extern size_t n_vcluster;
Pietro Incardona's avatar
Pietro Incardona committed
34
// Global MPI initialization
incardon's avatar
incardon committed
35
36
extern bool global_mpi_init;

incardon's avatar
incardon committed
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
///////////////////// Post functions /////////////

template<typename T> void assign(T * ptr1, T * ptr2)
{
	*ptr1 = *ptr2;
};

//////////////////////////////////////////////////

// temporal buffer for reductions
union red
{
	char c;
	unsigned char uc;
	short s;
	unsigned short us;
	int i;
	unsigned int ui;
	float f;
	double d;
};

/*! \brief Virtual Cluster exception
 *
 * This a a class that signal an exception on MPI_WaitAll
 *
 */

class exec_exception: public std::exception
{
  virtual const char* what() const throw()
  {
    return "One or more request has failed or reported an error";
  }
};

73
/*! \brief This class virtualize the cluster of PC as a set of processes that communicate
incardon's avatar
incardon committed
74
 *
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
 * At the moment it is an MPI-like interface, with a more type aware, and simple, interface.
 *  It also give some more complex communication functionalities like **Dynamic Sparse Data Exchange**
 *
 * Actually VCluster expose a Computation driven parallelism (MPI-like), with a plan of extending to
 * communication driven parallelism
 *
 * * In computation driven parallelism, the program compute than communicate to the other processors
 *
 * * In a communication driven parallelism, (Charm++ or HPX), the program receive messages, this receiving
 *   messages trigger computation
 *
 * ### An example of sending and receive plain buffers
 * \snippet VCluster_unit_test_util.hpp Send and receive plain buffer data
 * ### An example of sending vectors of primitives with (T=float,double,lont int,...)
 * \snippet VCluster_unit_test_util.hpp Sending and receiving primitives
 * ### An example of sending vectors of complexes object
 * \snippet VCluster_unit_test_util.hpp Send and receive vectors of complex
 * ### An example of gathering numbers from all processors
 * \snippet VCluster_unit_test_util.hpp allGather numbers
incardon's avatar
incardon committed
94
95
96
97
98
 *
 */

class Vcluster
{
incardon's avatar
incardon committed
99
100
	Vcluster_log log;

101
102
103
	//! NBX has a potential pitfall that must be addressed,
	//! NBX Send all the messages and probe for incoming messages,
	//! if there is an incoming message it receive it producing
104
	//! an acknowledge notification on the sending processor.
105
106
107
108
109
110
	//! When all the sends has been acknowledged, the processor call the MPI_Ibarrier
	//! when all the processors call MPI_Ibarrier all send has been received.
	//! While the processors are waiting for the MPI_Ibarrier to complete, all processors
	//! are still probing for incoming message, Unfortunately some processor
	//! can quit the MPI_Ibarrier before others and this mean that some
	//! processor can exit the probing status before others, these processors can in theory
111
112
113
114
	//! start new communications while the other processor are still in probing status producing
	//! a wrong send/recv association to
	//! resolve this problem an incremental NBX_cnt is used as message TAG to distinguish that the
	//! messages come from other send or subsequent NBX procedures
incardon's avatar
incardon committed
115
	size_t NBX_cnt;
116

incardon's avatar
incardon committed
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
	// temporal vector used for meta-communication
	// ( or meta-data before the real communication )
	openfpm::vector<size_t> proc_com;

	// vector that contain the scatter map (it is basically an array of one)
	openfpm::vector<int> map_scatter;

	// vector of MPI requests
	openfpm::vector<MPI_Request> req;

	// vector of MPI status
	openfpm::vector<MPI_Status> stat;

	// vector of functions to execute after all the request has been performed
	std::vector<int> post_exe;

	// Object array


	// Single objects

	// number of processes
	int size;
	// actual rank
	int rank;

	// number of processing unit per process
	int numPE = 1;

	/* This buffer is a temporal buffer for reductions
	 *
	 * MPI_Iallreduce does not accept recv and send buffer to be the same
	 * r is used to overcome this problem (is given as second parameter)
	 * after the execution the data is copied back
	 *
	 */
	std::vector<red> r;

155
156
157
158
159
160
161
162
163
164
165
166
167
168
	// vector of pointers of send buffers
	openfpm::vector<void *> ptr_send;

	// vector of the size of send buffers
	openfpm::vector<size_t> sz_send;

	// sending map
	openfpm::vector<size_t> map;

	// barrier request
	MPI_Request bar_req;
	// barrier status
	MPI_Status bar_stat;

incardon's avatar
incardon committed
169
170
171
172
173
public:

	// Finalize the MPI program
	~Vcluster()
	{
Pietro Incardona's avatar
Pietro Incardona committed
174
175
176
#ifdef SE_CLASS2
		check_delete(this);
#endif
Pietro Incardona's avatar
Pietro Incardona committed
177
		n_vcluster--;
incardon's avatar
incardon committed
178
179

		// if there are no other vcluster instances finalize
Pietro Incardona's avatar
Pietro Incardona committed
180
181
		if (n_vcluster == 0)
		{
182
183
184
185
			int already_finalised;

			MPI_Finalized(&already_finalised);
			if (!already_finalised)
Pietro Incardona's avatar
Pietro Incardona committed
186
			{
187
188
189
190
				if (MPI_Finalize() != 0)
				{
					std::cerr << __FILE__ << ":" << __LINE__  << " MPI_Finalize FAILED \n";
				}
Pietro Incardona's avatar
Pietro Incardona committed
191
			}
Pietro Incardona's avatar
Pietro Incardona committed
192
		}
incardon's avatar
incardon committed
193
194
	}

195
196
197
198
199
200
	/*! \brief Virtual cluster constructor
	 *
	 * \param argc pointer to arguments counts passed to the program
	 * \param argv pointer to arguments vector passed to the program
	 *
	 */
incardon's avatar
incardon committed
201
	Vcluster(int *argc, char ***argv)
incardon's avatar
incardon committed
202
	:NBX_cnt(0)
incardon's avatar
incardon committed
203
	{
Pietro Incardona's avatar
Pietro Incardona committed
204
205
206
207
#ifdef SE_CLASS2
		check_new(this,8,VCLUSTER_EVENT,PRJ_VCLUSTER);
#endif

Pietro Incardona's avatar
Pietro Incardona committed
208
		n_vcluster++;
incardon's avatar
incardon committed
209

210
211
212
		int already_initialised;
		MPI_Initialized(&already_initialised);

incardon's avatar
incardon committed
213
		// Check if MPI is already initialized
214
		if (!already_initialised)
incardon's avatar
incardon committed
215
		{
incardon's avatar
incardon committed
216

incardon's avatar
incardon committed
217
218
			MPI_Init(argc,argv);
		}
incardon's avatar
incardon committed
219

220
221
		// Get the total number of process
		// and the rank of this process
incardon's avatar
incardon committed
222
223
224
225

		MPI_Comm_size(MPI_COMM_WORLD, &size);
		MPI_Comm_rank(MPI_COMM_WORLD, &rank);

226
#ifdef SE_CLASS2
incardon's avatar
incardon committed
227
228
229
			process_v_cl = rank;
#endif

230
		// create and fill map scatter with one
incardon's avatar
incardon committed
231
232
233
234
235
236
		map_scatter.resize(size);

		for (size_t i = 0 ; i < map_scatter.size() ; i++)
		{
			map_scatter.get(i) = 1;
		}
incardon's avatar
incardon committed
237
238
239

		// open the log file
		log.openLog(rank);
incardon's avatar
incardon committed
240
241
	}

242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
#ifdef DEBUG

	/*! \brief Check for wrong types
	 *
	 * In general we do not know if a type T make sense to be sent or not, but if it has pointer
	 * inside it does not. This function check if the basic type T has a method called noPointers,
	 * This function in general notify if T has internally pointers. If T has pointer an error
	 * is printed, is T does not have the method a WARNING is printed
	 *
	 * \tparam T type to check
	 *
	 */
	template<typename T> void checkType()
	{
		// if T is a primitive like int, long int, float, double, ... make sense
		// (pointers, l-references and r-references are not fundamentals)
		if (std::is_fundamental<T>::value == true)
			return;

		// if it is a pointer make no sense
		if (std::is_pointer<T>::value == true)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " is a pointer, sending pointers values has no sense\n";

		// if it is an l-value reference make no send
		if (std::is_lvalue_reference<T>::value == true)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " is a pointer, sending pointers values has no sense\n";

		// if it is an r-value reference make no send
		if (std::is_rvalue_reference<T>::value == true)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " is a pointer, sending pointers values has no sense\n";

		// ... if not, check that T has a method called noPointers
		switch (check_no_pointers<T>::value())
		{
			case PNP::UNKNOWN:
			{
incardon's avatar
incardon committed
278
				std::cerr << "Warning: " << __FILE__ << ":" << __LINE__ << " impossible to check the type " << demangle(typeid(T).name()) << " please consider to add a static method \"static bool noPointers()\" \n" ;
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
				break;
			}
			case PNP::POINTERS:
			{
				std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " has pointers inside, sending pointers values has no sense\n";
				break;
			}
			default:
			{

			}
		}
	}

#endif

incardon's avatar
incardon committed
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
	//! Get the total number of processing units
	size_t getProcessingUnits()
	{
		return size*numPE;
	}

	//! Get the process unit id
	size_t getProcessUnitID()
	{
		return rank;
	}

	/*! \brief Allocate a set of objects
	 *
	 * \tparam obj
	 * \param n number of object
	 *
	 * \return an object representing an array of objects
	 *
	 */
	template <typename obj> Vcluster_object_array<obj> allocate(size_t n)
	{
		// Vcluster object array
		Vcluster_object_array<obj> vo;

		// resize the array
		vo.resize(n);

		// Create the object on memory and return a Vcluster_object_array
		return vo;
	}

327
	/*! \brief Sum the numbers across all processors and get the result
incardon's avatar
incardon committed
328
329
330
331
332
	 *
	 * \param num to reduce, input and output
	 *
	 */

333
	template<typename T> void sum(T & num)
incardon's avatar
incardon committed
334
	{
335
336
337
338
#ifdef DEBUG
		checkType<T>();
#endif

incardon's avatar
incardon committed
339
340
341
342
343
344
345
346
347
		// reduce over MPI

		// Create one request
		req.add();

		// reduce
		MPI_IallreduceW<T>::reduce(num,MPI_SUM,req.last());
	}

348
	/*! \brief Get the maximum number across all processors (or reduction with infinity norm)
incardon's avatar
incardon committed
349
350
351
352
353
354
	 *
	 * \param num to reduce
	 *
	 */
	template<typename T> void max(T & num)
	{
355
356
357
#ifdef DEBUG
		checkType<T>();
#endif
incardon's avatar
incardon committed
358
359
360
361
362
363
364
365
366
		// reduce over MPI

		// Create one request
		req.add();

		// reduce
		MPI_IallreduceW<T>::reduce(num,MPI_MAX,req.last());
	}

367
	/*! \brief Send and receive multiple messages
368
	 *
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
	 * It send multiple messages to a set of processors the and receive
	 * multiple messages from another set of processors, all the processor must call this
	 * function, NBX is more performant than PCX with more processors (1000+)
	 *
	 * suppose the following situation the calling processor want to communicate
	 * * 2 vector of 100 integers to processor 1
	 * * 1 vector of 50 integers to processor 6
	 * * 1 vector of 48 integers to processor 7
	 * * 1 vector of 70 integers to processor 8
	 *
	 * \param prc list of processors you should communicate with [1,1,6,7,8]
	 *
	 * \param v vector containing the data to send [v=vector<vector<int>>, v.size()=4, T=vector<int>], T at the moment
	 *          is only tested for vectors of 0 or more generic elements (without pointers)
	 *
	 * \param msg_alloc This is a call-back with the purpose to allocate space
	 *        for the incoming messages and give back a valid pointer, supposing that this call-back has been triggered by
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
	 *        in the call-back in order:
	 *        * message size required to receive the message (100)
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
	 *        * processor id (5)
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
394
	 *           every time message_alloc is called)
395
	 *        * void pointer, parameter for additional data to pass to the call-back
396
	 *
397
	 * \param opt options, only NONE supported
398
399
400
401
	 *
	 */
	template<typename T> void sendrecvMultipleMessagesNBX(openfpm::vector< size_t > & prc, openfpm::vector< T > & data, void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,void *), void * ptr_arg, long int opt=NONE)
	{
402
403
404
#ifdef DEBUG
		checkType<typename T::value_type>();
#endif
405
406
407
408
409
410
411
412
413
414
415
416
417
		// resize the pointer list
		ptr_send.resize(prc.size());
		sz_send.resize(prc.size());

		for (size_t i = 0 ; i < prc.size() ; i++)
		{
			ptr_send.get(i) = data.get(i).getPointer();
			sz_send.get(i) = data.get(i).size() * sizeof(typename T::value_type);
		}

		sendrecvMultipleMessagesNBX(prc.size(),(size_t *)sz_send.getPointer(),(size_t *)prc.getPointer(),(void **)ptr_send.getPointer(),msg_alloc,ptr_arg,opt);
	}

incardon's avatar
incardon committed
418
419
	/*! \brief Send and receive multiple messages
	 *
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
	 * It sends multiple messages to a set of processors and receives
	 * multiple messages from another set of processors, all the processor must call this
	 * function, NBX is more performant than PCX with more processors (1000+)
	 *
	 * suppose the following situation the calling processor want to communicate
	 * * 2 vector of 100 integers to processor 1
	 * * 1 vector of 50 integers to processor 6
	 * * 1 vector of 48 integers to processor 7
	 * * 1 vector of 70 integers to processor 8
	 *
	 * \param prc list of processors you should communicate with [1,1,6,7,8]
	 *
	 * \param v vector containing the data to send [v=vector<vector<int>>, v.size()=4, T=vector<int>], T at the moment
	 *          is only tested for vectors of 0 or more generic elements (without pointers)
	 *
	 * \param msg_alloc This is a call-back with the purpose to allocate space
	 *        for the incoming messages and give back a valid pointer, supposing that this call-back has been triggered by
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
	 *        in the call-back in order:
	 *        * message size required to receive the message [100]
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
	 *        * processor id [5]
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
445
	 *           every time message_alloc is called)
446
	 *        * void pointer, parameter for additional data to pass to the call-back
incardon's avatar
incardon committed
447
	 *
448
	 * \param opt options, only NONE supported
incardon's avatar
incardon committed
449
450
	 *
	 */
451
	template<typename T> void sendrecvMultipleMessagesPCX(openfpm::vector< size_t > & prc, openfpm::vector< T > & data, void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,void *), void * ptr_arg, long int opt=NONE)
incardon's avatar
incardon committed
452
	{
453
454
455
#ifdef DEBUG
		checkType<typename T::value_type>();
#endif
incardon's avatar
incardon committed
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
		// resize map with the number of processors
		map.resize(size);

		// reset the sending buffer
		map.fill(0);

		// create sending map
		for (size_t i = 0 ; i < prc.size() ; i++)
		{map.get(prc.get(i)) = 1;}

		// resize the pointer list
		ptr_send.resize(prc.size());
		sz_send.resize(prc.size());

		for (size_t i = 0 ; i < prc.size() ; i++)
		{
			ptr_send.get(i) = data.get(i).getPointer();
			sz_send.get(i) = data.get(i).size() * sizeof(typename T::value_type);
		}

476
477
478
		sendrecvMultipleMessagesPCX(prc.size(),(size_t *)map.getPointer(),(size_t *)sz_send.getPointer(),(size_t *)prc.getPointer(),(void **)ptr_send.getPointer(),msg_alloc,ptr_arg,opt);
	}

479
	/*! \brief Send and receive multiple messages
480
	 *
481
482
483
	 * It send multiple messages to a set of processors the and receive
	 * multiple messages from another set of processors, all the processor must call this
	 * function, NBX is more performant than PCX with more processors (1000+)
484
	 *
485
486
487
488
489
	 * suppose the following situation the calling processor want to communicate
	 * * 2 messages of size 100 byte to processor 1
	 * * 1 message of size 50 byte to processor 6
	 * * 1 message of size 48 byte to processor 7
	 * * 1 message of size 70 byte to processor 8
490
	 *
491
	 * \param n_send number of send for this processor [4]
492
493
	 *
	 * \param prc list of processor with which it should communicate
494
	 *        [1,1,6,7,8]
495
	 *
496
497
498
499
	 * \param sz the array contain the size of the message for each processor
	 *        (zeros must not be presents) [100,100,50,48,70]
	 *
	 * \param ptr array that contain the pointers to the message to send
500
501
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
502
503
504
	 *        for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
505
	 *        in the call-back are in order:
506
507
508
509
510
	 *        * message size required to receive the message [100]
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
	 *        * processor id [5]
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
511
	 *           every time message_alloc is called)
512
	 *        * void pointer, parameter for additional data to pass to the call-back
513
514
515
516
517
518
	 *
	 * \param opt options, NONE (ignored in this moment)
	 *
	 */
	void sendrecvMultipleMessagesNBX(size_t n_send , size_t sz[], size_t prc[] , void * ptr[], void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,void *), void * ptr_arg, long int opt)
	{
519
520
521
522
523
		if (stat.size() != 0 || req.size() != 0)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " this function must be called when no other requests are in progress \n";


		stat.clear();
524
525
526
527
528
		req.clear();
		// Do MPI_Issend

		for (size_t i = 0 ; i < n_send ; i++)
		{
incardon's avatar
incardon committed
529
530
531
			if (sz[i] != 0)
			{
				req.add();
532
				MPI_SAFE_CALL(MPI_Issend(ptr[i], sz[i], MPI_BYTE, prc[i], SEND_SPARSE + NBX_cnt, MPI_COMM_WORLD,&req.last()));
incardon's avatar
incardon committed
533
				log.logSend(prc[i]);
incardon's avatar
incardon committed
534
			}
535
536
537
538
539
540
541
		}

		size_t rid = 0;
		int flag = false;

		bool reached_bar_req = false;

incardon's avatar
incardon committed
542
543
		log.start(10);

544
545
546
		// Wait that all the send are acknowledge
		do
		{
incardon's avatar
incardon committed
547

548
549
550
551
552
			// flag that notify that this processor reach the barrier
			// Barrier request

			MPI_Status stat_t;
			int stat = false;
553
			MPI_SAFE_CALL(MPI_Iprobe(MPI_ANY_SOURCE,SEND_SPARSE + NBX_cnt,MPI_COMM_WORLD,&stat,&stat_t));
554

555
			// If I have an incoming message and is related to this NBX communication
556
557
558
559
560
561
562
563
564
			if (stat == true)
			{
				// Get the message size
				int msize;
				MPI_SAFE_CALL(MPI_Get_count(&stat_t,MPI_BYTE,&msize));

				// Get the pointer to receive the message
				void * ptr = msg_alloc(msize,0,0,stat_t.MPI_SOURCE,rid,ptr_arg);

incardon's avatar
incardon committed
565
566
567
				// Log the receiving request
				log.logRecv(stat_t);

568
569
				rid++;

570
				MPI_SAFE_CALL(MPI_Recv(ptr,msize,MPI_BYTE,stat_t.MPI_SOURCE,SEND_SPARSE+NBX_cnt,MPI_COMM_WORLD,&stat_t));
571
572
573
574
575
576
577
			}

			// Check the status of all the MPI_issend and call the barrier if finished

			if (reached_bar_req == false)
			{
				int flag = false;
incardon's avatar
incardon committed
578
579
580
581
				if (req.size() != 0)
				{MPI_SAFE_CALL(MPI_Testall(req.size(),&req.get(0),&flag,MPI_STATUSES_IGNORE));}
				else
					flag = true;
582
583
584
585
586
587

				// If all send has been completed
				if (flag == true)
				{MPI_SAFE_CALL(MPI_Ibarrier(MPI_COMM_WORLD,&bar_req));reached_bar_req = true;}
			}

incardon's avatar
incardon committed
588
			// Check if all processor reached the async barrier
589
			if (reached_bar_req)
incardon's avatar
incardon committed
590
			{MPI_SAFE_CALL(MPI_Test(&bar_req,&flag,&bar_stat))};
incardon's avatar
incardon committed
591
592

			// produce a report if communication get stuck
incardon's avatar
incardon committed
593
			log.NBXreport(NBX_cnt,req,reached_bar_req,bar_stat);
incardon's avatar
incardon committed
594

595
596
597
598
599
600
		} while (flag == false);

		// Remove the executed request

		req.clear();
		stat.clear();
incardon's avatar
incardon committed
601
		log.clear();
602
603
604

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
incardon's avatar
incardon committed
605
606
607
608
	}

	/*! \brief Send and receive multiple messages
	 *
609
610
611
	 * It send multiple messages to a set of processors the and receive
	 * multiple messages from another set of processors, all the processor must call this
	 * function, NBX is more performant than PCX with more processors (1000+)
incardon's avatar
incardon committed
612
	 *
613
614
615
616
617
	 * suppose the following situation the calling processor want to communicate
	 * * 2 messages of size 100 byte to processor 1
	 * * 1 message of size 50 byte to processor 6
	 * * 1 message of size 48 byte to processor 7
	 * * 1 message of size 70 byte to processor 8
incardon's avatar
incardon committed
618
	 *
619
	 * \param n_send number of send for this processor [4]
incardon's avatar
incardon committed
620
	 *
621
622
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
incardon's avatar
incardon committed
623
	 *
624
625
626
	 * \param map array containing an array of the number of messages for
	 *        each processor the colling processor want to communicate
	 *        [0 2 0 0 0 0 1 1 1]
incardon's avatar
incardon committed
627
	 *
628
629
	 * \param sz the array contain the size of the message for each processor
	 *        (zeros must not be presents) [100,100,50,48,70]
incardon's avatar
incardon committed
630
	 *
631
	 * \param ptr array that contain the pointers to the message to send
incardon's avatar
incardon committed
632
633
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
634
635
636
637
638
639
640
641
642
643
644
	 *        for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
	 *        in the call-back are in order:
	 *        * message size required to receive the message [100]
	 *        * total message size to receive from all the processors
	 *        * the total number of processor want to communicate with the calling processor
	 *        * processor id [5]
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
	 *           every time message_alloc is called)
	 *        * void pointer, parameter for additional data to pass to the call-back
incardon's avatar
incardon committed
645
	 *
646
	 * \param opt options, NONE (ignored in this moment)
incardon's avatar
incardon committed
647
648
	 *
	 */
649
	void sendrecvMultipleMessagesPCX(size_t n_send, size_t * map, size_t sz[], size_t prc[] , void * ptr[], void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,void *), void * ptr_arg, long int opt)
incardon's avatar
incardon committed
650
	{
651
652
653
654
		if (stat.size() != 0 || req.size() != 0)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " this function must be called when no other requests are in progress \n";

		stat.clear();
incardon's avatar
incardon committed
655
		req.clear();
656

incardon's avatar
incardon committed
657
		req.add();
incardon's avatar
incardon committed
658
		stat.add();
incardon's avatar
incardon committed
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

		proc_com.resize(1);

		MPI_SAFE_CALL(MPI_Ireduce_scatter(map,&proc_com.get(0),&map_scatter.get(0),MPI_UNSIGNED_LONG,MPI_SUM,MPI_COMM_WORLD,&req.last()));
		MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));

		// Remove the executed request

		req.clear();
		stat.clear();

		// Allocate the temporal buffer to store the message size for each processor

		size_t n_proc_com = proc_com.get(0);
		proc_com.resize(n_proc_com * 2);

		// queue n_proc_com MPI_Irecv with source ANY_SOURCE to get
		// the message length from each processor and
		//  send the message length to each processor

		for (size_t i = 0 ; i < n_proc_com ; i++)
		{
			req.add();
			MPI_SAFE_CALL(MPI_Irecv(&proc_com.get(i),1,MPI_UNSIGNED_LONG,MPI_ANY_SOURCE,MSG_LENGTH,MPI_COMM_WORLD,&req.last()));
		}

		for (size_t i = 0 ; i < n_send ; i++)
		{
			req.add();
			MPI_SAFE_CALL(MPI_Isend(&sz[i],1,MPI_UNSIGNED_LONG,prc[i],MSG_LENGTH,MPI_COMM_WORLD,&req.last()));
		}

		stat.resize(req.size());
692
		if (req.size() != 0) {MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));}
incardon's avatar
incardon committed
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718

		// Use proc_com to get the processor id that try to communicate

		for (size_t i = 0 ; i < n_proc_com ; i++)
		{
			proc_com.get(n_proc_com+i) = stat.get(i).MPI_SOURCE;
		}

		// Remove the executed request

		req.clear();
		stat.clear();

		// Calculate the total size of the message

		size_t total_msg = 0;

		for (size_t i = 0 ; i < n_proc_com ; i++)
		{
			total_msg += proc_com.get(i);
		}

		// Receive the message

		for (size_t i = 0 ; i < n_proc_com ; i++)
		{
incardon's avatar
incardon committed
719
			void * ptr = msg_alloc(proc_com.get(i),total_msg,n_proc_com,proc_com.get(n_proc_com+i),i,ptr_arg);
incardon's avatar
incardon committed
720
721
722
723
724
725
726
727
728
729
730
731
732
733

			req.add();
			MPI_SAFE_CALL(MPI_Irecv(ptr,proc_com.get(i),MPI_BYTE,proc_com.get(i+n_proc_com),MSG_SEND_RECV,MPI_COMM_WORLD,&req.last()));
		}

		// Send all the messages this processor must send

		for (size_t i = 0 ; i < n_send ; i++)
		{
			req.add();
			MPI_SAFE_CALL(MPI_Isend(ptr[i],sz[i],MPI_BYTE,prc[i],MSG_SEND_RECV,MPI_COMM_WORLD,&req.last()));
		}

		stat.resize(req.size());
734
		if (req.size() != 0) {MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));}
incardon's avatar
incardon committed
735
736
737
738
739
740
741

		// Remove the executed request

		req.clear();
		stat.clear();
	}

742
743
744
745
746
747
	/*! \brief Send data to a processor
	 *
	 * \warning In order to avoid deadlock every send must be coupled with a recv
	 *          in case you want to send data without knowledge from the other side
	 *          consider to use sendRecvMultipleMessages
	 *
incardon's avatar
incardon committed
748
749
750
751
752
753
754
755
756
757
758
759
	 * \warning operation is asynchronous execute must be called to ensure they are executed
	 *
	 * \see sendRecvMultipleMessages
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param mem buffer with the data to send
	 * \param sz size
	 *
	 * \return true if succeed false otherwise
	 *
	 */
760
	bool send(size_t proc, size_t tag, const void * mem, size_t sz)
incardon's avatar
incardon committed
761
762
763
764
765
766
767
768
769
770
771
772
	{
		// send over MPI

		// Create one request
		req.add();

		// send
		MPI_IsendWB::send(proc,SEND_RECV_BASE + tag,mem,sz,req.last());

		return true;
	}

incardon's avatar
incardon committed
773

incardon's avatar
incardon committed
774
775
776
777
778
779
780
	/*! \brief Send data to a processor
	 *
	 * \warning In order to avoid deadlock every send must be coupled with a recv
	 *          in case you want to send data without knowledge from the other side
	 *          consider to use sendRecvMultipleMessages
	 *
	 * \warning operation is asynchronous execute must be called to ensure they are executed
781
782
783
784
785
786
787
	 *
	 * \see sendRecvMultipleMessages
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
	 *
788
789
	 * \return true if succeed false otherwise
	 *
790
	 */
791
	template<typename T, typename Mem, typename gr> bool send(size_t proc, size_t tag, openfpm::vector<T,Mem,gr> & v)
792
793
794
795
796
797
798
799
800
801
	{
#ifdef DEBUG
		checkType<T>();
#endif

		// send over MPI

		// Create one request
		req.add();

802
		// send
803
		MPI_IsendW<T,Mem,gr>::send(proc,SEND_RECV_BASE + tag,v,req.last());
804
805

		return true;
806
807
808
809
810
811
	}

	/*! \brief Recv data from a processor
	 *
	 * \warning In order to avoid deadlock every recv must be coupled with a send
	 *          in case you want to send data without knowledge from the other side
812
	 *          consider to use sendrecvMultipleMessagesPCX or sendrecvMultipleMessagesNBX
813
	 *
incardon's avatar
incardon committed
814
	 * \warning operation is asynchronous execute must be called to ensure they are executed
815
	 *
816
	 * \see sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
817
818
819
820
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
incardon's avatar
incardon committed
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
	 * \param sz size of the buffer
	 *
	 * \return true if succeed false otherwise
	 *
	 */
	bool recv(size_t proc, size_t tag, void * v, size_t sz)
	{
		// recv over MPI

		// Create one request
		req.add();

		// receive
		MPI_IrecvWB::recv(proc,SEND_RECV_BASE + tag,v,sz,req.last());

		return true;
	}

incardon's avatar
incardon committed
839
840
841
842
    /*! \brief Recv data from a processor
     *
     * \warning In order to avoid deadlock every recv must be coupled with a send
     *          in case you want to send data without knowledge from the other side
843
     *          consider to use sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
844
845
846
     *
     * \warning operation is asynchronous execute must be called to ensure they are executed
     *
847
     * \see sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
     *
     * \param proc processor id
     * \param tag id
     * \param v vector to send
     *
     * \return true if succeed false otherwise
     *
     */
    template<typename T, typename Mem, typename gr> bool recv(size_t proc, size_t tag, openfpm::vector<T,Mem,gr> & v)
    {
#ifdef DEBUG
            checkType<T>();
#endif

            // recv over MPI

            // Create one request
            req.add();

            // receive
            MPI_IrecvW<T>::recv(proc,SEND_RECV_BASE + tag,v,req.last());

            return true;
    }

	/*! \brief Gather the data from all processors
874
875
	 *
	 * send a primitive data T receive the same primitive T from all the other processors
incardon's avatar
incardon committed
876
877
878
879
	 *
	 * \warning operation is asynchronous execute must be called to ensure they are executed
	 *
	 * \param tag id
880
	 * \param v vector to receive (automaticaly resized)
incardon's avatar
incardon committed
881
	 * \param send data to send
882
	 *
883
884
	 * \return true if succeed false otherwise
	 *
885
	 */
incardon's avatar
incardon committed
886
	template<typename T, typename Mem, typename gr> bool allGather(T & send, openfpm::vector<T,Mem,gr> & v)
887
888
889
890
891
892
893
894
	{
#ifdef DEBUG
		checkType<T>();
#endif

		// Create one request
		req.add();

incardon's avatar
incardon committed
895
896
897
		// Number of processors
		v.resize(getProcessingUnits());

898
		// gather
incardon's avatar
incardon committed
899
		MPI_IAllGatherW<T>::gather(&send,1,v.getPointer(),1,req.last());
900
901

		return true;
902
903
904
	}

	/*! \brief Execute all the requests
incardon's avatar
incardon committed
905
906
907
908
909
910
	 *
	 */
	void execute()
	{
		int err = 0;

911
912
913
914
		// if req == 0 return
		if (req.size() == 0)
			return;

incardon's avatar
incardon committed
915
		// Wait for all the requests
916
		stat.resize(req.size());
incardon's avatar
incardon committed
917
918
919
920
921
922
923
924
925
926
927
928
929
930
		err = MPI_Waitall(req.size(),&req.get(0),&stat.get(0));

		// MPI error get the message and abort MPI
		if (err != MPI_SUCCESS)
		{
			char * error_msg = NULL;
			int len;
			MPI_Error_string(err,error_msg,&len);

			std::cerr << "Error MPI rank " << rank << ": " << error_msg << "\n";

			MPI_Abort(MPI_COMM_WORLD,1);
		}

931
		// Remove executed request and status
incardon's avatar
incardon committed
932
933
934
		req.clear();
		stat.clear();
	}
935
936


incardon's avatar
incardon committed
937
938
};

Pietro Incardona's avatar
Pietro Incardona committed
939
// Function to initialize the global VCluster //
incardon's avatar
incardon committed
940
941
942

extern Vcluster * global_v_cluster;

Pietro Incardona's avatar
Pietro Incardona committed
943
944
945
946
947
948
949

/*! \brief Initialize a global instance of Runtime Virtual Cluster Machine
 *
 * Initialize a global instance of Runtime Virtual Cluster Machine
 *
 */

Pietro Incardona's avatar
Pietro Incardona committed
950
static inline void init_global_v_cluster(int *argc, char ***argv)
Pietro Incardona's avatar
Pietro Incardona committed
951
952
953
954
955
{
	if (global_v_cluster == NULL)
		global_v_cluster = new Vcluster(argc,argv);
}

Pietro Incardona's avatar
Pietro Incardona committed
956
static inline void delete_global_v_cluster()
Pietro Incardona's avatar
Pietro Incardona committed
957
958
959
960
{
	delete global_v_cluster;
}

incardon's avatar
incardon committed
961
962
#endif