model-net.c 19.7 KB
Newer Older
1
/*
Philip Carns's avatar
Philip Carns committed
2
 * Copyright (C) 2013 University of Chicago.
3
 * See COPYRIGHT notice in top-level directory.
Philip Carns's avatar
Philip Carns committed
4
 *
5
 */
Philip Carns's avatar
Philip Carns committed
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;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
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_custom_method;
25
extern struct model_net_method slimfly_method;
26
extern struct model_net_method fattree_method;
27
extern struct model_net_method dragonfly_router_method;
28
extern struct model_net_method dragonfly_custom_router_method;
Philip Carns's avatar
Philip Carns committed
29
extern struct model_net_method loggp_method;
30

31 32 33 34 35 36 37 38 39 40 41 42
#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

43
/* Global array initialization, terminated with a NULL entry */
44 45 46 47 48
#define X(a,b,c,d) d,
struct model_net_method* method_array[] = { 
    NETWORK_DEF
};
#undef X
49

50
// counter and offset for the MN_START_SEQ / MN_END_SEQ macros
51
int mn_in_sequence = 0;
52 53
tw_stime mn_msg_offset = 0.0;

54 55
// message parameters for use via model_net_set_msg_param
static int is_msg_params_set[MAX_MN_MSG_PARAM_TYPES];
56 57
static mn_sched_params sched_params; // MN_MSG_PARAM_SCHED
static tw_stime start_time_param; // MN_MSG_PARAM_START_TIME
58

59 60 61 62 63 64 65 66 67 68 69
// 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++){
70
            char const *nm = lpgroup->lptypes[lpt].name.ptr;
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
            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]){
91 92 93 94
            // some don't need configuration (dragonfly router is covered by
            // dragonfly)
            if (method_array[i]->mn_configure != NULL)
                method_array[i]->mn_configure();
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
            (*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);

134 135 136 137
    // init the per-msg params here
    memset(is_msg_params_set, 0,
            MAX_MN_MSG_PARAM_TYPES*sizeof(*is_msg_params_set));

138
    return ids;
139 140
}

141 142 143
int model_net_get_id(char *name){
    int i;
    for(i=0; method_array[i] != NULL; i++) {
144
        if(strcmp(model_net_method_names[i], name) == 0) {
145 146 147 148 149 150
            return i;
        }
    }
    return -1;
}

151 152 153
void model_net_write_stats(tw_lpid lpid, struct mn_stats* stat)
{
    int ret;
154
    char id[19+CATEGORY_NAME_MAX+1];
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
    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);
}

203
struct mn_stats* model_net_find_stats(char const * category, mn_stats mn_stats_array[])
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
{
    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]);
}

233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256
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++;
257 258 259 260
        /* 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);
261 262 263 264 265 266 267 268 269 270 271
        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(
272
        int net_id,
273 274
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
275
        char const * category, 
276 277 278 279 280
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        int is_pull,
        tw_stime offset,
        int remote_event_size,
281
        void const * remote_event,
282
        int self_event_size,
283
        void const * self_event,
284
        tw_lp *sender) {
285

286 287 288 289 290 291
    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);
292
        return -1;
293
    }
294

295 296 297 298 299
    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);

300 301 302 303 304
    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);

305
    tw_stime poffset = codes_local_latency(sender);
306
    if (mn_in_sequence){
307 308 309 310
        tw_stime tmp = mn_msg_offset;
        mn_msg_offset += poffset;
        poffset += tmp;
    }
311 312

    tw_event *e = tw_event_new(src_mn_lp, poffset+offset, sender);
313 314

    model_net_wrap_msg *m = tw_event_data(e);
315
    msg_set_header(model_net_base_magic, MN_BASE_NEW_MSG, sender->gid, &m->h);
316 317

    // set the request struct 
318
    model_net_request *r = &m->msg.m_base.req;
319
    r->final_dest_lp = final_dest_lp;
320
    r->dest_mn_lp = dest_mn_lp;
321
    r->src_lp = sender->gid;
322 323 324 325 326 327 328 329 330 331
    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;
332 333 334 335
    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';
336

337 338 339 340 341
    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);

342 343
    // this is an outgoing message
    m->msg.m_base.is_from_remote = 0;
344

345 346 347 348
    // 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
349
        model_net_sched_set_default_params(&m->msg.m_base.sched_params);
350 351 352 353
    // once params are set, clear the flags 
    memset(is_msg_params_set, 0,
            MAX_MN_MSG_PARAM_TYPES*sizeof(*is_msg_params_set));

354 355 356 357 358 359 360 361
    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);
    }
362

363 364
    //print_base(m);
    tw_event_send(e);
365 366

    return 1;
367 368 369
}
static void model_net_event_impl_base_rc(tw_lp *sender){
    codes_local_latency_reverse(sender);
370
}
371

372
model_net_event_return model_net_event(
373
    int net_id,
374
    char const * category, 
375 376 377 378
    tw_lpid final_dest_lp, 
    uint64_t message_size, 
    tw_stime offset,
    int remote_event_size,
379
    void const * remote_event,
380
    int self_event_size,
381
    void const * self_event,
382 383
    tw_lp *sender)
{
384 385 386 387
    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);
388 389
}

390
model_net_event_return model_net_event_annotated(
391
        int net_id,
392 393
        char const * annotation,
        char const * category, 
394 395 396 397
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        tw_stime offset,
        int remote_event_size,
398
        void const * remote_event,
399
        int self_event_size,
400
        void const * self_event,
401
        tw_lp *sender){
402
    struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, false);
403
    return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
404 405
            message_size, 0, offset, remote_event_size, remote_event,
            self_event_size, self_event, sender);
406 407
}

408
model_net_event_return model_net_event_mctx(
409 410 411 412 413 414 415 416 417 418 419 420
        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){
421 422 423 424
    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);
425 426
}

427
model_net_event_return model_net_pull_event(
428
        int net_id,
429
        char const *category,
430 431 432 433
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
434
        void const *self_event,
435 436 437
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
438 439 440
    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);
441 442
}

443
model_net_event_return model_net_pull_event_annotated(
444
        int net_id,
445 446
        char const * annotation,
        char const *category,
447 448 449 450
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
451
        void const *self_event,
452 453 454
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
455
    struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, false);
456
    return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
457 458
            message_size, 1, offset, self_event_size, self_event, 0, NULL,
            sender);
459 460
}

461
model_net_event_return model_net_pull_event_mctx(
462
        int net_id,
463 464
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
465 466 467 468 469 470 471 472 473
        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 */
474 475
    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,
476 477 478
            self_event, 0, NULL, sender);
}

479 480 481 482 483 484 485 486
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);
}

487 488 489 490
void model_net_event_rc(
        int net_id,
        tw_lp *sender,
        uint64_t message_size){
491 492
    (void)net_id;
    (void)message_size;
493
    model_net_event_impl_base_rc(sender);
494 495 496 497 498
}

void model_net_pull_event_rc(
        int net_id,
        tw_lp *sender) {
499
    (void)net_id;
500
    model_net_event_impl_base_rc(sender);
501 502
}

503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
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;
519 520 521 522 523 524 525 526 527 528 529
        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;
530 531 532 533 534
        default:
            tw_error(TW_LOC, "unknown or unsupported msg_param_type");
    }
}

535
/* returns the message size, can be either simplenet, dragonfly or torus message size*/
536
int model_net_get_msg_sz(int net_id)
537
{
538 539 540 541
    (void)net_id;
    // TODO: Add checks on network name
    // TODO: Add dragonfly and torus network models
    return sizeof(model_net_wrap_msg);
542
#if 0
543 544 545 546 547
    if(net_id < 0 || net_id >= MAX_NETS)
    {
        printf("%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
        exit(-1);
    }
548

549
    return method_array[net_id]->mn_get_msg_sz();
550
#endif
551 552 553
}

/* returns the packet size in the modelnet struct */
554
uint64_t model_net_get_packet_size(int net_id)
555
{
556
  if(net_id < 0 || net_id >= MAX_NETS)
557 558 559 560 561 562 563
     {
       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?
}

564
/* This event does a collective operation call for model-net */
565
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)
566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585
{
  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);
}

586 587 588
/* returns lp type for modelnet */
const tw_lptype* model_net_get_lp_type(int net_id)
{
589
    if(net_id < 0 || net_id >= MAX_NETS)
590 591 592 593 594 595 596 597 598 599
     {
       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();
}

600 601
void model_net_report_stats(int net_id)
{
602
  if(net_id < 0 || net_id >= MAX_NETS)
603 604 605 606 607 608 609
  {
    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
610 611
   method_array[net_id]->mn_report_stats();
   return;
612
}
613

614
tw_lpid model_net_find_local_device(
615 616 617
        int          net_id,
        const char * annotation,
        int          ignore_annotations,
618 619 620 621 622 623 624 625 626
        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;
627
    }
628
    return model_net_find_local_device_mctx(net_id, mc_p, sender_gid);
629 630
}

631 632 633 634
tw_lpid model_net_find_local_device_mctx(
        int net_id,
        struct codes_mctx const * map_ctx,
        tw_lpid sender_gid)
635
{
636 637
    return codes_mctx_to_lpid(map_ctx, model_net_lp_config_names[net_id],
            sender_gid);
638 639
}

640 641 642 643 644 645 646 647
/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */