VCluster.hpp 28.9 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
#include "VCluster_object.hpp"
#include "Vector/map_vector.hpp"
9 10 11
#include "MPI_wrapper/MPI_IallreduceW.hpp"
#include "MPI_wrapper/MPI_IrecvW.hpp"
#include "MPI_wrapper/MPI_IsendW.hpp"
incardon's avatar
incardon committed
12
#include "MPI_wrapper/MPI_IAllGather.hpp"
incardon's avatar
incardon committed
13
#include <exception>
14
#include "Vector/map_vector.hpp"
15 16
#ifdef DEBUG
#include "util/check_no_pointers.hpp"
17
#include "util/util_debug.hpp"
18
#endif
incardon's avatar
incardon committed
19
#include "util/Vcluster_log.hpp"
20
#include "memory/BHeapMemory.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
	// 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;

164
	// Receive buffers
165
	openfpm::vector<BHeapMemory> recv_buf;
166

167 168 169 170 171
	// barrier request
	MPI_Request bar_req;
	// barrier status
	MPI_Status bar_stat;

incardon's avatar
incardon committed
172 173 174 175 176
public:

	// Finalize the MPI program
	~Vcluster()
	{
Pietro Incardona's avatar
Pietro Incardona committed
177 178 179
#ifdef SE_CLASS2
		check_delete(this);
#endif
incardon's avatar
incardon committed
180 181 182 183
		n_vcluster--;

		// if there are no other vcluster instances finalize
		if (n_vcluster == 0)
Pietro Incardona's avatar
Pietro Incardona committed
184
		{
185 186 187 188
			int already_finalised;

			MPI_Finalized(&already_finalised);
			if (!already_finalised)
Pietro Incardona's avatar
Pietro Incardona committed
189
			{
190 191 192 193
				if (MPI_Finalize() != 0)
				{
					std::cerr << __FILE__ << ":" << __LINE__  << " MPI_Finalize FAILED \n";
				}
Pietro Incardona's avatar
Pietro Incardona committed
194
			}
Pietro Incardona's avatar
Pietro Incardona committed
195
		}
incardon's avatar
incardon committed
196 197
	}

198 199 200 201 202 203
	/*! \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
204
	Vcluster(int *argc, char ***argv)
incardon's avatar
incardon committed
205
	:NBX_cnt(0)
incardon's avatar
incardon committed
206
	{
Pietro Incardona's avatar
Pietro Incardona committed
207 208 209 210
#ifdef SE_CLASS2
		check_new(this,8,VCLUSTER_EVENT,PRJ_VCLUSTER);
#endif

incardon's avatar
incardon committed
211 212
		n_vcluster++;

213 214 215
		int already_initialised;
		MPI_Initialized(&already_initialised);

incardon's avatar
incardon committed
216
		// Check if MPI is already initialized
217
		if (!already_initialised)
incardon's avatar
incardon committed
218
		{
incardon's avatar
incardon committed
219

incardon's avatar
incardon committed
220 221
			MPI_Init(argc,argv);
		}
incardon's avatar
incardon committed
222

223 224
		// Get the total number of process
		// and the rank of this process
incardon's avatar
incardon committed
225 226 227 228

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

229
#ifdef SE_CLASS2
incardon's avatar
incardon committed
230 231 232
			process_v_cl = rank;
#endif

233
		// create and fill map scatter with one
incardon's avatar
incardon committed
234 235 236 237 238 239
		map_scatter.resize(size);

		for (size_t i = 0 ; i < map_scatter.size() ; i++)
		{
			map_scatter.get(i) = 1;
		}
incardon's avatar
incardon committed
240 241 242

		// open the log file
		log.openLog(rank);
243 244 245 246

		// Initialize bar_req
		bar_req = MPI_Request();
		bar_stat = MPI_Status();
incardon's avatar
incardon committed
247 248
	}

249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
#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
285
				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" ;
286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
				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

Pietro Incardona's avatar
Pietro Incardona committed
302 303 304 305 306 307 308 309 310 311
	/*! \brief Get the MPI_Communicator (or processor group) this VCluster is using
	 *
	 * \return MPI comunicator
	 *
	 */
	MPI_Comm getMPIComm()
	{
		return MPI_COMM_WORLD;
	}

incardon's avatar
incardon committed
312 313 314 315 316 317 318 319 320 321 322 323
	//! Get the total number of processing units
	size_t getProcessingUnits()
	{
		return size*numPE;
	}

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

324
	/*! \brief Sum the numbers across all processors and get the result
incardon's avatar
incardon committed
325 326 327 328 329
	 *
	 * \param num to reduce, input and output
	 *
	 */

330
	template<typename T> void sum(T & num)
incardon's avatar
incardon committed
331
	{
332 333 334 335
#ifdef DEBUG
		checkType<T>();
#endif

incardon's avatar
incardon committed
336 337 338 339 340 341 342 343 344
		// reduce over MPI

		// Create one request
		req.add();

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

345
	/*! \brief Get the maximum number across all processors (or reduction with infinity norm)
incardon's avatar
incardon committed
346 347 348 349 350 351
	 *
	 * \param num to reduce
	 *
	 */
	template<typename T> void max(T & num)
	{
352 353 354
#ifdef DEBUG
		checkType<T>();
#endif
incardon's avatar
incardon committed
355 356 357 358 359 360 361 362 363
		// reduce over MPI

		// Create one request
		req.add();

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

tonynsyde's avatar
tonynsyde committed
364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383
	/*! \brief Get the minimum number across all processors (or reduction with insinity norm)
	 *
	 * \param num to reduce
	 *
	 */

	template<typename T> void min(T & num)
	{
#ifdef DEBUG
		checkType<T>();
#endif
		// reduce over MPI

		// Create one request
		req.add();

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

384

385
	/*! \brief Send and receive multiple messages
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
	 * 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
412
	 *           every time message_alloc is called)
413
	 *        * void pointer, parameter for additional data to pass to the call-back
414
	 *
415
	 * \param opt options, only NONE supported
416 417 418 419
	 *
	 */
	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)
	{
420 421 422
#ifdef DEBUG
		checkType<typename T::value_type>();
#endif
423 424 425 426 427 428 429 430 431 432 433 434 435
		// 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
436 437
	/*! \brief Send and receive multiple messages
	 *
438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462
	 * 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
463
	 *           every time message_alloc is called)
464
	 *        * void pointer, parameter for additional data to pass to the call-back
incardon's avatar
incardon committed
465
	 *
466
	 * \param opt options, only NONE supported
incardon's avatar
incardon committed
467 468
	 *
	 */
469
	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
470
	{
471 472 473
#ifdef DEBUG
		checkType<typename T::value_type>();
#endif
incardon's avatar
incardon committed
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493
		// 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);
		}

494 495 496
		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);
	}

497
	/*! \brief Send and receive multiple messages
498
	 *
499 500 501
	 * 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+)
502
	 *
503 504 505 506 507
	 * 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
508
	 *
509
	 * \param n_send number of send for this processor [4]
510 511
	 *
	 * \param prc list of processor with which it should communicate
512
	 *        [1,1,6,7,8]
513
	 *
514 515 516 517
	 * \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
518 519
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
520 521 522
	 *        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
523
	 *        in the call-back are in order:
524 525 526 527 528
	 *        * 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
529
	 *           every time message_alloc is called)
530
	 *        * void pointer, parameter for additional data to pass to the call-back
531 532 533 534
	 *
	 * \param opt options, NONE (ignored in this moment)
	 *
	 */
535
	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 = NONE)
536
	{
537 538 539 540 541
		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();
542 543 544 545 546
		req.clear();
		// Do MPI_Issend

		for (size_t i = 0 ; i < n_send ; i++)
		{
incardon's avatar
incardon committed
547 548 549
			if (sz[i] != 0)
			{
				req.add();
550 551 552 553 554

#ifdef SE_CLASS2
				check_valid(ptr[i],sz[i]);
#endif

555
				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
556
				log.logSend(prc[i]);
incardon's avatar
incardon committed
557
			}
558 559 560 561 562 563 564
		}

		size_t rid = 0;
		int flag = false;

		bool reached_bar_req = false;

incardon's avatar
incardon committed
565 566
		log.start(10);

567 568 569
		// Wait that all the send are acknowledge
		do
		{
incardon's avatar
incardon committed
570

571 572 573 574 575
			// flag that notify that this processor reach the barrier
			// Barrier request

			MPI_Status stat_t;
			int stat = false;
576
			MPI_SAFE_CALL(MPI_Iprobe(MPI_ANY_SOURCE,SEND_SPARSE + NBX_cnt,MPI_COMM_WORLD,&stat,&stat_t));
577

578
			// If I have an incoming message and is related to this NBX communication
579 580 581 582 583 584 585 586 587
			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
588 589 590
				// Log the receiving request
				log.logRecv(stat_t);

591 592
				rid++;

593 594 595 596
				// Check the pointer
#ifdef SE_CLASS2
				check_valid(ptr,msize);
#endif
597
				MPI_SAFE_CALL(MPI_Recv(ptr,msize,MPI_BYTE,stat_t.MPI_SOURCE,SEND_SPARSE+NBX_cnt,MPI_COMM_WORLD,&stat_t));
598 599 600 601

#ifdef SE_CLASS2
				check_valid(ptr,msize);
#endif
602 603 604 605 606 607 608
			}

			// 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
609 610 611 612
				if (req.size() != 0)
				{MPI_SAFE_CALL(MPI_Testall(req.size(),&req.get(0),&flag,MPI_STATUSES_IGNORE));}
				else
					flag = true;
613 614 615 616 617 618

				// 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
619
			// Check if all processor reached the async barrier
620
			if (reached_bar_req)
incardon's avatar
incardon committed
621
			{MPI_SAFE_CALL(MPI_Test(&bar_req,&flag,&bar_stat))};
incardon's avatar
incardon committed
622 623

			// produce a report if communication get stuck
incardon's avatar
incardon committed
624
			log.NBXreport(NBX_cnt,req,reached_bar_req,bar_stat);
incardon's avatar
incardon committed
625

626 627 628 629 630 631
		} while (flag == false);

		// Remove the executed request

		req.clear();
		stat.clear();
incardon's avatar
incardon committed
632
		log.clear();
633 634 635

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
incardon's avatar
incardon committed
636 637 638 639
	}

	/*! \brief Send and receive multiple messages
	 *
640 641 642
	 * 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
643
	 *
644 645 646 647 648
	 * 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
649
	 *
650
	 * \param n_send number of send for this processor [4]
incardon's avatar
incardon committed
651
	 *
652 653
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
incardon's avatar
incardon committed
654
	 *
655 656 657
	 * \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
658
	 *
659 660
	 * \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
661
	 *
662
	 * \param ptr array that contain the pointers to the message to send
incardon's avatar
incardon committed
663 664
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
665 666 667 668 669 670 671 672 673 674 675
	 *        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
676
	 *
677
	 * \param opt options, NONE (ignored in this moment)
incardon's avatar
incardon committed
678 679
	 *
	 */
680
	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
681
	{
682 683 684 685
		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
686
		req.clear();
687

incardon's avatar
incardon committed
688
		req.add();
incardon's avatar
incardon committed
689
		stat.add();
incardon's avatar
incardon committed
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722

		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());
723
		if (req.size() != 0) {MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));}
incardon's avatar
incardon committed
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

		// 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
750
			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
751 752

			req.add();
753 754 755 756 757

#ifdef SE_CLASS2
				check_valid(ptr,proc_com.get(i));
#endif

incardon's avatar
incardon committed
758 759 760 761 762 763 764 765
			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();
766 767 768 769 770

#ifdef SE_CLASS2
			check_valid(ptr[i],sz[i]);
#endif

incardon's avatar
incardon committed
771 772 773 774
			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());
775
		if (req.size() != 0) {MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));}
incardon's avatar
incardon committed
776 777 778 779 780 781

		// Remove the executed request
		req.clear();
		stat.clear();
	}

782 783 784 785 786 787
	/*! \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
788 789 790 791 792 793 794 795 796 797 798 799
	 * \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
	 *
	 */
800
	bool send(size_t proc, size_t tag, const void * mem, size_t sz)
incardon's avatar
incardon committed
801 802 803 804 805 806 807 808 809 810 811 812
	{
		// 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
813

incardon's avatar
incardon committed
814 815 816 817 818 819 820
	/*! \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
821 822 823 824 825 826 827
	 *
	 * \see sendRecvMultipleMessages
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
	 *
828 829
	 * \return true if succeed false otherwise
	 *
830
	 */
831
	template<typename T, typename Mem, typename gr> bool send(size_t proc, size_t tag, openfpm::vector<T,Mem,gr> & v)
832 833 834 835 836 837 838 839 840 841
	{
#ifdef DEBUG
		checkType<T>();
#endif

		// send over MPI

		// Create one request
		req.add();

842
		// send
843
		MPI_IsendW<T,Mem,gr>::send(proc,SEND_RECV_BASE + tag,v,req.last());
844 845

		return true;
846 847 848 849 850 851
	}

	/*! \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
852
	 *          consider to use sendrecvMultipleMessagesPCX or sendrecvMultipleMessagesNBX
853
	 *
incardon's avatar
incardon committed
854
	 * \warning operation is asynchronous execute must be called to ensure they are executed
855
	 *
856
	 * \see sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
857 858 859 860
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
incardon's avatar
incardon committed
861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878
	 * \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
879 880 881 882
    /*! \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
883
     *          consider to use sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
884 885 886
     *
     * \warning operation is asynchronous execute must be called to ensure they are executed
     *
887
     * \see sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913
     *
     * \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
914 915
	 *
	 * send a primitive data T receive the same primitive T from all the other processors
incardon's avatar
incardon committed
916 917 918
	 *
	 * \warning operation is asynchronous execute must be called to ensure they are executed
	 *
919
	 * \param v vector to receive (automaticaly resized)
incardon's avatar
incardon committed
920
	 * \param send data to send
921
	 *
922 923
	 * \return true if succeed false otherwise
	 *
924
	 */
incardon's avatar
incardon committed
925
	template<typename T, typename Mem, typename gr> bool allGather(T & send, openfpm::vector<T,Mem,gr> & v)
926 927 928 929 930 931 932 933
	{
#ifdef DEBUG
		checkType<T>();
#endif

		// Create one request
		req.add();

incardon's avatar
incardon committed
934 935 936
		// Number of processors
		v.resize(getProcessingUnits());

937
		// gather
incardon's avatar
incardon committed
938
		MPI_IAllGatherW<T>::gather(&send,1,v.getPointer(),1,req.last());
939 940

		return true;
941 942 943
	}

	/*! \brief Execute all the requests
incardon's avatar
incardon committed
944 945 946 947 948 949
	 *
	 */
	void execute()
	{
		int err = 0;

950 951 952 953
		// if req == 0 return
		if (req.size() == 0)
			return;

incardon's avatar
incardon committed
954
		// Wait for all the requests
955
		stat.resize(req.size());
incardon's avatar
incardon committed
956 957 958 959 960 961
		err = MPI_Waitall(req.size(),&req.get(0),&stat.get(0));

		// MPI error get the message and abort MPI
		if (err != MPI_SUCCESS)
			MPI_Abort(MPI_COMM_WORLD,1);

962
		// Remove executed request and status
incardon's avatar
incardon committed
963 964 965
		req.clear();
		stat.clear();
	}
966 967


968 969
	/////////////////////// Semantic communication ///////////////////////
	#include "VCluster_semantic.ipp"
incardon's avatar
incardon committed
970 971
};

Pietro Incardona's avatar
Pietro Incardona committed
972
// Function to initialize the global VCluster //
incardon's avatar
incardon committed
973 974 975

extern Vcluster * global_v_cluster;

Pietro Incardona's avatar
Pietro Incardona committed
976 977 978 979 980 981 982

/*! \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
983
static inline void init_global_v_cluster(int *argc, char ***argv)
Pietro Incardona's avatar
Pietro Incardona committed
984 985 986 987 988
{
	if (global_v_cluster == NULL)
		global_v_cluster = new Vcluster(argc,argv);
}

Pietro Incardona's avatar
Pietro Incardona committed
989
static inline void delete_global_v_cluster()
Pietro Incardona's avatar
Pietro Incardona committed
990 991 992 993
{
	delete global_v_cluster;
}

incardon's avatar
incardon committed
994 995
#endif