VCluster.hpp 33 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"
Yaroslav's avatar
Yaroslav committed
21
#include "Packer_Unpacker/has_max_prop.hpp"
incardon's avatar
incardon committed
22
#include "data_type/aggregate.hpp"
incardon's avatar
incardon committed
23

Pietro Incardona's avatar
Pietro Incardona committed
24 25 26 27
#ifdef HAVE_PETSC
#include <petscvec.h>
#endif

incardon's avatar
incardon committed
28 29
#define MSG_LENGTH 1024
#define MSG_SEND_RECV 1025
30
#define SEND_SPARSE 4096
incardon's avatar
incardon committed
31 32 33
#define NONE 1
#define NEED_ALL_SIZE 2

34 35
#define SERIVCE_MESSAGE_TAG 16384
#define SEND_RECV_BASE 8192
incardon's avatar
incardon committed
36
#define GATHER_BASE 24576
incardon's avatar
incardon committed
37

Pietro Incardona's avatar
Pietro Incardona committed
38
// number of vcluster instances
incardon's avatar
incardon committed
39
extern size_t n_vcluster;
Pietro Incardona's avatar
Pietro Incardona committed
40
// Global MPI initialization
incardon's avatar
incardon committed
41
extern bool global_mpi_init;
Pietro Incardona's avatar
Pietro Incardona committed
42 43
// initialization flag
extern bool ofp_initialized;
incardon's avatar
incardon committed
44

incardon's avatar
incardon committed
45 46 47 48 49 50 51
///////////////////// Post functions /////////////

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

incardon's avatar
incardon committed
52
//! Helper class to add data without serialization
incardon's avatar
incardon committed
53 54 55
template<bool sr>
struct op_ssend_recv_add_sr
{
incardon's avatar
incardon committed
56
	//! Add data
incardon's avatar
incardon committed
57 58 59 60 61 62 63
	template<typename T, typename D, typename S, int ... prp> static void execute(D & recv,S & v2, size_t i)
	{
		// Merge the information
		recv.template add_prp<typename T::value_type,PtrMemory,openfpm::grow_policy_identity,openfpm::vect_isel<typename T::value_type>::value,prp...>(v2);
	}
};

incardon's avatar
incardon committed
64
//! Helper class to add data with serialization
incardon's avatar
incardon committed
65 66 67
template<>
struct op_ssend_recv_add_sr<true>
{
incardon's avatar
incardon committed
68
	//! Add data
incardon's avatar
incardon committed
69 70 71 72 73 74 75
	template<typename T, typename D, typename S, int ... prp> static void execute(D & recv,S & v2, size_t i)
	{
		// Merge the information
		recv.template add_prp<typename T::value_type,HeapMemory,openfpm::grow_policy_double,openfpm::vect_isel<typename T::value_type>::value, prp...>(v2);
	}
};

incardon's avatar
incardon committed
76
//! Helper class to add data
incardon's avatar
incardon committed
77 78 79
template<typename op>
struct op_ssend_recv_add
{
incardon's avatar
incardon committed
80
	//! Add data
incardon's avatar
incardon committed
81 82 83 84 85 86 87
	template<bool sr, typename T, typename D, typename S, int ... prp> static void execute(D & recv,S & v2, size_t i)
	{
		// Merge the information
		op_ssend_recv_add_sr<sr>::template execute<T,D,S,prp...>(recv,v2,i);
	}
};

incardon's avatar
incardon committed
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
//! Helper class to merge data without serialization
template<bool sr,template<typename,typename> class op>
struct op_ssend_recv_merge_impl
{
	//! Merge the
	template<typename T, typename D, typename S, int ... prp> inline static void execute(D & recv,S & v2,size_t i,openfpm::vector<openfpm::vector<aggregate<size_t,size_t>>> & opart)
	{
		// Merge the information
		recv.template merge_prp_v<op,typename T::value_type, PtrMemory, openfpm::grow_policy_identity, prp...>(v2,opart.get(i));
	}
};

//! Helper class to merge data with serialization
template<template<typename,typename> class op>
struct op_ssend_recv_merge_impl<true,op>
{
	//! merge the data
	template<typename T, typename D, typename S, int ... prp> inline static void execute(D & recv,S & v2,size_t i,openfpm::vector<openfpm::vector<aggregate<size_t,size_t>>> & opart)
	{
		// Merge the information
		recv.template merge_prp_v<op,typename T::value_type, HeapMemory, openfpm::grow_policy_double, prp...>(v2,opart.get(i));
	}
};

//! Helper class to merge data
incardon's avatar
incardon committed
113 114 115
template<template<typename,typename> class op>
struct op_ssend_recv_merge
{
incardon's avatar
incardon committed
116
	//! For each processor contain the list of the particles with which I must merge the information
incardon's avatar
incardon committed
117 118
	openfpm::vector<openfpm::vector<aggregate<size_t,size_t>>> & opart;

incardon's avatar
incardon committed
119
	//! constructor
incardon's avatar
incardon committed
120 121 122 123
	op_ssend_recv_merge(openfpm::vector<openfpm::vector<aggregate<size_t,size_t>>> & opart)
	:opart(opart)
	{}

incardon's avatar
incardon committed
124
	//! execute the merge
incardon's avatar
incardon committed
125 126
	template<bool sr, typename T, typename D, typename S, int ... prp> void execute(D & recv,S & v2,size_t i)
	{
incardon's avatar
incardon committed
127
		op_ssend_recv_merge_impl<sr,op>::template execute<T,D,S,prp...>(recv,v2,i,opart);
incardon's avatar
incardon committed
128 129 130
	}
};

incardon's avatar
incardon committed
131 132
//////////////////////////////////////////////////

incardon's avatar
incardon committed
133
//! temporal buffer for reductions
incardon's avatar
incardon committed
134 135
union red
{
incardon's avatar
incardon committed
136
	//! char
incardon's avatar
incardon committed
137
	char c;
incardon's avatar
incardon committed
138
	//! unsigned char
incardon's avatar
incardon committed
139
	unsigned char uc;
incardon's avatar
incardon committed
140
	//! signed
incardon's avatar
incardon committed
141
	short s;
incardon's avatar
incardon committed
142
	//! unsigned short
incardon's avatar
incardon committed
143
	unsigned short us;
incardon's avatar
incardon committed
144
	//! integer
incardon's avatar
incardon committed
145
	int i;
incardon's avatar
incardon committed
146
	//! unsigned integer
incardon's avatar
incardon committed
147
	unsigned int ui;
incardon's avatar
incardon committed
148
	//! float
incardon's avatar
incardon committed
149
	float f;
incardon's avatar
incardon committed
150
	//! double
incardon's avatar
incardon committed
151 152 153 154
	double d;
};


155
/*! \brief This class virtualize the cluster of PC as a set of processes that communicate
incardon's avatar
incardon committed
156
 *
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
 * 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
176 177 178 179 180
 *
 */

class Vcluster
{
incardon's avatar
incardon committed
181
	//! log file
incardon's avatar
incardon committed
182 183
	Vcluster_log log;

184 185 186
	//! 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
187
	//! an acknowledge notification on the sending processor.
188 189 190 191 192 193
	//! 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
194 195 196 197
	//! 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
198
	size_t NBX_cnt;
199

incardon's avatar
incardon committed
200 201
	//! temporal vector used for meta-communication
	//! ( or meta-data before the real communication )
incardon's avatar
incardon committed
202 203
	openfpm::vector<size_t> proc_com;

incardon's avatar
incardon committed
204
	//! vector that contain the scatter map (it is basically an array of one)
incardon's avatar
incardon committed
205 206
	openfpm::vector<int> map_scatter;

incardon's avatar
incardon committed
207
	//! vector of MPI requests
incardon's avatar
incardon committed
208 209
	openfpm::vector<MPI_Request> req;

incardon's avatar
incardon committed
210
	//! vector of MPI status
incardon's avatar
incardon committed
211 212
	openfpm::vector<MPI_Status> stat;

incardon's avatar
incardon committed
213
	//! vector of functions to execute after all the request has been performed
incardon's avatar
incardon committed
214 215 216 217 218 219 220
	std::vector<int> post_exe;

	// Object array


	// Single objects

incardon's avatar
incardon committed
221
	//! number of processes
incardon's avatar
incardon committed
222
	int size;
incardon's avatar
incardon committed
223
	//! actual rank
incardon's avatar
incardon committed
224 225
	int rank;

incardon's avatar
incardon committed
226
	//! number of processing unit per process
incardon's avatar
incardon committed
227 228
	int numPE = 1;

incardon's avatar
incardon committed
229
	/*! This buffer is a temporal buffer for reductions
incardon's avatar
incardon committed
230 231 232 233 234 235 236 237
	 *
	 * 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;

incardon's avatar
incardon committed
238
	//! vector of pointers of send buffers
239 240
	openfpm::vector<void *> ptr_send;

incardon's avatar
incardon committed
241
	//! vector of the size of send buffers
242 243
	openfpm::vector<size_t> sz_send;

incardon's avatar
incardon committed
244
	//! sending map
245 246
	openfpm::vector<size_t> map;

incardon's avatar
incardon committed
247
	//! Receive buffers
248
	openfpm::vector<BHeapMemory> recv_buf;
249

incardon's avatar
incardon committed
250
	//! barrier request
251
	MPI_Request bar_req;
incardon's avatar
incardon committed
252 253

	//! barrier status
254 255
	MPI_Status bar_stat;

incardon's avatar
incardon committed
256
	//! disable operator=
Pietro Incardona's avatar
Pietro Incardona committed
257 258
	Vcluster & operator=(const Vcluster &)	{return *this;};

incardon's avatar
incardon committed
259 260 261 262
	//! disable copy constructor
	Vcluster(const Vcluster &)
	:NBX_cnt(0)
	{};
Pietro Incardona's avatar
Pietro Incardona committed
263

incardon's avatar
incardon committed
264 265 266 267 268
public:

	// Finalize the MPI program
	~Vcluster()
	{
Pietro Incardona's avatar
Pietro Incardona committed
269 270 271
#ifdef SE_CLASS2
		check_delete(this);
#endif
incardon's avatar
incardon committed
272 273 274 275
		n_vcluster--;

		// if there are no other vcluster instances finalize
		if (n_vcluster == 0)
Pietro Incardona's avatar
Pietro Incardona committed
276
		{
277 278 279 280
			int already_finalised;

			MPI_Finalized(&already_finalised);
			if (!already_finalised)
Pietro Incardona's avatar
Pietro Incardona committed
281
			{
282 283 284 285
				if (MPI_Finalize() != 0)
				{
					std::cerr << __FILE__ << ":" << __LINE__  << " MPI_Finalize FAILED \n";
				}
Pietro Incardona's avatar
Pietro Incardona committed
286
			}
Pietro Incardona's avatar
Pietro Incardona committed
287
		}
incardon's avatar
incardon committed
288 289
	}

290 291 292 293 294 295
	/*! \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
296
	Vcluster(int *argc, char ***argv)
incardon's avatar
incardon committed
297
	:NBX_cnt(0)
incardon's avatar
incardon committed
298
	{
Pietro Incardona's avatar
Pietro Incardona committed
299 300 301 302
#ifdef SE_CLASS2
		check_new(this,8,VCLUSTER_EVENT,PRJ_VCLUSTER);
#endif

incardon's avatar
incardon committed
303 304
		n_vcluster++;

305 306 307
		int already_initialised;
		MPI_Initialized(&already_initialised);

incardon's avatar
incardon committed
308
		// Check if MPI is already initialized
309
		if (!already_initialised)
incardon's avatar
incardon committed
310
		{
incardon's avatar
incardon committed
311

incardon's avatar
incardon committed
312 313
			MPI_Init(argc,argv);
		}
incardon's avatar
incardon committed
314

315 316
		// Get the total number of process
		// and the rank of this process
incardon's avatar
incardon committed
317 318 319 320

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

321
#ifdef SE_CLASS2
incardon's avatar
incardon committed
322 323 324
			process_v_cl = rank;
#endif

325
		// create and fill map scatter with one
incardon's avatar
incardon committed
326 327 328 329 330 331
		map_scatter.resize(size);

		for (size_t i = 0 ; i < map_scatter.size() ; i++)
		{
			map_scatter.get(i) = 1;
		}
incardon's avatar
incardon committed
332 333 334

		// open the log file
		log.openLog(rank);
335 336 337 338

		// Initialize bar_req
		bar_req = MPI_Request();
		bar_stat = MPI_Status();
incardon's avatar
incardon committed
339 340
	}

341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376
#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
377
				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" ;
378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
				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
394 395 396 397 398 399 400 401 402 403
	/*! \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
404 405 406 407 408 409 410 411 412 413 414 415
	//! Get the total number of processing units
	size_t getProcessingUnits()
	{
		return size*numPE;
	}

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

416
	/*! \brief Sum the numbers across all processors and get the result
incardon's avatar
incardon committed
417 418 419 420 421
	 *
	 * \param num to reduce, input and output
	 *
	 */

422
	template<typename T> void sum(T & num)
incardon's avatar
incardon committed
423
	{
424 425 426 427
#ifdef DEBUG
		checkType<T>();
#endif

incardon's avatar
incardon committed
428 429 430 431 432 433 434 435 436
		// reduce over MPI

		// Create one request
		req.add();

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

437
	/*! \brief Get the maximum number across all processors (or reduction with infinity norm)
incardon's avatar
incardon committed
438 439 440 441 442 443
	 *
	 * \param num to reduce
	 *
	 */
	template<typename T> void max(T & num)
	{
444 445 446
#ifdef DEBUG
		checkType<T>();
#endif
incardon's avatar
incardon committed
447 448 449 450 451 452 453 454 455
		// reduce over MPI

		// Create one request
		req.add();

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

tonynsyde's avatar
tonynsyde committed
456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475
	/*! \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());
	}

476

477
	/*! \brief Send and receive multiple messages
478
	 *
479 480 481 482 483 484 485 486 487 488 489 490
	 * 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]
	 *
incardon's avatar
incardon committed
491
	 * \param data vector containing the data to send [v=vector<vector<int>>, v.size()=4, T=vector<int>], T at the moment
492 493 494 495 496 497 498 499 500 501 502 503
	 *          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
504
	 *           every time message_alloc is called)
505
	 *        * void pointer, parameter for additional data to pass to the call-back
506
	 *
incardon's avatar
incardon committed
507 508
	 * \param ptr_arg data passed to the call-back function specified
	 *
509
	 * \param opt options, only NONE supported
510 511 512 513
	 *
	 */
	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)
	{
514 515 516
#ifdef DEBUG
		checkType<typename T::value_type>();
#endif
517 518 519 520 521 522 523 524 525 526 527 528 529
		// 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
530 531
	/*! \brief Send and receive multiple messages
	 *
532 533 534 535 536 537 538 539 540 541 542 543
	 * 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]
	 *
incardon's avatar
incardon committed
544
	 * \param data vector containing the data to send [v=vector<vector<int>>, v.size()=4, T=vector<int>], T at the moment
545 546 547 548 549 550 551 552 553 554 555 556
	 *          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
557
	 *           every time message_alloc is called)
558
	 *        * void pointer, parameter for additional data to pass to the call-back
incardon's avatar
incardon committed
559
	 *
incardon's avatar
incardon committed
560 561
	 * \param ptr_arg data passed to the call-back function specified
	 *
562
	 * \param opt options, only NONE supported
incardon's avatar
incardon committed
563 564
	 *
	 */
565
	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
566
	{
567 568 569
#ifdef DEBUG
		checkType<typename T::value_type>();
#endif
incardon's avatar
incardon committed
570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589
		// 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);
		}

590 591 592
		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);
	}

593
	/*! \brief Send and receive multiple messages
594
	 *
595 596 597
	 * 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+)
598
	 *
599 600 601 602 603
	 * 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
604
	 *
605
	 * \param n_send number of send for this processor [4]
606 607
	 *
	 * \param prc list of processor with which it should communicate
608
	 *        [1,1,6,7,8]
609
	 *
610 611 612 613
	 * \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
614 615
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
616 617 618
	 *        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
619
	 *        in the call-back are in order:
620 621 622 623 624
	 *        * 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
625
	 *           every time message_alloc is called)
626
	 *        * void pointer, parameter for additional data to pass to the call-back
627
	 *
incardon's avatar
incardon committed
628 629
	 * \param ptr_arg data passed to the call-back function specified
	 *
630 631 632
	 * \param opt options, NONE (ignored in this moment)
	 *
	 */
633
	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)
634
	{
635 636 637 638 639
		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();
640 641 642 643 644
		req.clear();
		// Do MPI_Issend

		for (size_t i = 0 ; i < n_send ; i++)
		{
incardon's avatar
incardon committed
645 646 647
			if (sz[i] != 0)
			{
				req.add();
648 649 650 651 652

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

653
				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
654
				log.logSend(prc[i]);
incardon's avatar
incardon committed
655
			}
656 657 658 659 660 661 662
		}

		size_t rid = 0;
		int flag = false;

		bool reached_bar_req = false;

incardon's avatar
incardon committed
663 664
		log.start(10);

665 666 667
		// Wait that all the send are acknowledge
		do
		{
incardon's avatar
incardon committed
668

669 670 671 672 673
			// flag that notify that this processor reach the barrier
			// Barrier request

			MPI_Status stat_t;
			int stat = false;
674
			MPI_SAFE_CALL(MPI_Iprobe(MPI_ANY_SOURCE,SEND_SPARSE + NBX_cnt,MPI_COMM_WORLD,&stat,&stat_t));
675

676
			// If I have an incoming message and is related to this NBX communication
677 678 679 680 681 682 683 684 685
			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
686 687 688
				// Log the receiving request
				log.logRecv(stat_t);

689 690
				rid++;

691 692 693 694
				// Check the pointer
#ifdef SE_CLASS2
				check_valid(ptr,msize);
#endif
695
				MPI_SAFE_CALL(MPI_Recv(ptr,msize,MPI_BYTE,stat_t.MPI_SOURCE,SEND_SPARSE+NBX_cnt,MPI_COMM_WORLD,&stat_t));
696 697 698 699

#ifdef SE_CLASS2
				check_valid(ptr,msize);
#endif
700 701 702 703 704 705 706
			}

			// 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
707 708 709 710
				if (req.size() != 0)
				{MPI_SAFE_CALL(MPI_Testall(req.size(),&req.get(0),&flag,MPI_STATUSES_IGNORE));}
				else
					flag = true;
711 712 713 714 715 716

				// 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
717
			// Check if all processor reached the async barrier
718
			if (reached_bar_req)
incardon's avatar
incardon committed
719
			{MPI_SAFE_CALL(MPI_Test(&bar_req,&flag,&bar_stat))};
incardon's avatar
incardon committed
720 721

			// produce a report if communication get stuck
incardon's avatar
incardon committed
722
			log.NBXreport(NBX_cnt,req,reached_bar_req,bar_stat);
incardon's avatar
incardon committed
723

724 725 726 727 728 729
		} while (flag == false);

		// Remove the executed request

		req.clear();
		stat.clear();
incardon's avatar
incardon committed
730
		log.clear();
731 732 733

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
incardon's avatar
incardon committed
734 735 736 737
	}

	/*! \brief Send and receive multiple messages
	 *
738 739 740
	 * 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
741
	 *
742 743 744 745 746
	 * 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
747
	 *
748
	 * \param n_send number of send for this processor [4]
incardon's avatar
incardon committed
749
	 *
750 751
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
incardon's avatar
incardon committed
752
	 *
753 754 755
	 * \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
756
	 *
757 758
	 * \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
759
	 *
760
	 * \param ptr array that contain the pointers to the message to send
incardon's avatar
incardon committed
761 762
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
763 764 765 766 767 768 769 770 771 772 773
	 *        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
774
	 *
incardon's avatar
incardon committed
775 776
	 * \param ptr_arg pointer passed to the call-back function
	 *
777
	 * \param opt options, NONE (ignored in this moment)
incardon's avatar
incardon committed
778 779
	 *
	 */
780
	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
781
	{
782 783 784 785
		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
786
		req.clear();
787

incardon's avatar
incardon committed
788
		req.add();
incardon's avatar
incardon committed
789
		stat.add();
incardon's avatar
incardon committed
790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822

		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());
823
		if (req.size() != 0) {MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));}
incardon's avatar
incardon committed
824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849

		// 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
850
			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
851 852

			req.add();
853 854 855 856 857

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

incardon's avatar
incardon committed
858 859 860 861 862 863 864 865
			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();
866 867 868 869 870

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

incardon's avatar
incardon committed
871 872 873 874
			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());
875
		if (req.size() != 0) {MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));}
incardon's avatar
incardon committed
876 877 878 879 880 881

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

882 883 884 885 886 887
	/*! \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
888 889 890 891 892 893 894 895 896 897 898 899
	 * \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
	 *
	 */
900
	bool send(size_t proc, size_t tag, const void * mem, size_t sz)
incardon's avatar
incardon committed
901 902 903 904 905 906 907 908 909 910 911 912
	{
		// 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
913

incardon's avatar
incardon committed
914 915 916 917 918 919 920
	/*! \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
921 922 923 924 925 926 927
	 *
	 * \see sendRecvMultipleMessages
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
	 *
928 929
	 * \return true if succeed false otherwise
	 *
930
	 */
931
	template<typename T, typename Mem, typename gr> bool send(size_t proc, size_t tag, openfpm::vector<T,Mem,gr> & v)
932 933 934 935 936 937 938 939 940 941
	{
#ifdef DEBUG
		checkType<T>();
#endif

		// send over MPI

		// Create one request
		req.add();

942
		// send
943
		MPI_IsendW<T,Mem,gr>::send(proc,SEND_RECV_BASE + tag,v,req.last());
944 945

		return true;
946 947 948 949 950 951
	}

	/*! \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
952
	 *          consider to use sendrecvMultipleMessagesPCX or sendrecvMultipleMessagesNBX
953
	 *
incardon's avatar
incardon committed
954
	 * \warning operation is asynchronous execute must be called to ensure they are executed
955
	 *
956
	 * \see sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
957 958 959 960
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
incardon's avatar
incardon committed
961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978
	 * \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
979 980 981 982
    /*! \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
983
     *          consider to use sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
984 985 986
     *
     * \warning operation is asynchronous execute must be called to ensure they are executed
     *
987
     * \see sendrecvMultipleMessagesPCX sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013
     *
     * \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
1014 1015
	 *
	 * send a primitive data T receive the same primitive T from all the other processors
incardon's avatar
incardon committed
1016 1017 1018
	 *
	 * \warning operation is asynchronous execute must be called to ensure they are executed
	 *
1019
	 * \param v vector to receive (automaticaly resized)
incardon's avatar
incardon committed
1020
	 * \param send data to send
1021
	 *
1022 1023
	 * \return true if succeed false otherwise
	 *
1024
	 */
incardon's avatar
incardon committed
1025
	template<typename T, typename Mem, typename gr> bool allGather(T & send, openfpm::vector<T,Mem,gr> & v)
1026 1027 1028 1029 1030 1031 1032 1033
	{
#ifdef DEBUG
		checkType<T>();
#endif

		// Create one request
		req.add();

incardon's avatar
incardon committed
1034 1035 1036
		// Number of processors
		v.resize(getProcessingUnits());

1037
		// gather
incardon's avatar
incardon committed
1038
		MPI_IAllGatherW<T>::gather(&send,1,v.getPointer(),1,req.last());
1039 1040

		return true;
1041 1042 1043
	}

	/*! \brief Execute all the requests
incardon's avatar
incardon committed
1044 1045 1046 1047
	 *
	 */
	void execute()
	{
1048 1049 1050 1051
		// if req == 0 return
		if (req.size() == 0)
			return;

incardon's avatar
incardon committed
1052
		// Wait for all the requests
1053
		stat.resize(req.size());
1054
		MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));
incardon's avatar
incardon committed
1055

1056
		// Remove executed request and status
incardon's avatar
incardon committed
1057 1058 1059
		req.clear();
		stat.clear();
	}
1060 1061


1062 1063
	/////////////////////// Semantic communication ///////////////////////
	#include "VCluster_semantic.ipp"
incardon's avatar
incardon committed
1064 1065
};

Pietro Incardona's avatar
Pietro Incardona committed
1066
// Function to initialize the global VCluster //
incardon's avatar
incardon committed
1067

1068
extern Vcluster * global_v_cluster_private;
Pietro Incardona's avatar
Pietro Incardona committed
1069 1070 1071 1072 1073 1074 1075

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

1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098
static inline void init_global_v_cluster_private(int *argc, char ***argv)
{
	if (global_v_cluster_private == NULL)
		global_v_cluster_private = new Vcluster(argc,argv);
}

static inline void delete_global_v_cluster_private()
{
	delete global_v_cluster_private;
}

static inline Vcluster & create_vcluster()
{
#ifdef SE_CLASS1

	if (global_v_cluster_private == NULL)
		std::cerr << __FILE__ << ":" << __LINE__ << " Error you must call openfpm_init before using any distributed data structures";

#endif

	return *global_v_cluster_private;
}

Pietro Incardona's avatar
Pietro Incardona committed
1099 1100 1101 1102 1103 1104 1105 1106 1107 1108
/*! \brief Check if the library has been initialized
 *
 * \return true if the library has been initialized
 *
 */
static inline bool is_openfpm_init()
{
	return ofp_initialized;
}

1109 1110 1111 1112 1113 1114
/*! \brief Initialize the library
 *
 * This function MUST be called before any other function
 *
 */
static inline void openfpm_init(int *argc, char ***argv)
Pietro Incardona's avatar
Pietro Incardona committed
1115
{
Pietro Incardona's avatar
Pietro Incardona committed
1116 1117 1118 1119 1120 1121
#ifdef HAVE_PETSC

	PetscInitialize(argc,argv,NULL,NULL);

#endif

1122
	init_global_v_cluster_private(argc,argv);
Pietro Incardona's avatar
Pietro Incardona committed
1123
	ofp_initialized = true;
Pietro Incardona's avatar
Pietro Incardona committed
1124 1125
}

1126 1127 1128 1129 1130 1131
/*! \brief Finalize the library
 *
 * This function MUST be called at the end of the program
 *
 */
static inline void openfpm_finalize()
Pietro Incardona's avatar
Pietro Incardona committed
1132
{
Pietro Incardona's avatar
Pietro Incardona committed
1133 1134 1135 1136 1137 1138
#ifdef HAVE_PETSC

	PetscFinalize();

#endif

1139
	delete_global_v_cluster_private();
Pietro Incardona's avatar
Pietro Incardona committed
1140
	ofp_initialized = false;
Pietro Incardona's avatar
Pietro Incardona committed
1141 1142
}

incardon's avatar
incardon committed
1143 1144
#endif