model-net.c 19.5 KB
Newer Older
1
/*
2
 * Copyright (C) 2013 University of Chicago.
3
 * See COPYRIGHT notice in top-level directory.
4
 *
5
 */
6

7 8 9
#include <string.h>
#include <assert.h>

10
#include "codes/model-net.h"
11
#include "codes/model-net-method.h"
12
#include "codes/model-net-lp.h"
13
#include "codes/model-net-sched.h"
14
#include "codes/codes.h"
15
#include <codes/codes_mapping.h>
16

17 18 19 20
#define STR_SIZE 16
#define PROC_TIME 10.0

extern struct model_net_method simplenet_method;
21
extern struct model_net_method simplep2p_method;
22
extern struct model_net_method torus_method;
23
extern struct model_net_method dragonfly_method;
24
extern struct model_net_method dragonfly_router_method;
Philip Carns's avatar
Philip Carns committed
25
extern struct model_net_method loggp_method;
26

27 28 29 30 31 32 33 34 35 36 37 38
#define X(a,b,c,d) b,
char * model_net_lp_config_names[] = {
    NETWORK_DEF
};
#undef X

#define X(a,b,c,d) c,
char * model_net_method_names[] = {
    NETWORK_DEF
};
#undef X

39
/* Global array initialization, terminated with a NULL entry */
40 41 42 43 44
#define X(a,b,c,d) d,
struct model_net_method* method_array[] = { 
    NETWORK_DEF
};
#undef X
45

46
// counter and offset for the MN_START_SEQ / MN_END_SEQ macros
47
int mn_in_sequence = 0;
48 49
tw_stime mn_msg_offset = 0.0;

50 51
// message parameters for use via model_net_set_msg_param
static int is_msg_params_set[MAX_MN_MSG_PARAM_TYPES];
52 53
static mn_sched_params sched_params; // MN_MSG_PARAM_SCHED
static tw_stime start_time_param; // MN_MSG_PARAM_START_TIME
54

55 56 57 58 59 60 61 62 63 64 65
// global listing of lp types found by model_net_register
// - needs to be held between the register and configure calls
static int do_config_nets[MAX_NETS];

void model_net_register(){
    // first set up which networks need to be registered, then pass off to base
    // LP to do its thing
    memset(do_config_nets, 0, MAX_NETS * sizeof(*do_config_nets));
    for (int grp = 0; grp < lpconf.lpgroups_count; grp++){
        config_lpgroup_t *lpgroup = &lpconf.lpgroups[grp];
        for (int lpt = 0; lpt < lpgroup->lptypes_count; lpt++){
66
            char const *nm = lpgroup->lptypes[lpt].name.ptr;
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
            for (int n = 0; n < MAX_NETS; n++){
                if (!do_config_nets[n] && 
                        strcmp(model_net_lp_config_names[n], nm) == 0){
                    do_config_nets[n] = 1;
                    break;
                }
            }
        }
    }
    model_net_base_register(do_config_nets);
}

int* model_net_configure(int *id_count){
    // first call the base LP configure, which sets up the general parameters
    model_net_base_configure();

    // do network-specific configures
    *id_count = 0;
    for (int i = 0; i < MAX_NETS; i++) {
        if (do_config_nets[i]){
87 88 89 90
            // some don't need configuration (dragonfly router is covered by
            // dragonfly)
            if (method_array[i]->mn_configure != NULL)
                method_array[i]->mn_configure();
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
            (*id_count)++;
        }
    }

    // allocate the output
    int *ids = malloc(*id_count * sizeof(int));
    // read the ordering provided by modelnet_order
    char **values;
    size_t length;
    int ret = configuration_get_multivalue(&config, "PARAMS", "modelnet_order",
            NULL, &values, &length);
    if (ret != 1){
        tw_error(TW_LOC, "unable to read PARAMS:modelnet_order variable\n");
    }
    if (length != (size_t) *id_count){
        tw_error(TW_LOC, "number of networks in PARAMS:modelnet_order "
                "do not match number in LPGROUPS\n");
    }
    // set the index
    for (int i = 0; i < *id_count; i++){
        ids[i] = -1;
        for (int n = 0; n < MAX_NETS; n++){
            if (strcmp(values[i], model_net_method_names[n]) == 0){
                if (!do_config_nets[n]){
                    tw_error(TW_LOC, "network in PARAMS:modelnet_order not "
                            "present in LPGROUPS: %s\n", values[i]);
                }
                ids[i] = n;
                break;
            }
        }
        if (ids[i] == -1){
            tw_error(TW_LOC, "unknown network in PARAMS:modelnet_order: %s\n",
                    values[i]);
        }
        free(values[i]);
    }
    free(values);

130 131 132 133
    // init the per-msg params here
    memset(is_msg_params_set, 0,
            MAX_MN_MSG_PARAM_TYPES*sizeof(*is_msg_params_set));

134
    return ids;
135 136
}

137 138 139
int model_net_get_id(char *name){
    int i;
    for(i=0; method_array[i] != NULL; i++) {
140
        if(strcmp(model_net_method_names[i], name) == 0) {
141 142 143 144 145 146
            return i;
        }
    }
    return -1;
}

147 148 149
void model_net_write_stats(tw_lpid lpid, struct mn_stats* stat)
{
    int ret;
150
    char id[19+CATEGORY_NAME_MAX+1];
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
    char data[1024];

    sprintf(id, "model-net-category-%s", stat->category);
    sprintf(data, "lp:%ld\tsend_count:%ld\tsend_bytes:%ld\tsend_time:%f\t" 
        "recv_count:%ld\trecv_bytes:%ld\trecv_time:%f\tmax_event_size:%ld\n",
        (long)lpid,
        stat->send_count,
        stat->send_bytes,
        stat->send_time,
        stat->recv_count,
        stat->recv_bytes,
        stat->recv_time,
        stat->max_event_size);

    ret = lp_io_write(lpid, id, strlen(data), data);
    assert(ret == 0);

    return;
}

void model_net_print_stats(tw_lpid lpid, mn_stats mn_stats_array[])
{

    int i;
    struct mn_stats all;

    memset(&all, 0, sizeof(all));
    sprintf(all.category, "all");

    for(i=0; i<CATEGORY_MAX; i++)
    {
        if(strlen(mn_stats_array[i].category) > 0)
        {
            all.send_count += mn_stats_array[i].send_count;
            all.send_bytes += mn_stats_array[i].send_bytes;
            all.send_time += mn_stats_array[i].send_time;
            all.recv_count += mn_stats_array[i].recv_count;
            all.recv_bytes += mn_stats_array[i].recv_bytes;
            all.recv_time += mn_stats_array[i].recv_time;
            if(mn_stats_array[i].max_event_size > all.max_event_size)
                all.max_event_size = mn_stats_array[i].max_event_size;

            model_net_write_stats(lpid, &mn_stats_array[i]);
        }
    }
    model_net_write_stats(lpid, &all);
}

199
struct mn_stats* model_net_find_stats(char const * category, mn_stats mn_stats_array[])
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
{
    int i;
    int new_flag = 0;
    int found_flag = 0;

    for(i=0; i<CATEGORY_MAX; i++)
    {
        if(strlen(mn_stats_array[i].category) == 0)
        {
            found_flag = 1;
            new_flag = 1;
            break;
        }
        if(strcmp(category, mn_stats_array[i].category) == 0)
        {
            found_flag = 1;
            new_flag = 0;
            break;
        }
    }
    assert(found_flag);

    if(new_flag)
    {
        strcpy(mn_stats_array[i].category, category);
    }
    return(&mn_stats_array[i]);
}

229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252
static model_net_event_return model_net_noop_event(
        tw_lpid final_dest_lp,
        int is_pull,
        tw_stime offset,
        int remote_event_size,
        void const * remote_event,
        int self_event_size,
        void const * self_event,
        tw_lp *sender)
{
    model_net_event_return num_rng_calls = 0;
    tw_stime poffset = mn_in_sequence ? mn_msg_offset : 0.0;

    if (self_event_size && self_event != NULL) {
        poffset += codes_local_latency(sender);
        num_rng_calls++;
        tw_event *e = tw_event_new(sender->gid, poffset+offset, sender);
        memcpy(tw_event_data(e), self_event, self_event_size);
        tw_event_send(e);
    }

    if (remote_event_size && remote_event != NULL) {
        poffset += codes_local_latency(sender);
        num_rng_calls++;
253 254 255 256
        /* special case - in a "pull" event, the "remote" message is actually
         * to self */
        tw_event *e = tw_event_new(is_pull ? sender->gid : final_dest_lp,
                poffset+offset, sender);
257 258 259 260 261 262 263 264 265 266 267
        memcpy(tw_event_data(e), remote_event, remote_event_size);
        tw_event_send(e);
    }

    if (mn_in_sequence)
        mn_msg_offset += poffset;

    return num_rng_calls;
}

static model_net_event_return model_net_event_impl_base(
268
        int net_id,
269 270
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
271
        char const * category, 
272 273 274 275 276
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        int is_pull,
        tw_stime offset,
        int remote_event_size,
277
        void const * remote_event,
278
        int self_event_size,
279
        void const * self_event,
280
        tw_lp *sender) {
281

282 283 284 285 286 287
    if (remote_event_size + self_event_size + sizeof(model_net_wrap_msg) 
            > g_tw_msg_sz){
        tw_error(TW_LOC, "Error: model_net trying to transmit an event of size "
                         "%d but ROSS is configured for events of size %zd\n",
                         remote_event_size+self_event_size+sizeof(model_net_wrap_msg),
                         g_tw_msg_sz);
288
        return -1;
289
    }
290

291 292 293 294 295
    tw_lpid src_mn_lp = model_net_find_local_device_mctx(net_id, send_map_ctx,
            sender->gid);
    tw_lpid dest_mn_lp = model_net_find_local_device_mctx(net_id, recv_map_ctx,
            final_dest_lp);

296 297 298 299 300
    if (src_mn_lp == dest_mn_lp)
        return model_net_noop_event(final_dest_lp, is_pull, offset,
                remote_event_size, remote_event, self_event_size, self_event,
                sender);

301
    tw_stime poffset = codes_local_latency(sender);
302
    if (mn_in_sequence){
303 304 305 306
        tw_stime tmp = mn_msg_offset;
        mn_msg_offset += poffset;
        poffset += tmp;
    }
307 308

    tw_event *e = tw_event_new(src_mn_lp, poffset+offset, sender);
309 310

    model_net_wrap_msg *m = tw_event_data(e);
311
    msg_set_header(model_net_base_magic, MN_BASE_NEW_MSG, sender->gid, &m->h);
312 313

    // set the request struct 
314
    model_net_request *r = &m->msg.m_base.req;
315
    r->final_dest_lp = final_dest_lp;
316
    r->dest_mn_lp = dest_mn_lp;
317
    r->src_lp = sender->gid;
318 319 320 321 322 323 324 325 326 327
    r->is_pull = is_pull;
    if (r->is_pull) {
        r->msg_size = PULL_MSG_SIZE;
        r->pull_size = message_size;
    }
    else {
        r->msg_size = message_size;
        r->pull_size = 0;
    }
    r->net_id = net_id;
328 329 330 331
    r->remote_event_size = remote_event_size;
    r->self_event_size = self_event_size;
    strncpy(r->category, category, CATEGORY_NAME_MAX-1);
    r->category[CATEGORY_NAME_MAX-1]='\0';
332

333 334 335 336 337
    if (is_msg_params_set[MN_MSG_PARAM_START_TIME])
        r->msg_start_time = start_time_param;
    else
        r->msg_start_time = tw_now(sender);

338 339
    // this is an outgoing message
    m->msg.m_base.is_from_remote = 0;
340

341 342 343 344
    // set the msg-specific params
    if (is_msg_params_set[MN_SCHED_PARAM_PRIO])
        m->msg.m_base.sched_params = sched_params;
    else // set the default
345
        model_net_sched_set_default_params(&m->msg.m_base.sched_params);
346 347 348 349
    // once params are set, clear the flags 
    memset(is_msg_params_set, 0,
            MAX_MN_MSG_PARAM_TYPES*sizeof(*is_msg_params_set));

350 351 352 353 354 355 356 357
    void *e_msg = (m+1);
    if (remote_event_size > 0){
        memcpy(e_msg, remote_event, remote_event_size);
        e_msg = (char*)e_msg + remote_event_size; 
    }
    if (self_event_size > 0){
        memcpy(e_msg, self_event, self_event_size);
    }
358

359 360
    //print_base(m);
    tw_event_send(e);
361 362

    return 1;
363 364 365
}
static void model_net_event_impl_base_rc(tw_lp *sender){
    codes_local_latency_reverse(sender);
366
}
367

368
model_net_event_return model_net_event(
369
    int net_id,
370
    char const * category, 
371 372 373 374
    tw_lpid final_dest_lp, 
    uint64_t message_size, 
    tw_stime offset,
    int remote_event_size,
375
    void const * remote_event,
376
    int self_event_size,
377
    void const * self_event,
378 379
    tw_lp *sender)
{
380 381 382 383
    return model_net_event_impl_base(net_id, CODES_MCTX_DEFAULT,
            CODES_MCTX_DEFAULT, category, final_dest_lp, message_size, 0,
            offset, remote_event_size, remote_event, self_event_size,
            self_event, sender);
384 385
}

386
model_net_event_return model_net_event_annotated(
387
        int net_id,
388 389
        char const * annotation,
        char const * category, 
390 391 392 393
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        tw_stime offset,
        int remote_event_size,
394
        void const * remote_event,
395
        int self_event_size,
396
        void const * self_event,
397
        tw_lp *sender){
398
    struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, true);
399
    return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
400 401
            message_size, 0, offset, remote_event_size, remote_event,
            self_event_size, self_event, sender);
402 403
}

404
model_net_event_return model_net_event_mctx(
405 406 407 408 409 410 411 412 413 414 415 416
        int net_id,
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
        char const * category, 
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        tw_stime offset,
        int remote_event_size,
        void const * remote_event,
        int self_event_size,
        void const * self_event,
        tw_lp *sender){
417 418 419 420
    return model_net_event_impl_base(net_id, send_map_ctx, recv_map_ctx,
            category, final_dest_lp, message_size, 0, offset,
            remote_event_size, remote_event, self_event_size, self_event,
            sender);
421 422
}

423
model_net_event_return model_net_pull_event(
424
        int net_id,
425
        char const *category,
426 427 428 429
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
430
        void const *self_event,
431 432 433
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
434 435 436
    return model_net_event_impl_base(net_id, CODES_MCTX_DEFAULT,
            CODES_MCTX_DEFAULT, category, final_dest_lp, message_size, 1,
            offset, self_event_size, self_event, 0, NULL, sender);
437 438
}

439
model_net_event_return model_net_pull_event_annotated(
440
        int net_id,
441 442
        char const * annotation,
        char const *category,
443 444 445 446
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
447
        void const *self_event,
448 449 450
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
451
    struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, true);
452
    return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
453 454
            message_size, 1, offset, self_event_size, self_event, 0, NULL,
            sender);
455 456
}

457
model_net_event_return model_net_pull_event_mctx(
458
        int net_id,
459 460
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
461 462 463 464 465 466 467 468 469
        char const *category,
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
        void const *self_event,
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
470 471
    return model_net_event_impl_base(net_id, send_map_ctx, recv_map_ctx,
            category, final_dest_lp, message_size, 1, offset, self_event_size,
472 473 474
            self_event, 0, NULL, sender);
}

475 476 477 478 479 480 481 482
void model_net_event_rc2(
        tw_lp *sender,
        model_net_event_return const * ret)
{
    for (int i = 0; i < *ret; i++)
        codes_local_latency_reverse(sender);
}

483 484 485 486
void model_net_event_rc(
        int net_id,
        tw_lp *sender,
        uint64_t message_size){
487 488
    (void)net_id;
    (void)message_size;
489
    model_net_event_impl_base_rc(sender);
490 491 492 493 494
}

void model_net_pull_event_rc(
        int net_id,
        tw_lp *sender) {
495
    (void)net_id;
496
    model_net_event_impl_base_rc(sender);
497 498
}

499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514
void model_net_set_msg_param(
        enum msg_param_type type,
        int sub_type,
        const void * params){
    switch(type){
        case MN_MSG_PARAM_SCHED:
            is_msg_params_set[MN_MSG_PARAM_SCHED] = 1;
            switch(sub_type){
                case MN_SCHED_PARAM_PRIO:
                    sched_params.prio = *(int*)params;
                    break;
                default:
                    tw_error(TW_LOC, "unknown or unsupported "
                            "MN_MSG_PARAM_SCHED parameter type");
            }
            break;
515 516 517 518 519 520 521 522 523 524 525
        case MN_MSG_PARAM_START_TIME:
            is_msg_params_set[MN_MSG_PARAM_START_TIME] = 1;
            switch(sub_type){
                case MN_MSG_PARAM_START_TIME_VAL:
                    start_time_param = *(tw_stime*)params;
                    break;
                default:
                    tw_error(TW_LOC, "unknown or unsupported "
                            "MN_MSG_PARAM_START_TIME parameter type");
            }
            break;
526 527 528 529 530
        default:
            tw_error(TW_LOC, "unknown or unsupported msg_param_type");
    }
}

531
/* returns the message size, can be either simplenet, dragonfly or torus message size*/
532
int model_net_get_msg_sz(int net_id)
533
{
534 535 536 537
    (void)net_id;
    // TODO: Add checks on network name
    // TODO: Add dragonfly and torus network models
    return sizeof(model_net_wrap_msg);
538
#if 0
539 540 541 542 543
    if(net_id < 0 || net_id >= MAX_NETS)
    {
        printf("%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
        exit(-1);
    }
544

545
    return method_array[net_id]->mn_get_msg_sz();
546
#endif
547 548 549
}

/* returns the packet size in the modelnet struct */
550
uint64_t model_net_get_packet_size(int net_id)
551
{
552
  if(net_id < 0 || net_id >= MAX_NETS)
553 554 555 556 557 558 559
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }
  return method_array[net_id]->packet_size; // TODO: where to set the packet size?
}

560
/* This event does a collective operation call for model-net */
561
void model_net_event_collective(int net_id, char const * category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581
{
  if(net_id < 0 || net_id > MAX_NETS)
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }
  return method_array[net_id]->mn_collective_call(category, message_size, remote_event_size, remote_event, sender);
}

/* reverse event of the collective operation call */
void model_net_event_collective_rc(int net_id, int message_size, tw_lp* sender)
{
  if(net_id < 0 || net_id > MAX_NETS)
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }
  return method_array[net_id]->mn_collective_call_rc(message_size, sender);
}

582 583 584
/* returns lp type for modelnet */
const tw_lptype* model_net_get_lp_type(int net_id)
{
585
    if(net_id < 0 || net_id >= MAX_NETS)
586 587 588 589 590 591 592 593 594 595
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }

   // TODO: ADd checks by network names
   // Add dragonfly and torus network models
   return method_array[net_id]->mn_get_lp_type();
}

596 597
void model_net_report_stats(int net_id)
{
598
  if(net_id < 0 || net_id >= MAX_NETS)
599 600 601 602 603 604 605
  {
    fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
    exit(-1);
   }

     // TODO: ADd checks by network names
     //    // Add dragonfly and torus network models
606 607
   method_array[net_id]->mn_report_stats();
   return;
608
}
609

610
tw_lpid model_net_find_local_device(
611 612 613
        int          net_id,
        const char * annotation,
        int          ignore_annotations,
614 615 616 617 618 619 620 621 622
        tw_lpid      sender_gid)
{
    struct codes_mctx const * mc_p;
    struct codes_mctx mc;
    if (ignore_annotations)
        mc_p = CODES_MCTX_DEFAULT;
    else {
        mc = codes_mctx_set_group_modulo(annotation, ignore_annotations);
        mc_p = &mc;
623
    }
624
    return model_net_find_local_device_mctx(net_id, mc_p, sender_gid);
625 626
}

627 628 629 630
tw_lpid model_net_find_local_device_mctx(
        int net_id,
        struct codes_mctx const * map_ctx,
        tw_lpid sender_gid)
631
{
632 633
    return codes_mctx_to_lpid(map_ctx, model_net_lp_config_names[net_id],
            sender_gid);
634 635
}

636 637 638 639 640 641 642 643
/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */