VCluster.hpp 10.8 KB
Newer Older
incardon's avatar
incardon committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
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
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
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
113
114
115
116
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
#ifndef VCLUSTER
#define VCLUSTER

#include <mpi.h>
#include "VCluster_object.hpp"
#include "VCluster_object_array.hpp"
#include "Vector/map_vector.hpp"
#include "MPI_IallreduceW.hpp"
#include <exception>
#include <Vector/map_vector.hpp>

#define MSG_LENGTH 1024
#define MSG_SEND_RECV 1025
#define NONE 1
#define NEED_ALL_SIZE 2

#define MPI_SAFE_CALL(call) {\
	int err = call;\
	if (MPI_SUCCESS != err) {\
		std::cerr << "MPI error: "<< __FILE__ << " " << __LINE__ << "\n";\
		error_handler(err);\
	}\
}

///////////////////// 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";
  }
};

/*! \brief This class is virtualize the cluster as a set of processing unit
 *         and communication unit
 *
 * This class virtualize the cluster as a set of processing unit and communication
 * unit. It can execute any vcluster_exe
 *
 */

class Vcluster
{
	// 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;


	void error_handler(int error_code)
	{
	   char error_string[BUFSIZ];
	   int length_of_error_string, error_class;

	   MPI_Error_class(error_code, &error_class);
	   MPI_Error_string(error_class, error_string, &length_of_error_string);
	   std::cerr << getProcessUnitID() << ": " << error_string;
	   MPI_Error_string(error_code, error_string, &length_of_error_string);
	   std::cerr << getProcessUnitID() << ": " << error_string;
	}


public:

	// Finalize the MPI program
	~Vcluster()
	{
	}

	//! \brief Virtual cluster constructor
	Vcluster(int *argc, char ***argv)
	{

		//! Get the total number of process
		//! and the rank of this process

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

#ifdef MEMLEAK_CHECK
			process_v_cl = rank;
#endif

		//! create and fill map scatter with one
		map_scatter.resize(size);

		for (size_t i = 0 ; i < map_scatter.size() ; i++)
		{
			map_scatter.get(i) = 1;
		}
	}

	//! 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;
	}

	/*! \brief Sum the number across all processors and get the result
	 *
	 * \param num to reduce, input and output
	 *
	 */

	template<typename T> void reduce(T & num)
	{
		// reduce over MPI

		// Create one request
		req.add();

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

incardon's avatar
incardon committed
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
	/*! \brief Get the maximum number across all processors (or reduction with insinity norm)
	 *
	 * \param num to reduce
	 *
	 */

	template<typename T> void max(T & num)
	{
		// reduce over MPI

		// Create one request
		req.add();

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

incardon's avatar
incardon committed
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
	// 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;


	/*! \brief Send and receive multiple messages
	 *
	 * It send multiple (to more than one) messages and receive
	 * other multiple messages, all the processor must call this
	 * function
	 *
	 * \param prc list of processors with which it should communicate
	 *
	 * \param v vector containing the data to send
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
	 *        for the incoming message and give back a valid pointer, the 3 parameters
	 *        in the call-back are  , total message to receive, i processor id from witch
	 *        to receive
	 *
	 * \param opt options, NONE or NEED_ALL_SIZE, with NEED_ALL_SIZE the allocation
	 *        callback will not be called until all the message size will be
	 *        gathered, [usefull for example with you want to allocate one big buffer
	 *        to gather all the messages]
	 *
	 */

incardon's avatar
incardon committed
251
	template<typename T> void sendrecvMultipleMessages(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
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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
	{
		// 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);
		}

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

	/*! \brief Send and receive multiple messages
	 *
	 * It send multiple (to more than one) messages and receive
	 * other multiple messages, all the processor must call this
	 * function
	 *
	 * \param n_send number of send this processor must do
	 *
	 * \param map array containing an array of unsigned chars that
	 *        specify the communication pattern of the processor
	 *
	 *        [Example]   0 1 0 0 0 0 1 1 1 mean that the processor
	 *        communicate with the processor 1 6 7 8
	 *
	 * \param sz the array contain the size of the message for each processor
	 *        (zeros must be omitted)
	 *
	 *        [Example] for the previous patter 5 10 15 4 mean processor 1
	 *        message size 5 byte, processor 6 message size 10 , ......
	 *
	 * \param prc list of processor with which it should communicate
	 *        [Example] for the previous case should be
	 *        1 6 7 8 (prc and mp contain the same information in different
	 *        format, giving both reduce the computation)
	 *
	 * \param ptr array that contain the message (zero lengh must be omitted)
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
	 *        for the incoming message and give back a valid pointer, the 3 parameters
	 *        in the call-back are  , total message to receive, i processor id from witch
	 *        to receive
	 *
	 * \param opt options, NONE or NEED_ALL_SIZE, with NEED_ALL_SIZE the allocation
	 *        callback will not be called until all the message size will be
	 *        gathered, [usefull for example with you want to allocate one big buffer
	 *        to gather all the messages]
	 *
	 */

incardon's avatar
incardon committed
315
	void sendrecvMultipleMessages(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
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
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
377
378
	{
		req.clear();
		req.add();

		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());
		MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));

		// 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
379
			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
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408

			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());
		MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));

		// Remove the executed request

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

	/*! \brief Execute all the request
	 *
	 */
	void execute()
	{
		int err = 0;

409
410
411
412
		// if req == 0 return
		if (req.size() == 0)
			return;

incardon's avatar
incardon committed
413
		// Wait for all the requests
414
		stat.resize(req.size());
incardon's avatar
incardon committed
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
		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);
		}

		//! Remove executed request and status
		req.clear();
		stat.clear();
	}
};

void init_global_v_cluster(int *argc, char ***argv);

extern Vcluster * global_v_cluster;

#endif