model-net.c 19.9 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 21
#define STR_SIZE 16
#define PROC_TIME 10.0

extern struct model_net_method simplenet_method;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
22
extern struct model_net_method simplep2p_method;
23
extern struct model_net_method torus_method;
24
extern struct model_net_method dragonfly_method;
25
extern struct model_net_method dragonfly_custom_method;
26
extern struct model_net_method slimfly_method;
27
extern struct model_net_method fattree_method;
28
extern struct model_net_method dragonfly_router_method;
29
extern struct model_net_method dragonfly_custom_router_method;
Philip Carns's avatar
Philip Carns committed
30
extern struct model_net_method loggp_method;
31

32 33 34 35 36 37 38 39 40 41 42 43
#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

44
/* Global array initialization, terminated with a NULL entry */
45 46 47 48 49
#define X(a,b,c,d) d,
struct model_net_method* method_array[] = { 
    NETWORK_DEF
};
#undef X
50

51
// counter and offset for the MN_START_SEQ / MN_END_SEQ macros
52
int mn_in_sequence = 0;
53 54
tw_stime mn_msg_offset = 0.0;

55 56
// message parameters for use via model_net_set_msg_param
static int is_msg_params_set[MAX_MN_MSG_PARAM_TYPES];
57 58
static mn_sched_params sched_params; // MN_MSG_PARAM_SCHED
static tw_stime start_time_param; // MN_MSG_PARAM_START_TIME
59

60 61 62 63 64 65 66 67 68 69 70
// 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++){
71
            char const *nm = lpgroup->lptypes[lpt].name.ptr;
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
            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]){
92 93 94 95
            // some don't need configuration (dragonfly router is covered by
            // dragonfly)
            if (method_array[i]->mn_configure != NULL)
                method_array[i]->mn_configure();
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
            (*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);

135 136 137 138
    // init the per-msg params here
    memset(is_msg_params_set, 0,
            MAX_MN_MSG_PARAM_TYPES*sizeof(*is_msg_params_set));

139
    return ids;
140 141
}

142 143 144
int model_net_get_id(char *name){
    int i;
    for(i=0; method_array[i] != NULL; i++) {
145
        if(strcmp(model_net_method_names[i], name) == 0) {
146 147 148 149 150 151
            return i;
        }
    }
    return -1;
}

152 153 154
void model_net_write_stats(tw_lpid lpid, struct mn_stats* stat)
{
    int ret;
155
    char id[19+CATEGORY_NAME_MAX+1];
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
    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);
}

204
struct mn_stats* model_net_find_stats(char const * category, mn_stats mn_stats_array[])
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
{
    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]);
}

234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257
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++;
258 259 260 261
        /* 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);
262 263 264 265 266 267 268 269 270 271 272
        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(
273
        int net_id,
274 275
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
276
        char const * category, 
277 278 279 280 281
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        int is_pull,
        tw_stime offset,
        int remote_event_size,
282
        void const * remote_event,
283
        int self_event_size,
284
        void const * self_event,
285
        tw_lp *sender) {
286

287 288 289 290 291 292
    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);
293
        return -1;
294
    }
295

296 297 298 299 300
    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);

301 302 303 304 305
    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);

306
    tw_stime poffset = codes_local_latency(sender);
307
    if (mn_in_sequence){
308 309 310 311
        tw_stime tmp = mn_msg_offset;
        mn_msg_offset += poffset;
        poffset += tmp;
    }
312 313

    tw_event *e = tw_event_new(src_mn_lp, poffset+offset, sender);
314 315

    model_net_wrap_msg *m = tw_event_data(e);
316
    msg_set_header(model_net_base_magic, MN_BASE_NEW_MSG, sender->gid, &m->h);
317 318

    // set the request struct 
319
    model_net_request *r = &m->msg.m_base.req;
320
    r->final_dest_lp = final_dest_lp;
321
    r->dest_mn_lp = dest_mn_lp;
322
    r->src_lp = sender->gid;
323 324 325 326 327 328 329 330 331 332
    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;
333 334 335 336
    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';
337

338 339 340 341 342
    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);

343 344
    // this is an outgoing message
    m->msg.m_base.is_from_remote = 0;
345

346 347 348 349
    // 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
350
        model_net_sched_set_default_params(&m->msg.m_base.sched_params);
351 352 353 354
    // once params are set, clear the flags 
    memset(is_msg_params_set, 0,
            MAX_MN_MSG_PARAM_TYPES*sizeof(*is_msg_params_set));

355 356 357 358 359 360 361 362
    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);
    }
363

364 365
    //print_base(m);
    tw_event_send(e);
366 367

    return 1;
368 369 370
}
static void model_net_event_impl_base_rc(tw_lp *sender){
    codes_local_latency_reverse(sender);
371
}
372

373
model_net_event_return model_net_event(
374
    int net_id,
375
    char const * category, 
376 377 378 379
    tw_lpid final_dest_lp, 
    uint64_t message_size, 
    tw_stime offset,
    int remote_event_size,
380
    void const * remote_event,
381
    int self_event_size,
382
    void const * self_event,
383 384
    tw_lp *sender)
{
385 386 387 388
    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);
389 390
}

391
model_net_event_return model_net_event_annotated(
392
        int net_id,
393 394
        char const * annotation,
        char const * category, 
395 396 397 398
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        tw_stime offset,
        int remote_event_size,
399
        void const * remote_event,
400
        int self_event_size,
401
        void const * self_event,
402
        tw_lp *sender){
403
    struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, false);
404
    return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
405 406
            message_size, 0, offset, remote_event_size, remote_event,
            self_event_size, self_event, sender);
407 408
}

409
model_net_event_return model_net_event_mctx(
410 411 412 413 414 415 416 417 418 419 420 421
        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){
422 423 424 425
    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);
426 427
}

428
model_net_event_return model_net_pull_event(
429
        int net_id,
430
        char const *category,
431 432 433 434
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
435
        void const *self_event,
436 437 438
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
439 440 441
    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);
442 443
}

444
model_net_event_return model_net_pull_event_annotated(
445
        int net_id,
446 447
        char const * annotation,
        char const *category,
448 449 450 451
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
452
        void const *self_event,
453 454 455
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
456
    struct codes_mctx mc = codes_mctx_set_group_modulo(annotation, false);
457
    return model_net_event_impl_base(net_id, &mc, &mc, category, final_dest_lp,
458 459
            message_size, 1, offset, self_event_size, self_event, 0, NULL,
            sender);
460 461
}

462
model_net_event_return model_net_pull_event_mctx(
463
        int net_id,
464 465
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
466 467 468 469 470 471 472 473 474
        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 */
475 476
    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,
477 478 479
            self_event, 0, NULL, sender);
}

480 481 482 483 484 485 486 487
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);
}

488 489 490 491
void model_net_event_rc(
        int net_id,
        tw_lp *sender,
        uint64_t message_size){
492 493
    (void)net_id;
    (void)message_size;
494
    model_net_event_impl_base_rc(sender);
495 496 497 498 499
}

void model_net_pull_event_rc(
        int net_id,
        tw_lp *sender) {
500
    (void)net_id;
501
    model_net_event_impl_base_rc(sender);
502 503
}

504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519
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;
520 521 522 523 524 525 526 527 528 529 530
        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;
531 532 533 534 535
        default:
            tw_error(TW_LOC, "unknown or unsupported msg_param_type");
    }
}

536
/* returns the message size, can be either simplenet, dragonfly or torus message size*/
537
int model_net_get_msg_sz(int net_id)
538
{
539 540 541 542
    (void)net_id;
    // TODO: Add checks on network name
    // TODO: Add dragonfly and torus network models
    return sizeof(model_net_wrap_msg);
543
#if 0
544 545 546 547 548
    if(net_id < 0 || net_id >= MAX_NETS)
    {
        printf("%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
        exit(-1);
    }
549

550
    return method_array[net_id]->mn_get_msg_sz();
551
#endif
552 553 554
}

/* returns the packet size in the modelnet struct */
555
uint64_t model_net_get_packet_size(int net_id)
556
{
557
  if(net_id < 0 || net_id >= MAX_NETS)
558 559 560 561 562 563 564
     {
       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?
}

565
/* This event does a collective operation call for model-net */
566
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)
567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586
{
  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);
}

587 588 589
/* returns lp type for modelnet */
const tw_lptype* model_net_get_lp_type(int net_id)
{
590
    if(net_id < 0 || net_id >= MAX_NETS)
591 592 593 594 595 596 597 598 599 600
     {
       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();
}

601
const st_model_types* model_net_get_model_stat_type(int net_id)
602
{
603
   return method_array[net_id]->mn_get_model_stat_types();
604 605
}

606 607
void model_net_report_stats(int net_id)
{
608
  if(net_id < 0 || net_id >= MAX_NETS)
609 610 611 612 613 614 615
  {
    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
616 617
   method_array[net_id]->mn_report_stats();
   return;
618
}
619

620
tw_lpid model_net_find_local_device(
621 622 623
        int          net_id,
        const char * annotation,
        int          ignore_annotations,
624 625 626 627 628 629 630 631 632
        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;
633
    }
634
    return model_net_find_local_device_mctx(net_id, mc_p, sender_gid);
635 636
}

637 638 639 640
tw_lpid model_net_find_local_device_mctx(
        int net_id,
        struct codes_mctx const * map_ctx,
        tw_lpid sender_gid)
641
{
642 643
    return codes_mctx_to_lpid(map_ctx, model_net_lp_config_names[net_id],
            sender_gid);
644 645
}

646 647 648 649 650
#ifdef ENABLE_CORTEX
/* structure that gives access to the topology functions */
cortex_topology model_net_topology;
#endif

651 652 653 654 655 656 657 658
/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */