model-net-lp.c 31.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

#include <stddef.h>
#include <assert.h>
#include "codes/model-net.h"
#include "codes/model-net-method.h"
#include "codes/model-net-lp.h"
12
#include "codes/model-net-sched.h"
13 14 15 16 17 18 19 20
#include "codes/codes_mapping.h"
#include "codes/jenkins-hash.h"

#define MN_NAME "model_net_base"

/**** BEGIN SIMULATION DATA STRUCTURES ****/

int model_net_base_magic;
21
int mn_sample_enabled = 0;
22 23 24 25 26

// message-type specific offsets - don't want to get bitten later by alignment
// issues...
static int msg_offsets[MAX_NETS];

27 28
typedef struct model_net_base_params_s {
    model_net_sched_cfg_params sched_params;
29
    uint64_t packet_size;
Nikhil's avatar
Nikhil committed
30
    int num_queues;
31
    int use_recv_queue;
Nikhil's avatar
Nikhil committed
32 33
    tw_stime nic_seq_delay;
    int node_copy_queues;
34 35
} model_net_base_params;

36
/* annotation-specific parameters (unannotated entry occurs at the
37 38 39 40 41
 * last index) */
static int                       num_params = 0;
static const char              * annos[CONFIGURATION_MAX_ANNOS];
static model_net_base_params     all_params[CONFIGURATION_MAX_ANNOS];

42
static tw_stime mn_sample_interval = 0.0;
43
static tw_stime mn_sample_end = 0.0;
Nikhil's avatar
Nikhil committed
44 45
static int servers_per_node_queue = -1;
extern tw_stime codes_cn_delay;
46

47
typedef struct model_net_base_state {
Nikhil's avatar
Nikhil committed
48
    int net_id, nics_per_router;
49
    // whether scheduler loop is running
Nikhil's avatar
Nikhil committed
50
    int *in_sched_send_loop, in_sched_recv_loop;
51 52 53
    // unique message id counter. This doesn't get decremented on RC to prevent
    // optimistic orderings using "stale" ids
    uint64_t msg_id;
54
    // model-net schedulers
Nikhil's avatar
Nikhil committed
55
    model_net_sched **sched_send, *sched_recv;
56 57
    // parameters
    const model_net_base_params * params;
58 59 60
    // lp type and state of underlying model net method - cache here so we
    // don't have to constantly look up
    const tw_lptype *sub_type;
61
    const st_model_types *sub_model_type;
62
    void *sub_state;
Nikhil's avatar
Nikhil committed
63 64
    tw_stime next_available_time;
    tw_stime *node_copy_next_available_time;
65 66
} model_net_base_state;

67

68 69 70 71
/**** END SIMULATION DATA STRUCTURES ****/

/**** BEGIN LP, EVENT PROCESSING FUNCTION DECLS ****/

72
/* ROSS LP processing functions */
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
static void model_net_base_lp_init(
        model_net_base_state * ns,
        tw_lp * lp);
static void model_net_base_event(
        model_net_base_state * ns,
        tw_bf * b,
        model_net_wrap_msg * m,
        tw_lp * lp);
static void model_net_base_event_rc(
        model_net_base_state * ns,
        tw_bf * b,
        model_net_wrap_msg * m,
        tw_lp * lp);
static void model_net_base_finalize(
        model_net_base_state * ns,
        tw_lp * lp);

/* event type handlers */
static void handle_new_msg(
        model_net_base_state * ns,
        tw_bf *b,
        model_net_wrap_msg * m,
        tw_lp * lp);
static void handle_sched_next(
        model_net_base_state * ns,
        tw_bf *b,
        model_net_wrap_msg * m,
        tw_lp * lp);
static void handle_new_msg_rc(
        model_net_base_state * ns,
        tw_bf *b,
        model_net_wrap_msg * m,
        tw_lp * lp);
static void handle_sched_next_rc(
        model_net_base_state * ns,
        tw_bf *b,
        model_net_wrap_msg * m,
        tw_lp * lp);

/* ROSS function pointer table for this LP */
tw_lptype model_net_base_lp = {
114 115 116 117
    (init_f) model_net_base_lp_init,
    (pre_run_f) NULL,
    (event_f) model_net_base_event,
    (revent_f) model_net_base_event_rc,
118 119
    (commit_f) NULL,
    (final_f)  model_net_base_finalize,
120 121
    (map_f) codes_mapping,
    sizeof(model_net_base_state),
122 123
};

124 125 126 127
/* setup for the ROSS event tracing
 * can have a different function for  rbev_trace_f and ev_trace_f
 * but right now it is set to the same function for both
 */
128
void mn_event_collect(model_net_wrap_msg *m, tw_lp *lp, char *buffer, int *collect_flag)
129
{
130 131 132
    // assigning large numbers to message types to make it easier to
    // determine which messages are model net base LP msgs
    int type;
133
    void * sub_msg;
134 135 136
    switch (m->h.event_type){
        case MN_BASE_NEW_MSG:
            type = 9000;
137
            memcpy(buffer, &type, sizeof(type));
138 139 140
            break;
        case MN_BASE_SCHED_NEXT:
            type = 9001;
141
            memcpy(buffer, &type, sizeof(type));
142 143 144
            break;
        case MN_BASE_SAMPLE: 
            type = 9002;
145
            memcpy(buffer, &type, sizeof(type));
146 147
            break;
        case MN_BASE_PASS:
148
            sub_msg = ((char*)m)+msg_offsets[((model_net_base_state*)lp->cur_state)->net_id];
149 150
            if (g_st_ev_trace == RB_TRACE || g_st_ev_trace == COMMIT_TRACE)
                (((model_net_base_state*)lp->cur_state)->sub_model_type->rbev_trace)(sub_msg, lp, buffer, collect_flag);
151
            else if (g_st_ev_trace == FULL_TRACE)
152
                (((model_net_base_state*)lp->cur_state)->sub_model_type->ev_trace)(sub_msg, lp, buffer, collect_flag);
153
            break;
154
        default:  // this shouldn't happen, but can help detect an issue
155 156 157
            type = 9004;
            break;
    }
158 159
}

160 161 162 163 164 165 166 167 168 169
void mn_model_stat_collect(model_net_base_state *s, tw_lp *lp, char *buffer)
{
    // need to call the model level stats collection fn
    (*s->sub_model_type->model_stat_fn)(s->sub_state, lp, buffer);
    return;
}

st_model_types mn_model_types[MAX_NETS];

st_model_types mn_model_base_type = {
170
    (rbev_trace_f) mn_event_collect,
171
     sizeof(int),
172
     (ev_trace_f) mn_event_collect,
173
     sizeof(int),
174 175
     (model_stat_f) mn_model_stat_collect,
     0
176 177
};

178 179 180 181
/**** END LP, EVENT PROCESSING FUNCTION DECLS ****/

/**** BEGIN IMPLEMENTATIONS ****/

182
void model_net_enable_sampling(tw_stime interval, tw_stime end)
183 184
{
    mn_sample_interval = interval;
185
    mn_sample_end = end;
186 187 188 189 190 191 192 193 194 195 196
    mn_sample_enabled = 1;
}

int model_net_sampling_enabled(void)
{
    return mn_sample_enabled;
}

// schedule sample event - want to be precise, so no noise here
static void issue_sample_event(tw_lp *lp)
{
197 198 199 200 201 202
    if (tw_now(lp) + mn_sample_interval < mn_sample_end + 0.0001) {
        tw_event *e = tw_event_new(lp->gid, mn_sample_interval, lp);
        model_net_wrap_msg *m = tw_event_data(e);
        msg_set_header(model_net_base_magic, MN_BASE_SAMPLE, lp->gid, &m->h);
        tw_event_send(e);
    }
203 204
}

205 206 207 208
void model_net_base_register(int *do_config_nets){
    // here, we initialize ALL lp types to use the base type
    for (int i = 0; i < MAX_NETS; i++){
        if (do_config_nets[i]){
209 210 211 212 213 214 215
            // some model-net lps need custom registration hooks (dragonfly).
            // Those that don't NULL out the reg. function
            if (method_array[i]->mn_register == NULL)
                lp_type_register(model_net_lp_config_names[i],
                        &model_net_base_lp);
            else
                method_array[i]->mn_register(&model_net_base_lp);
216
            if (g_st_ev_trace || g_st_model_stats) // for ROSS event tracing
217
            {
218 219 220 221
                memcpy(&mn_model_types[i], &mn_model_base_type, sizeof(st_model_types));

                if (method_array[i]->mn_model_stat_register == NULL)
                    st_model_type_register(model_net_lp_config_names[i], &mn_model_types[i]);
222
                else
223
                    method_array[i]->mn_model_stat_register(&mn_model_types[i]);
224
            }
225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240
        }
    }
}

static void base_read_config(const char * anno, model_net_base_params *p){
    char sched[MAX_NAME_LENGTH];
    long int packet_size_l = 0;
    uint64_t packet_size;
    int ret;

    ret = configuration_get_value(&config, "PARAMS", "modelnet_scheduler",
            anno, sched, MAX_NAME_LENGTH);
    configuration_get_value_longint(&config, "PARAMS", "packet_size", anno,
            &packet_size_l);
    packet_size = packet_size_l;

Nikhil's avatar
Nikhil committed
241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
    p->num_queues = 1;
    ret = configuration_get_value_int(&config, "PARAMS", "num_injection_queues", anno,
            &p->num_queues);
    if(ret && !g_tw_mynode) {
        fprintf(stdout, "NIC num injection port not specified, "
                "setting to %d\n", p->num_queues);
    }

    p->nic_seq_delay = 200;
    ret = configuration_get_value_double(&config, "PARAMS", "nic_seq_delay", anno,
            &p->nic_seq_delay);
    if(ret && !g_tw_mynode) {
        fprintf(stdout, "NIC seq delay not specified, "
                "setting to %lf\n", p->nic_seq_delay);
    }

    p->node_copy_queues = 4;
    ret = configuration_get_value_int(&config, "PARAMS", "node_copy_queues", anno,
            &p->node_copy_queues);
    if(ret && !g_tw_mynode) {
        fprintf(stdout, "NIC num copy queues not specified, "
                "setting to %d\n", p->node_copy_queues);
    }

265 266 267 268
    if (ret > 0){
        int i;
        for (i = 0; i < MAX_SCHEDS; i++){
            if (strcmp(sched_names[i], sched) == 0){
269
                p->sched_params.type = i;
270 271 272 273 274
                break;
            }
        }
        if (i == MAX_SCHEDS){
            tw_error(TW_LOC,"Unknown value for PARAMS:modelnet-scheduler : "
275
                    "%s", sched);
276 277 278 279
        }
    }
    else{
        // default: FCFS
280
        p->sched_params.type = MN_SCHED_FCFS;
281 282
    }

283 284
    // get scheduler-specific parameters
    if (p->sched_params.type == MN_SCHED_PRIO){
285
        // prio scheduler uses default parameters
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
        int             * num_prios = &p->sched_params.u.prio.num_prios;
        enum sched_type * sub_stype = &p->sched_params.u.prio.sub_stype;
        // number of priorities to allocate
        ret = configuration_get_value_int(&config, "PARAMS",
                "prio-sched-num-prios", anno, num_prios);
        if (ret != 0)
            *num_prios = 10;

        ret = configuration_get_value(&config, "PARAMS",
                "prio-sched-sub-sched", anno, sched, MAX_NAME_LENGTH);
        if (ret == 0)
            *sub_stype = MN_SCHED_FCFS;
        else{
            int i;
            for (i = 0; i < MAX_SCHEDS; i++){
                if (strcmp(sched_names[i], sched) == 0){
                    *sub_stype = i;
                    break;
                }
            }
            if (i == MAX_SCHEDS){
                tw_error(TW_LOC, "Unknown value for "
                        "PARAMS:prio-sched-sub-sched %s", sched);
            }
            else if (i == MN_SCHED_PRIO){
                tw_error(TW_LOC, "priority scheduler cannot be used as a "
                        "priority scheduler's sub sched "
                        "(PARAMS:prio-sched-sub-sched)");
            }
        }
316 317
    }

318 319 320 321 322 323 324 325 326 327 328 329 330
    if (p->sched_params.type == MN_SCHED_FCFS_FULL ||
            (p->sched_params.type == MN_SCHED_PRIO &&
             p->sched_params.u.prio.sub_stype == MN_SCHED_FCFS_FULL)){
        // override packet size to something huge (leave a bit in the unlikely
        // case that an op using packet size causes overflow)
        packet_size = 1ull << 62;
    }
    else if (!packet_size &&
            (p->sched_params.type != MN_SCHED_FCFS_FULL ||
             (p->sched_params.type == MN_SCHED_PRIO &&
              p->sched_params.u.prio.sub_stype != MN_SCHED_FCFS_FULL))){
        packet_size = 512;
        fprintf(stderr, "WARNING, no packet size specified, setting packet "
Jonathan Jenkins's avatar
Jonathan Jenkins committed
331
                "size to %llu\n", LLU(packet_size));
332 333 334
    }


335 336 337 338
    p->packet_size = packet_size;
}

void model_net_base_configure(){
339 340 341 342 343
    uint32_t h1=0, h2=0;

    bj_hashlittle2(MN_NAME, strlen(MN_NAME), &h1, &h2);
    model_net_base_magic = h1+h2;

344 345 346
    // set up offsets - doesn't matter if they are actually used or not
    msg_offsets[SIMPLENET] =
        offsetof(model_net_wrap_msg, msg.m_snet);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
347 348
    msg_offsets[SIMPLEP2P] =
        offsetof(model_net_wrap_msg, msg.m_sp2p);
349 350 351
    msg_offsets[TORUS] =
        offsetof(model_net_wrap_msg, msg.m_torus);
    msg_offsets[DRAGONFLY] =
352
        offsetof(model_net_wrap_msg, msg.m_dfly);
353 354
    // note: dragonfly router uses the same event struct
    msg_offsets[DRAGONFLY_ROUTER] =
355
        offsetof(model_net_wrap_msg, msg.m_dfly);
356
    msg_offsets[DRAGONFLY_CUSTOM] =
357
        offsetof(model_net_wrap_msg, msg.m_custom_dfly);
358
    msg_offsets[DRAGONFLY_CUSTOM_ROUTER] =
359
        offsetof(model_net_wrap_msg, msg.m_custom_dfly);
360 361
    msg_offsets[SLIMFLY] =
        offsetof(model_net_wrap_msg, msg.m_slim);
362 363
    msg_offsets[FATTREE] =
	offsetof(model_net_wrap_msg, msg.m_fat);
364 365
    msg_offsets[LOGGP] =
        offsetof(model_net_wrap_msg, msg.m_loggp);
366 367 368 369
    msg_offsets[EXPRESS_MESH] =
        offsetof(model_net_wrap_msg, msg.m_em);
    msg_offsets[EXPRESS_MESH_ROUTER] =
        offsetof(model_net_wrap_msg, msg.m_em);
370

371 372 373 374 375 376 377 378 379 380
    // perform the configuration(s)
    // This part is tricky, as we basically have to look up all annotations that
    // have LP names of the form modelnet_*. For each of those, we need to read
    // the base parameters
    // - the init is a little easier as we can use the LP-id to look up the
    // annotation

    // first grab all of the annotations and store locally
    for (int c = 0; c < lpconf.lpannos_count; c++){
        const config_anno_map_t *amap = &lpconf.lpannos[c];
381
        if (strncmp("modelnet_", amap->lp_name.ptr, 9) == 0){
382 383 384
            for (int n = 0; n < amap->num_annos; n++){
                int a;
                for (a = 0; a < num_params; a++){
385 386
                    if (annos[a] != NULL && amap->annotations[n].ptr != NULL &&
                            strcmp(amap->annotations[n].ptr, annos[a]) == 0){
387 388 389 390 391
                        break;
                    }
                }
                if (a == num_params){
                    // found a new annotation
392
                    annos[num_params++] = amap->annotations[n].ptr;
393 394 395 396 397 398 399 400 401 402 403 404 405
                }
            }
            if (amap->has_unanno_lp){
                int a;
                for (a = 0; a < num_params; a++){
                    if (annos[a] == NULL)
                        break;
                }
                if (a == num_params){
                    // found a new (empty) annotation
                    annos[num_params++] = NULL;
                }
            }
406 407
        }
    }
408 409 410 411 412 413

    // now that we have all of the annos for all of the networks, loop through
    // and read the configs
    for (int i = 0; i < num_params; i++){
        base_read_config(annos[i], &all_params[i]);
    }
414 415 416 417 418 419
}

void model_net_base_lp_init(
        model_net_base_state * ns,
        tw_lp * lp){
    // obtain the underlying lp type through codes-mapping
Nikhil's avatar
Nikhil committed
420
    char lp_type_name[MAX_NAME_LENGTH], anno[MAX_NAME_LENGTH], group[MAX_NAME_LENGTH];
421 422
    int dummy;

Nikhil's avatar
Nikhil committed
423
    codes_mapping_get_lp_info(lp->gid, group, &dummy,
424 425 426 427 428 429 430 431 432 433
            lp_type_name, &dummy, anno, &dummy, &dummy);

    // get annotation-specific parameters
    for (int i = 0; i < num_params; i++){
        if ((anno[0]=='\0' && annos[i] == NULL) ||
                strcmp(anno, annos[i]) == 0){
            ns->params = &all_params[i];
            break;
        }
    }
434 435 436 437 438 439 440 441

    // find the corresponding method name / index
    for (int i = 0; i < MAX_NETS; i++){
        if (strcmp(model_net_lp_config_names[i], lp_type_name) == 0){
            ns->net_id = i;
            break;
        }
    }
Nikhil's avatar
Nikhil committed
442 443 444 445 446 447 448 449 450 451
    
    ns->nics_per_router = codes_mapping_get_lp_count(group, 1,
            lp_type_name, NULL, 1);
    
    ns->msg_id = 0;
    ns->next_available_time = 0;
    ns->node_copy_next_available_time = (tw_stime*)malloc(ns->params->node_copy_queues * sizeof(tw_stime));
    for(int i = 0; i < ns->params->node_copy_queues; i++) {
        ns->node_copy_next_available_time[i] = 0;
    }
452

Nikhil's avatar
Nikhil committed
453 454 455 456 457 458 459 460
    ns->in_sched_send_loop = (int *)malloc(ns->params->num_queues * sizeof(int));
    ns->sched_send = (model_net_sched**)malloc(ns->params->num_queues * sizeof(model_net_sched*));
    for(int i = 0; i < ns->params->num_queues; i++) {
        ns->sched_send[i] = (model_net_sched*)malloc(sizeof(model_net_sched));
        model_net_sched_init(&ns->params->sched_params, 0, method_array[ns->net_id],
                ns->sched_send[i]);
        ns->in_sched_send_loop[i] = 0;
    }
461 462 463
    ns->sched_recv = malloc(sizeof(model_net_sched));
    model_net_sched_init(&ns->params->sched_params, 1, method_array[ns->net_id],
            ns->sched_recv);
464

465
    ns->sub_type = model_net_get_lp_type(ns->net_id);
466

467 468 469 470 471 472
    /* some ROSS instrumentation setup */
    if (g_st_ev_trace || g_st_model_stats)
    {
        ns->sub_model_type = model_net_get_model_stat_type(ns->net_id);
        mn_model_types[ns->net_id].mstat_sz = ns->sub_model_type->mstat_sz;
    }
473

474 475 476 477 478 479
    // NOTE: some models actually expect LP state to be 0 initialized...
    // *cough anything that uses mn_stats_array cough*
    ns->sub_state = calloc(1, ns->sub_type->state_sz);

    // initialize the model-net method
    ns->sub_type->init(ns->sub_state, lp);
480 481 482 483 484 485

    // check validity of sampling function
    event_f  sample  = method_array[ns->net_id]->mn_sample_fn;
    revent_f rsample = method_array[ns->net_id]->mn_sample_rc_fn;
    if (model_net_sampling_enabled()) {
        if (sample == NULL) {
486 487 488
            /* MM: Commented out temporarily--- */
            //tw_error(TW_LOC,
            //        "Sampling requested for a model that doesn't provide it\n");
489 490 491 492
        }
        else if (rsample == NULL &&
                (g_tw_synchronization_protocol == OPTIMISTIC ||
                 g_tw_synchronization_protocol == OPTIMISTIC_DEBUG)) {
493 494 495
            /* MM: Commented out temporarily--- */
            //tw_error(TW_LOC,
            //        "Sampling requested for a model that doesn't provide it\n");
496 497 498 499 500 501 502 503
        }
        else {
            init_f sinit = method_array[ns->net_id]->mn_sample_init_fn;
            if (sinit != NULL)
                sinit(ns->sub_state, lp);
            issue_sample_event(lp);
        }
    }
504 505 506 507 508 509 510
}

void model_net_base_event(
        model_net_base_state * ns,
        tw_bf * b,
        model_net_wrap_msg * m,
        tw_lp * lp){
511 512

    if(m->h.magic != model_net_base_magic)
513
        printf("\n LP ID mismatched %llu ", lp->gid);
514

515
    assert(m->h.magic == model_net_base_magic);
516 517

    void * sub_msg;
518
    switch (m->h.event_type){
519 520 521 522 523 524
        case MN_BASE_NEW_MSG:
            handle_new_msg(ns, b, m, lp);
            break;
        case MN_BASE_SCHED_NEXT:
            handle_sched_next(ns, b, m, lp);
            break;
525 526 527 528 529 530 531
        case MN_BASE_SAMPLE: ;
            event_f sample = method_array[ns->net_id]->mn_sample_fn;
            assert(model_net_sampling_enabled() && sample != NULL);
            sub_msg = ((char*)m)+msg_offsets[ns->net_id];
            sample(ns->sub_state, b, sub_msg, lp);
            issue_sample_event(lp);
            break;
532
        case MN_BASE_PASS: ;
533
            sub_msg = ((char*)m)+msg_offsets[ns->net_id];
534 535 536 537 538 539 540 541 542 543 544 545 546 547
            ns->sub_type->event(ns->sub_state, b, sub_msg, lp);
            break;
        /* ... */
        default:
            assert(!"model_net_base event type not known");
            break;
    }
}

void model_net_base_event_rc(
        model_net_base_state * ns,
        tw_bf * b,
        model_net_wrap_msg * m,
        tw_lp * lp){
548
    assert(m->h.magic == model_net_base_magic);
549 550

    void * sub_msg;
551
    switch (m->h.event_type){
552 553 554 555 556 557
        case MN_BASE_NEW_MSG:
            handle_new_msg_rc(ns, b, m, lp);
            break;
        case MN_BASE_SCHED_NEXT:
            handle_sched_next_rc(ns, b, m, lp);
            break;
558 559 560 561 562 563
        case MN_BASE_SAMPLE: ;
            revent_f sample_rc = method_array[ns->net_id]->mn_sample_rc_fn;
            assert(model_net_sampling_enabled() && sample_rc != NULL);
            sub_msg = ((char*)m)+msg_offsets[ns->net_id];
            sample_rc(ns->sub_state, b, sub_msg, lp);
            break;
564
        case MN_BASE_PASS: ;
565
            sub_msg = ((char*)m)+msg_offsets[ns->net_id];
566 567 568 569 570 571 572 573 574 575 576 577
            ns->sub_type->revent(ns->sub_state, b, sub_msg, lp);
            break;
        /* ... */
        default:
            assert(!"model_net_base event type not known");
            break;
    }
}

void model_net_base_finalize(
        model_net_base_state * ns,
        tw_lp * lp){
578 579 580
    final_f sfini = method_array[ns->net_id]->mn_sample_fini_fn;
    if (sfini != NULL)
        sfini(ns->sub_state, lp);
581 582 583 584
    ns->sub_type->final(ns->sub_state, lp);
    free(ns->sub_state);
}

585
/// bitfields used:
586
/// c31 - we initiated a sched_next event
587 588 589 590 591
void handle_new_msg(
        model_net_base_state * ns,
        tw_bf *b,
        model_net_wrap_msg * m,
        tw_lp * lp){
Nikhil's avatar
Nikhil committed
592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
    static int num_servers = -1;
    static int servers_per_node = -1;
    if(num_servers == -1) {
        char const *sender_group;
        char const *sender_lpname;
        int rep_id, offset;
        model_net_request *r = &m->msg.m_base.req;
        codes_mapping_get_lp_info2(r->src_lp, &sender_group, &sender_lpname, 
                NULL, &rep_id, &offset);
        num_servers = codes_mapping_get_lp_count(sender_group, 1,
                sender_lpname, NULL, 1);
        servers_per_node = num_servers/ns->params->num_queues; //this is for entire switch
        servers_per_node_queue = num_servers/ns->nics_per_router/ns->params->node_copy_queues;
        if(!g_tw_mynode) {
            fprintf(stdout, "Set num_servers per router %d, servers per "
                "injection queue per router %d, servers per node copy queue "
                "per node %d\n", num_servers, servers_per_node, 
                servers_per_node_queue);
        }
    } 

    if(lp->gid == m->msg.m_base.req.dest_mn_lp) {
        model_net_request *r = &m->msg.m_base.req;
        int rep_id, offset;
        codes_mapping_get_lp_info2(r->src_lp, NULL, NULL, NULL, &rep_id, &offset);
        int queue = offset/ns->nics_per_router/servers_per_node_queue;
        m->msg.m_base.save_ts = ns->node_copy_next_available_time[queue];
        tw_stime exp_time = ((ns->node_copy_next_available_time[queue] 
                            > tw_now(lp)) ? ns->node_copy_next_available_time[queue] : tw_now(lp));
        exp_time += r->msg_size * codes_cn_delay;
        ns->node_copy_next_available_time[queue] = exp_time;
        int remote_event_size = r->remote_event_size;
        int self_event_size = r->self_event_size;
        void *e_msg = (m+1);
        if (remote_event_size > 0) {
            tw_event *e = tw_event_new(r->final_dest_lp, exp_time - tw_now(lp), lp);
            memcpy(tw_event_data(e), e_msg, remote_event_size);
            tw_event_send(e);
            e_msg = (char*)e_msg + remote_event_size; 
        }
        if (self_event_size > 0) {
            tw_event *e = tw_event_new(r->src_lp, exp_time - tw_now(lp), lp);
            memcpy(tw_event_data(e), e_msg, self_event_size);
            tw_event_send(e);
        }
        return;
    }

    if(m->msg.m_base.isQueueReq) {
        m->msg.m_base.save_ts = ns->next_available_time;
        tw_stime exp_time = ((ns->next_available_time > tw_now(lp)) ? ns->next_available_time : tw_now(lp));
        exp_time += ns->params->nic_seq_delay;
        ns->next_available_time = exp_time;
        tw_event *e = tw_event_new(lp->gid, exp_time - tw_now(lp), lp);
        model_net_wrap_msg *m_new = tw_event_data(e);
        memcpy(m_new, m, sizeof(model_net_wrap_msg));
        void *e_msg = (m+1);
        void *e_new_msg = (m_new+1);
        model_net_request *r = &m->msg.m_base.req;
        int remote_event_size = r->remote_event_size;
        int self_event_size = r->self_event_size;
        if (remote_event_size > 0){
            memcpy(e_new_msg, e_msg, remote_event_size);
            e_msg = (char*)e_msg + remote_event_size; 
            e_new_msg = (char*)e_new_msg + remote_event_size; 
        }
        if (self_event_size > 0){
            memcpy(e_new_msg, e_msg, self_event_size);
        }
        m_new->msg.m_base.isQueueReq = 0;
        tw_event_send(e);
        return;
    } 
665
    // simply pass down to the scheduler
666
    model_net_request *r = &m->msg.m_base.req;
667 668
    // don't forget to set packet size, now that we're responsible for it!
    r->packet_size = ns->params->packet_size;
669
    r->msg_id = ns->msg_id++;
670 671
    void * m_data = m+1;
    void *remote = NULL, *local = NULL;
672
    if (r->remote_event_size > 0){
673 674
        remote = m_data;
        m_data = (char*)m_data + r->remote_event_size;
675 676
    }
    if (r->self_event_size > 0){
677
        local = m_data;
678
    }
679

Nikhil's avatar
Nikhil committed
680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697
    int queue_offset = 0;
    if(!m->msg.m_base.is_from_remote && ns->params->num_queues != 1) {
        int rep_id, offset;
        if(num_servers == -1) {
            char const *sender_group;
            char const *sender_lpname;
            codes_mapping_get_lp_info2(r->src_lp, &sender_group, &sender_lpname, 
                    NULL, &rep_id, &offset);
            num_servers = codes_mapping_get_lp_count(sender_group, 1,
                    sender_lpname, NULL, 1);
            servers_per_node = num_servers/ns->params->num_queues;
        } else {
            codes_mapping_get_lp_info2(r->src_lp, NULL, NULL, NULL, &rep_id, &offset);
        }
        queue_offset = offset/servers_per_node;
    }
    r->queue_offset = queue_offset;

698
    // set message-specific params
699
    int is_from_remote = m->msg.m_base.is_from_remote;
Nikhil's avatar
Nikhil committed
700
    model_net_sched *ss = is_from_remote ? ns->sched_recv : ns->sched_send[queue_offset];
701
    int *in_sched_loop = is_from_remote  ?
Nikhil's avatar
Nikhil committed
702
        &ns->in_sched_recv_loop : &ns->in_sched_send_loop[queue_offset];
703 704
    model_net_sched_add(r, &m->msg.m_base.sched_params, r->remote_event_size,
            remote, r->self_event_size, local, ss, &m->msg.m_base.rc, lp);
705

706
    if (*in_sched_loop == 0){
707 708
        b->c31 = 1;
        /* No need to issue an extra sched-next event if we're currently idle */
709
        *in_sched_loop = 1;
710 711 712 713 714
        /* NOTE: we can do this because the sched rc struct in the event is
         * *very* lightly used (there's harmless overlap in usage for the
         * priority scheduler) */
        handle_sched_next(ns, b, m, lp);
        assert(*in_sched_loop); // we shouldn't have fallen out of the loop
715 716
    }
}
717 718 719

void handle_new_msg_rc(
        model_net_base_state *ns,
720
        tw_bf *b,
721 722
        model_net_wrap_msg *m,
        tw_lp *lp){
Nikhil's avatar
Nikhil committed
723 724 725 726 727 728 729 730 731 732 733 734 735
    if(lp->gid == m->msg.m_base.req.dest_mn_lp) {
        model_net_request *r = &m->msg.m_base.req;
        int rep_id, offset;
        codes_mapping_get_lp_info2(r->src_lp, NULL, NULL, NULL, &rep_id, &offset);
        int queue = offset/ns->nics_per_router/servers_per_node_queue;
        ns->node_copy_next_available_time[queue] = m->msg.m_base.save_ts;
        return;
    }
    if(m->msg.m_base.isQueueReq) {
        ns->next_available_time = m->msg.m_base.save_ts;
        return;
    }
    model_net_request *r = &m->msg.m_base.req;
736
    int is_from_remote = m->msg.m_base.is_from_remote;
Nikhil's avatar
Nikhil committed
737
    model_net_sched *ss = is_from_remote ? ns->sched_recv : ns->sched_send[r->queue_offset];
738
    int *in_sched_loop = is_from_remote  ?
Nikhil's avatar
Nikhil committed
739
        &ns->in_sched_recv_loop : &ns->in_sched_send_loop[r->queue_offset];
740

741 742
    if (b->c31) {
        handle_sched_next_rc(ns, b, m, lp);
743
        *in_sched_loop = 0;
744
    }
745
    model_net_sched_add_rc(ss, &m->msg.m_base.rc, lp);
746
}
747 748 749 750

/// bitfields used
/// c0 - scheduler loop is finished
void handle_sched_next(
751 752 753 754
        model_net_base_state * ns,
        tw_bf *b,
        model_net_wrap_msg * m,
        tw_lp * lp){
755
    tw_stime poffset;
Nikhil's avatar
Nikhil committed
756
    model_net_request *r = &m->msg.m_base.req;
757
    int is_from_remote = m->msg.m_base.is_from_remote;
Nikhil's avatar
Nikhil committed
758
    model_net_sched * ss = is_from_remote ? ns->sched_recv : ns->sched_send[r->queue_offset];
759
    int *in_sched_loop = is_from_remote ?
Nikhil's avatar
Nikhil committed
760
        &ns->in_sched_recv_loop : &ns->in_sched_send_loop[r->queue_offset];
761
    int ret = model_net_sched_next(&poffset, ss, m+1, &m->msg.m_base.rc, lp);
762 763 764 765
    // we only need to know whether scheduling is finished or not - if not,
    // go to the 'next iteration' of the loop
    if (ret == -1){
        b->c0 = 1;
766
        *in_sched_loop = 0;
767
    }
768 769 770
    // Currently, only a subset of the network implementations use the
    // callback-based scheduling loop (model_net_method_idle_event).
    // For all others, we need to schedule the next packet
771
    // immediately
Jonathan Jenkins's avatar
Jonathan Jenkins committed
772
    else if (ns->net_id == SIMPLEP2P || ns->net_id == TORUS){
773
        tw_event *e = tw_event_new(lp->gid,
774
                poffset+codes_local_latency(lp), lp);
775
        model_net_wrap_msg *m_wrap = tw_event_data(e);
Nikhil's avatar
Nikhil committed
776
        model_net_request *r_wrap = &m_wrap->msg.m_base.req;
777
        msg_set_header(model_net_base_magic, MN_BASE_SCHED_NEXT, lp->gid,
778 779
                &m_wrap->h);
        m_wrap->msg.m_base.is_from_remote = is_from_remote;
Nikhil's avatar
Nikhil committed
780
        r_wrap->queue_offset = r->queue_offset;
781 782
        // no need to set m_base here
        tw_event_send(e);
783 784
    }
}
785

786 787 788 789 790
void handle_sched_next_rc(
        model_net_base_state * ns,
        tw_bf *b,
        model_net_wrap_msg * m,
        tw_lp * lp){
Nikhil's avatar
Nikhil committed
791
    model_net_request *r = &m->msg.m_base.req;
792
    int is_from_remote = m->msg.m_base.is_from_remote;
Nikhil's avatar
Nikhil committed
793
    model_net_sched * ss = is_from_remote ? ns->sched_recv : ns->sched_send[r->queue_offset];
794
    int *in_sched_loop = is_from_remote ?
Nikhil's avatar
Nikhil committed
795
        &ns->in_sched_recv_loop : &ns->in_sched_send_loop[r->queue_offset];
796

797
    model_net_sched_next_rc(ss, m+1, &m->msg.m_base.rc, lp);
798
    if (b->c0){
799
        *in_sched_loop = 1;
800
    }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
801
    else if (ns->net_id == SIMPLEP2P || ns->net_id == TORUS){
802 803
        codes_local_latency_reverse(lp);
    }
804 805 806 807 808 809 810 811 812 813 814 815 816
}

/**** END IMPLEMENTATIONS ****/

tw_event * model_net_method_event_new(
        tw_lpid dest_gid,
        tw_stime offset_ts,
        tw_lp *sender,
        int net_id,
        void **msg_data,
        void **extra_data){
    tw_event *e = tw_event_new(dest_gid, offset_ts, sender);
    model_net_wrap_msg *m_wrap = tw_event_data(e);
817 818
    msg_set_header(model_net_base_magic, MN_BASE_PASS, sender->gid,
            &m_wrap->h);
819 820 821 822 823 824 825 826
    *msg_data = ((char*)m_wrap)+msg_offsets[net_id];
    // extra_data is optional
    if (extra_data != NULL){
        *extra_data = m_wrap + 1;
    }
    return e;
}

827 828 829 830 831 832 833 834 835 836 837 838
void model_net_method_send_msg_recv_event(
        tw_lpid final_dest_lp,
        tw_lpid dest_mn_lp,
        tw_lpid src_lp, // the "actual" source (as opposed to the model net lp)
        uint64_t msg_size,
        int is_pull,
        uint64_t pull_size,
        int remote_event_size,
        const mn_sched_params *sched_params,
        const char * category,
        int net_id,
        void * msg,
839
        tw_stime offset,
840
        tw_lp *sender){
841
    tw_event *e =
842
        tw_event_new(dest_mn_lp, offset+codes_local_latency(sender), sender);
843 844 845 846 847 848 849 850
    model_net_wrap_msg *m = tw_event_data(e);
    msg_set_header(model_net_base_magic, MN_BASE_NEW_MSG, sender->gid, &m->h);

    if (sched_params != NULL)
        m->msg.m_base.sched_params = *sched_params;
    else
        model_net_sched_set_default_params(&m->msg.m_base.sched_params);

851 852 853
    model_net_request *r = &m->msg.m_base.req;
    r->final_dest_lp = final_dest_lp;
    r->src_lp = src_lp;
854
    // for "recv" events, set the "dest" to this LP in the case of a pull event
855 856 857 858 859 860 861 862 863
    r->dest_mn_lp = sender->gid;
    r->pull_size = pull_size;
    r->msg_size = msg_size;
    // TODO: document why we're setting packet_size this way
    r->packet_size = msg_size;
    r->net_id = net_id;
    r->is_pull = is_pull;
    r->remote_event_size = remote_event_size;
    r->self_event_size = 0;
864 865
    m->msg.m_base.is_from_remote = 1;

866 867
    strncpy(r->category, category, CATEGORY_NAME_MAX-1);
    r->category[CATEGORY_NAME_MAX-1] = '\0';
868 869 870 871 872 873 874 875 876

    if (remote_event_size > 0){
        void * m_dat = model_net_method_get_edata(net_id, msg);
        memcpy(m+1, m_dat, remote_event_size);
    }

    tw_event_send(e);
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
877 878 879 880
void model_net_method_send_msg_recv_event_rc(tw_lp *sender){
    codes_local_latency_reverse(sender);
}

881 882 883

void model_net_method_idle_event(tw_stime offset_ts, int is_recv_queue,
        tw_lp * lp){
Nikhil's avatar
Nikhil committed
884 885 886 887 888
    model_net_method_idle_event2(offset_ts, is_recv_queue, 0, lp);
}

void model_net_method_idle_event2(tw_stime offset_ts, int is_recv_queue, 
        int queue_offset, tw_lp * lp){
889 890
    tw_event *e = tw_event_new(lp->gid, offset_ts, lp);
    model_net_wrap_msg *m_wrap = tw_event_data(e);
Nikhil's avatar
Nikhil committed
891
    model_net_request *r_wrap = &m_wrap->msg.m_base.req;
892 893
    msg_set_header(model_net_base_magic, MN_BASE_SCHED_NEXT, lp->gid,
            &m_wrap->h);
894
    m_wrap->msg.m_base.is_from_remote = is_recv_queue;
Nikhil's avatar
Nikhil committed
895
    r_wrap->queue_offset = queue_offset;
896 897 898
    tw_event_send(e);
}

899 900 901 902 903 904 905 906 907 908 909 910
void * model_net_method_get_edata(int net_id, void *msg){
    return (char*)msg + sizeof(model_net_wrap_msg) - msg_offsets[net_id];
}

/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */