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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
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
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
|
/*
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2015 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2008 Sun Microsystems, Inc. All rights reserved.
* Copyright (c) 2016 Intel, Inc. All rights reserved.
* Copyright (c) 2018 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "coll_tuned.h"
#include <stdio.h>
#include "mpi.h"
#include "ompi/communicator/communicator.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/mca/coll/base/base.h"
#include "ompi/mca/coll/base/coll_base_topo.h"
#include "coll_tuned.h"
#include "coll_tuned_dynamic_rules.h"
#include "coll_tuned_dynamic_file.h"
static int tuned_module_enable(mca_coll_base_module_t *module,
struct ompi_communicator_t *comm);
/*
* Initial query function that is invoked during MPI_INIT, allowing
* this component to disqualify itself if it doesn't support the
* required level of thread support.
*/
int ompi_coll_tuned_init_query(bool enable_progress_threads,
bool enable_mpi_threads)
{
return OMPI_SUCCESS;
}
/*
* Invoked when there's a new communicator that has been created.
* Look at the communicator and decide which set of functions and
* priority we want to return.
*/
mca_coll_base_module_t *
ompi_coll_tuned_comm_query(struct ompi_communicator_t *comm, int *priority)
{
mca_coll_tuned_module_t *tuned_module;
OPAL_OUTPUT((ompi_coll_tuned_stream, "coll:tuned:module_tuned query called"));
/**
* No support for inter-communicator yet.
*/
if (OMPI_COMM_IS_INTER(comm)) {
*priority = 0;
return NULL;
}
/**
* If it is inter-communicator and size is less than 2 we have specialized modules
* to handle the intra collective communications.
*/
if (OMPI_COMM_IS_INTRA(comm) && ompi_comm_size(comm) < 2) {
*priority = 0;
return NULL;
}
tuned_module = OBJ_NEW(mca_coll_tuned_module_t);
if (NULL == tuned_module) return NULL;
*priority = ompi_coll_tuned_priority;
/*
* Choose whether to use [intra|inter] decision functions
* and if using fixed OR dynamic rule sets.
* Right now you cannot mix them, maybe later on it can be changed
* but this would probably add an extra if and funct call to the path
*/
tuned_module->super.coll_module_enable = tuned_module_enable;
/* By default stick with the fixed version of the tuned collectives. Later on,
* when the module get enabled, set the correct version based on the availability
* of the dynamic rules.
*/
tuned_module->super.coll_allgather = ompi_coll_tuned_allgather_intra_dec_fixed;
tuned_module->super.coll_allgatherv = ompi_coll_tuned_allgatherv_intra_dec_fixed;
tuned_module->super.coll_allreduce = ompi_coll_tuned_allreduce_intra_dec_fixed;
tuned_module->super.coll_alltoall = ompi_coll_tuned_alltoall_intra_dec_fixed;
tuned_module->super.coll_alltoallv = ompi_coll_tuned_alltoallv_intra_dec_fixed;
tuned_module->super.coll_alltoallw = NULL;
tuned_module->super.coll_barrier = ompi_coll_tuned_barrier_intra_dec_fixed;
tuned_module->super.coll_bcast = ompi_coll_tuned_bcast_intra_dec_fixed;
tuned_module->super.coll_exscan = NULL;
tuned_module->super.coll_gather = ompi_coll_tuned_gather_intra_dec_fixed;
tuned_module->super.coll_gatherv = NULL;
tuned_module->super.coll_reduce = ompi_coll_tuned_reduce_intra_dec_fixed;
tuned_module->super.coll_reduce_scatter = ompi_coll_tuned_reduce_scatter_intra_dec_fixed;
tuned_module->super.coll_reduce_scatter_block = ompi_coll_tuned_reduce_scatter_block_intra_dec_fixed;
tuned_module->super.coll_scan = NULL;
tuned_module->super.coll_scatter = ompi_coll_tuned_scatter_intra_dec_fixed;
tuned_module->super.coll_scatterv = NULL;
return &(tuned_module->super);
}
/* We put all routines that handle the MCA user forced algorithm and parameter choices here */
/* recheck the setting of forced, called on module create (i.e. for each new comm) */
static int
ompi_coll_tuned_forced_getvalues( enum COLLTYPE type,
coll_tuned_force_algorithm_params_t *forced_values )
{
coll_tuned_force_algorithm_mca_param_indices_t* mca_params;
const int *tmp = NULL;
mca_params = &(ompi_coll_tuned_forced_params[type]);
/**
* Set the selected algorithm to 0 by default. Later on we can check this against 0
* to see if it was set explicitly (if we suppose that setting it to 0 enable the
* default behavior) or not.
*/
mca_base_var_get_value(mca_params->algorithm_param_index, &tmp, NULL, NULL);
forced_values->algorithm = tmp ? tmp[0] : 0;
if( BARRIER != type ) {
mca_base_var_get_value(mca_params->segsize_param_index, &tmp, NULL, NULL);
if (tmp) forced_values->segsize = tmp[0];
mca_base_var_get_value(mca_params->tree_fanout_param_index, &tmp, NULL, NULL);
if (tmp) forced_values->tree_fanout = tmp[0];
mca_base_var_get_value(mca_params->chain_fanout_param_index, &tmp, NULL, NULL);
if (tmp) forced_values->chain_fanout = tmp[0];
mca_base_var_get_value(mca_params->max_requests_param_index, &tmp, NULL, NULL);
if (tmp) forced_values->max_requests = tmp[0];
}
return (MPI_SUCCESS);
}
#define COLL_TUNED_EXECUTE_IF_DYNAMIC(TMOD, TYPE, EXECUTE) \
{ \
int need_dynamic_decision = 0; \
ompi_coll_tuned_forced_getvalues( (TYPE), &((TMOD)->user_forced[(TYPE)]) ); \
(TMOD)->com_rules[(TYPE)] = NULL; \
if( 0 != (TMOD)->user_forced[(TYPE)].algorithm ) { \
need_dynamic_decision = 1; \
} \
if( NULL != mca_coll_tuned_component.all_base_rules ) { \
(TMOD)->com_rules[(TYPE)] \
= ompi_coll_tuned_get_com_rule_ptr( mca_coll_tuned_component.all_base_rules, \
(TYPE), size ); \
if( NULL != (TMOD)->com_rules[(TYPE)] ) { \
need_dynamic_decision = 1; \
} \
} \
if( 1 == need_dynamic_decision ) { \
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned: enable dynamic selection for "#TYPE)); \
EXECUTE; \
} \
}
/*
* Init module on the communicator
*/
static int
tuned_module_enable( mca_coll_base_module_t *module,
struct ompi_communicator_t *comm )
{
int size;
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t *) module;
mca_coll_base_comm_t *data = NULL;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init called."));
/* Allocate the data that hangs off the communicator */
if (OMPI_COMM_IS_INTER(comm)) {
size = ompi_comm_remote_size(comm);
} else {
size = ompi_comm_size(comm);
}
/**
* we still malloc data as it is used by the TUNED modules
* if we don't allocate it and fall back to a BASIC module routine then confuses debuggers
* we place any special info after the default data
*
* BUT on very large systems we might not be able to allocate all this memory so
* we do check a MCA parameter to see if if we should allocate this memory
*
* The default is set very high
*/
/* prepare the placeholder for the array of request* */
data = OBJ_NEW(mca_coll_base_comm_t);
if (NULL == data) {
return OMPI_ERROR;
}
if (ompi_coll_tuned_use_dynamic_rules) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init MCW & Dynamic"));
/**
* next dynamic state, recheck all forced rules as well
* warning, we should check to make sure this is really an INTRA comm here...
*/
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, ALLGATHER,
tuned_module->super.coll_allgather = ompi_coll_tuned_allgather_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, ALLGATHERV,
tuned_module->super.coll_allgatherv = ompi_coll_tuned_allgatherv_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, ALLREDUCE,
tuned_module->super.coll_allreduce = ompi_coll_tuned_allreduce_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, ALLTOALL,
tuned_module->super.coll_alltoall = ompi_coll_tuned_alltoall_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, ALLTOALLV,
tuned_module->super.coll_alltoallv = ompi_coll_tuned_alltoallv_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, ALLTOALLW,
tuned_module->super.coll_alltoallw = NULL);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, BARRIER,
tuned_module->super.coll_barrier = ompi_coll_tuned_barrier_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, BCAST,
tuned_module->super.coll_bcast = ompi_coll_tuned_bcast_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, EXSCAN,
tuned_module->super.coll_exscan = ompi_coll_tuned_exscan_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, GATHER,
tuned_module->super.coll_gather = ompi_coll_tuned_gather_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, GATHERV,
tuned_module->super.coll_gatherv = NULL);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, REDUCE,
tuned_module->super.coll_reduce = ompi_coll_tuned_reduce_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, REDUCESCATTER,
tuned_module->super.coll_reduce_scatter = ompi_coll_tuned_reduce_scatter_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, REDUCESCATTERBLOCK,
tuned_module->super.coll_reduce_scatter_block = ompi_coll_tuned_reduce_scatter_block_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, SCAN,
tuned_module->super.coll_scan = ompi_coll_tuned_scan_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, SCATTER,
tuned_module->super.coll_scatter = ompi_coll_tuned_scatter_intra_dec_dynamic);
COLL_TUNED_EXECUTE_IF_DYNAMIC(tuned_module, SCATTERV,
tuned_module->super.coll_scatterv = NULL);
}
/* general n fan out tree */
data->cached_ntree = NULL;
/* binary tree */
data->cached_bintree = NULL;
/* binomial tree */
data->cached_bmtree = NULL;
/* binomial tree */
data->cached_in_order_bmtree = NULL;
/* k-nomial tree */
data->cached_kmtree = NULL;
/* chains (fanout followed by pipelines) */
data->cached_chain = NULL;
/* standard pipeline */
data->cached_pipeline = NULL;
/* in-order binary tree */
data->cached_in_order_bintree = NULL;
/* All done */
tuned_module->super.base_data = data;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_init Tuned is in use"));
return OMPI_SUCCESS;
}
|
