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
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
315
316
317
|
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2007 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) 2009 IBM Corporation. All rights reserved.
* Copyright (c) 2009 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_PML_CSUM_H
#define MCA_PML_CSUM_H
#include "ompi_config.h"
#include "ompi/class/ompi_free_list.h"
#include "ompi/request/request.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/pml/base/pml_base_request.h"
#include "ompi/mca/pml/base/pml_base_bsend.h"
#include "ompi/mca/pml/base/pml_base_sendreq.h"
#include "ompi/datatype/ompi_datatype.h"
#include "pml_csum_hdr.h"
#include "ompi/mca/bml/base/base.h"
#include "ompi/proc/proc.h"
#include "ompi/mca/allocator/base/base.h"
BEGIN_C_DECLS
/**
* CSUM PML module
*/
struct mca_pml_csum_t {
mca_pml_base_module_t super;
int free_list_num; /* initial size of free list */
int free_list_max; /* maximum size of free list */
int free_list_inc; /* number of elements to grow free list */
size_t send_pipeline_depth;
size_t recv_pipeline_depth;
size_t rdma_put_retries_limit;
int max_rdma_per_request;
int max_send_per_range;
bool leave_pinned;
int leave_pinned_pipeline;
/* lock queue access */
opal_mutex_t lock;
/* free lists */
ompi_free_list_t rdma_frags;
ompi_free_list_t recv_frags;
ompi_free_list_t pending_pckts;
ompi_free_list_t buffers;
ompi_free_list_t send_ranges;
/* list of pending operations */
opal_list_t pckt_pending;
opal_list_t send_pending;
opal_list_t recv_pending;
opal_list_t rdma_pending;
/* List of pending fragments without a matching communicator */
opal_list_t non_existing_communicator_pending;
bool enabled;
char* allocator_name;
mca_allocator_base_module_t* allocator;
uint32_t unexpected_limit;
};
typedef struct mca_pml_csum_t mca_pml_csum_t;
extern mca_pml_csum_t mca_pml_csum;
/*
* PML interface functions.
*/
extern int mca_pml_csum_add_comm(
struct ompi_communicator_t* comm
);
extern int mca_pml_csum_del_comm(
struct ompi_communicator_t* comm
);
extern int mca_pml_csum_add_procs(
struct ompi_proc_t **procs,
size_t nprocs
);
extern int mca_pml_csum_del_procs(
struct ompi_proc_t **procs,
size_t nprocs
);
extern int mca_pml_csum_enable( bool enable );
extern int mca_pml_csum_progress(void);
extern int mca_pml_csum_iprobe( int dst,
int tag,
struct ompi_communicator_t* comm,
int *matched,
ompi_status_public_t* status );
extern int mca_pml_csum_probe( int dst,
int tag,
struct ompi_communicator_t* comm,
ompi_status_public_t* status );
extern int mca_pml_csum_isend_init( void *buf,
size_t count,
ompi_datatype_t *datatype,
int dst,
int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm,
struct ompi_request_t **request );
extern int mca_pml_csum_isend( void *buf,
size_t count,
ompi_datatype_t *datatype,
int dst,
int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm,
struct ompi_request_t **request );
extern int mca_pml_csum_send( void *buf,
size_t count,
ompi_datatype_t *datatype,
int dst,
int tag,
mca_pml_base_send_mode_t mode,
struct ompi_communicator_t* comm );
extern int mca_pml_csum_irecv_init( void *buf,
size_t count,
ompi_datatype_t *datatype,
int src,
int tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request );
extern int mca_pml_csum_irecv( void *buf,
size_t count,
ompi_datatype_t *datatype,
int src,
int tag,
struct ompi_communicator_t* comm,
struct ompi_request_t **request );
extern int mca_pml_csum_recv( void *buf,
size_t count,
ompi_datatype_t *datatype,
int src,
int tag,
struct ompi_communicator_t* comm,
ompi_status_public_t* status );
extern int mca_pml_csum_dump( struct ompi_communicator_t* comm,
int verbose );
extern int mca_pml_csum_start( size_t count,
ompi_request_t** requests );
extern int mca_pml_csum_ft_event( int state );
END_C_DECLS
struct mca_pml_csum_pckt_pending_t {
ompi_free_list_item_t super;
ompi_proc_t* proc;
mca_pml_csum_hdr_t hdr;
struct mca_bml_base_btl_t *bml_btl;
uint8_t order;
};
typedef struct mca_pml_csum_pckt_pending_t mca_pml_csum_pckt_pending_t;
OBJ_CLASS_DECLARATION(mca_pml_csum_pckt_pending_t);
#define MCA_PML_CSUM_PCKT_PENDING_ALLOC(pckt,rc) \
do { \
ompi_free_list_item_t* item; \
OMPI_FREE_LIST_WAIT(&mca_pml_csum.pending_pckts, item, rc); \
pckt = (mca_pml_csum_pckt_pending_t*)item; \
} while (0)
#define MCA_PML_CSUM_PCKT_PENDING_RETURN(pckt) \
do { \
/* return packet */ \
OMPI_FREE_LIST_RETURN(&mca_pml_csum.pending_pckts, \
(ompi_free_list_item_t*)pckt); \
} while(0)
#define MCA_PML_CSUM_ADD_FIN_TO_PENDING(P, D, B, O, S) \
do { \
mca_pml_csum_pckt_pending_t *_pckt; \
int _rc; \
\
MCA_PML_CSUM_PCKT_PENDING_ALLOC(_pckt,_rc); \
_pckt->hdr.hdr_common.hdr_type = MCA_PML_CSUM_HDR_TYPE_FIN; \
_pckt->hdr.hdr_fin.hdr_des.pval = (D); \
_pckt->hdr.hdr_fin.hdr_fail = (S); \
_pckt->proc = (P); \
_pckt->bml_btl = (B); \
_pckt->order = (O); \
OPAL_THREAD_LOCK(&mca_pml_csum.lock); \
opal_list_append(&mca_pml_csum.pckt_pending, \
(opal_list_item_t*)_pckt); \
OPAL_THREAD_UNLOCK(&mca_pml_csum.lock); \
} while(0)
int mca_pml_csum_send_fin(ompi_proc_t* proc, mca_bml_base_btl_t* bml_btl,
void *hdr_des, uint8_t order, uint32_t status);
/* This function tries to resend FIN/ACK packets from pckt_pending queue.
* Packets are added to the queue when sending of FIN or ACK is failed due to
* resource unavailability. bml_btl passed to the function doesn't represents
* packet's destination, it represents BTL on which resource was freed, so only
* this BTL should be considered for resending packets */
void mca_pml_csum_process_pending_packets(mca_bml_base_btl_t* bml_btl);
/* This function retries failed PUT/GET operations on frag. When RDMA operation
* cannot be accomplished for some reason, frag is put on the rdma_pending list.
* Later the operation is retried. The destination of RDMA operation is stored
* inside the frag structure */
void mca_pml_csum_process_pending_rdma(void);
#define MCA_PML_CSUM_PROGRESS_PENDING(bml_btl) \
do { \
if(opal_list_get_size(&mca_pml_csum.pckt_pending)) \
mca_pml_csum_process_pending_packets(bml_btl); \
if(opal_list_get_size(&mca_pml_csum.recv_pending)) \
mca_pml_csum_recv_request_process_pending(); \
if(opal_list_get_size(&mca_pml_csum.send_pending)) \
mca_pml_csum_send_request_process_pending(bml_btl); \
if(opal_list_get_size(&mca_pml_csum.rdma_pending)) \
mca_pml_csum_process_pending_rdma(); \
} while (0)
/*
* Compute the total number of bytes on supplied descriptor
*/
#define MCA_PML_CSUM_COMPUTE_SEGMENT_LENGTH(segments, count, hdrlen, length) \
do { \
size_t i; \
\
for( i = 0; i < count; i++ ) { \
length += segments[i].seg_len; \
} \
length -= hdrlen; \
} while(0)
/* represent BTL chosen for sending request */
struct mca_pml_csum_com_btl_t {
mca_bml_base_btl_t *bml_btl;
struct mca_mpool_base_registration_t* btl_reg;
size_t length;
};
typedef struct mca_pml_csum_com_btl_t mca_pml_csum_com_btl_t;
int mca_pml_csum_com_btl_comp(const void *v1, const void *v2);
/* Calculate what percentage of a message to send through each BTL according to
* relative weight */
static inline void
mca_pml_csum_calc_weighted_length( mca_pml_csum_com_btl_t *btls, int num_btls, size_t size,
double weight_total )
{
int i;
size_t length_left;
/* shortcut for common case for only one BTL */
if( OPAL_LIKELY(1 == num_btls) ) {
btls[0].length = size;
return;
}
/* sort BTLs according of their weights so BTLs with smaller weight will
* not hijack all of the traffic */
qsort( btls, num_btls, sizeof(mca_pml_csum_com_btl_t),
mca_pml_csum_com_btl_comp );
for(length_left = size, i = 0; i < num_btls; i++) {
mca_bml_base_btl_t* bml_btl = btls[i].bml_btl;
size_t length = 0;
if( OPAL_UNLIKELY(0 != length_left) ) {
length = (length_left > bml_btl->btl->btl_eager_limit)?
((size_t)(size * (bml_btl->btl_weight / weight_total))) :
length_left;
if(length > length_left)
length = length_left;
length_left -= length;
}
btls[i].length = length;
}
/* account for rounding errors */
btls[0].length += length_left;
}
#endif
|
