torus.c 67 KB
Newer Older
Philip Carns's avatar
Philip Carns committed
1 2 3 4 5 6
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

7 8
#include <ross.h>
#include <assert.h>
9
#include <string.h>
10 11 12 13 14 15

#include "codes/lp-io.h"
#include "codes/codes_mapping.h"
#include "codes/codes.h"
#include "codes/model-net.h"
#include "codes/model-net-method.h"
16 17
#include "codes/model-net-lp.h"
#include "codes/net/torus.h"
18
#include "codes/rc-stack.h"
19

20 21 22 23 24
#ifdef ENABLE_CORTEX
#include <cortex/cortex.h>
#include <cortex/topology.h>
#endif

25 26
#define DEBUG 1
#define MEAN_INTERVAL 100
27 28 29
// type casts to make compiler happy
#define TRACE ((unsigned long long)(-1))
#define TRACK ((tw_lpid)(-1))
30

31
#define STATICQ 0
32 33 34
/* collective specific parameters */
#define TREE_DEGREE 4
#define LEVEL_DELAY 1000
35
#define TORUS_COLLECTIVE_DEBUG 0
36 37 38 39
#define NUM_COLLECTIVES  1
#define COLLECTIVE_COMPUTATION_DELAY 5700
#define TORUS_FAN_OUT_DELAY 20.0

40 41 42
#define LP_CONFIG_NM (model_net_lp_config_names[TORUS])
#define LP_METHOD_NM (model_net_method_names[TORUS])

43 44
static double maxd(double a, double b) { return a < b ? b : a; }

45
/* Torus network model implementation of codes, implements the modelnet API */
46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
typedef struct nodes_message_list nodes_message_list;
struct nodes_message_list {
    nodes_message msg;
    char* event_data;
    nodes_message_list *next;
    nodes_message_list *prev;
};

void init_nodes_message_list(nodes_message_list *this, nodes_message *inmsg) {
    this->msg = *inmsg;
    this->event_data = NULL;
    this->next = NULL;
    this->prev = NULL;
}

void delete_nodes_message_list(nodes_message_list *this) {
    if(this->event_data != NULL) free(this->event_data);
    free(this);
}

66 67 68 69 70 71 72 73
static void free_tmp(void * ptr)
{
    nodes_message_list * entry = ptr;
    if(entry->event_data != NULL)
        free(entry->event_data);

    free(entry);
}
74 75 76 77 78 79
typedef struct torus_param torus_param;
struct torus_param
{
    int n_dims; /*Dimension of the torus network, 5-D, 7-D or any other*/
    int* dim_length; /*Length of each torus dimension*/
    double link_bandwidth;/* bandwidth for each torus link */
80
    double cn_bandwidth; /* injection bandwidth */
81 82 83 84 85 86 87 88 89 90 91 92 93 94
    int buffer_size; /* number of buffer slots for each vc in flits*/
    int num_vc; /* number of virtual channels for each torus link */
    float mean_process;/* mean process time for each flit  */
    int chunk_size; /* chunk is the smallest unit--default set to 32 */

    /* "derived" torus parameters */

    /* factor, used in torus coordinate calculation */
    int * factor;
    /* half length of each dimension, used in torus coordinates calculation */
    int * half_length;

    double head_delay;
    double credit_delay;
95 96 97 98
    double cn_delay;

    double router_delay;
    int routing;
99
};
100 101

/* codes mapping group name, lp type name */
102
static char grp_name[MAX_NAME_LENGTH];
103
/* codes mapping group id, lp type id, repetition id and offset */
104
int mapping_grp_id, mapping_rep_id, mapping_type_id, mapping_offset;
105 106 107 108

/* for calculating torus model statistics, average and maximum travel time of a packet */
static tw_stime         total_time = 0;
static tw_stime         max_latency = 0;
109
static tw_stime         max_collective = 0;
110 111 112 113 114 115

/* number of finished packets on each PE */
static long long       N_finished_packets = 0;
/* total number of hops traversed by a message on each PE */
static long long       total_hops = 0;

116
/* annotation-specific parameters (unannotated entry occurs at the
117
 * last index) */
118
static int                       num_params = 0;
119 120
static torus_param             * all_params = NULL;
static const config_anno_map_t * anno_map   = NULL;
121 122 123 124 125 126 127

typedef struct nodes_state nodes_state;

/* state of a torus node */
struct nodes_state
{
  /* counts the number of packets sent from this compute node */
128
  unsigned long long packet_counter;
129
  /* availability time of each torus link */
130
  tw_stime** next_link_available_time;
131 132 133 134 135 136
  /* availability of each torus credit link */
  tw_stime** next_credit_available_time;
  /* next flit generate time */
  tw_stime** next_flit_generate_time;
  /* buffer size for each torus virtual channel */
  int** buffer;
137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
  /* Head and tail of the terminal messages list */
  nodes_message_list **terminal_msgs;
  nodes_message_list **terminal_msgs_tail;
  int all_term_length;
  int *terminal_length, *queued_length;
  /* pending packets to be sent out */
  nodes_message_list ***pending_msgs;
  nodes_message_list ***pending_msgs_tail;
  nodes_message_list ***queued_msgs;
  nodes_message_list ***queued_msgs_tail;
  nodes_message_list **other_msgs;
  nodes_message_list **other_msgs_tail;
  int *in_send_loop;
  /* traffic through each torus link */
  int64_t *link_traffic;
152 153 154 155 156 157 158 159
  /* coordinates of the current torus node */
  int* dim_position;
  /* neighbor LP ids for this torus node */
  int* neighbour_minus_lpID;
  int* neighbour_plus_lpID;

  /* records torus statistics for this LP having different communication categories */
  struct mn_stats torus_stats_array[CATEGORY_MAX];
160 161 162 163 164
   /* for collective operations */

  /* collective init time */
  tw_stime collective_init_time;

165
  /* node ID in the tree */
166 167
   tw_lpid node_id;

168
   /* messages sent & received in collectives may get interchanged several times so we have to save the
169
     origin server information in the node's state */
170 171
   tw_lpid origin_svr;

172 173 174 175 176 177 178 179 180 181 182 183 184 185
  /* parent node ID of the current node */
   tw_lpid parent_node_id;
   /* array of children to be allocated in terminal_init*/
   tw_lpid* children;

   /* children of a node can be less than or equal to the tree degree */
   int num_children;

   short is_root;
   short is_leaf;

   /* to maintain a count of child nodes that have fanned in at the parent during the collective
      fan-in phase*/
   int num_fan_nodes;
186 187 188 189 190

   /* LPs annotation */
   const char * anno;
   /* LPs configuration */
   const torus_param * params;
191 192 193

   /* create the RC stack */
   struct rc_stack * st;
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215

   /* finished chunks */
   long finished_chunks;

   /* finished packets */
   long finished_packets;

   /* total hops */
   double total_hops;

   /* total time */
   double total_time;

   /* total data */
   long total_data_sz;
   /* output buf */
   char output_buf[2048];
   char output_busy_buf[2048];

   /* busy time */
   tw_stime * busy_time;
   tw_stime * last_buf_full;
216 217
};

218
static void append_to_node_message_list(
219 220
        nodes_message_list ** thisq,
        nodes_message_list ** thistail,
221
        int index,
222 223 224 225 226 227
        nodes_message_list *msg) {
    if(thisq[index] == NULL) {
        thisq[index] = msg;
    } else {
        thistail[index]->next = msg;
        msg->prev = thistail[index];
228
    }
229 230
    thistail[index] = msg;
}
231
static void prepend_to_node_message_list(
232 233
        nodes_message_list ** thisq,
        nodes_message_list ** thistail,
234
        int index,
235 236 237 238 239 240
        nodes_message_list *msg) {
    if(thisq[index] == NULL) {
        thistail[index] = msg;
    } else {
        thisq[index]->prev = msg;
        msg->next = thisq[index];
241
    }
242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
    thisq[index] = msg;
}

static nodes_message_list* return_head(
        nodes_message_list ** thisq,
        nodes_message_list ** thistail,
        int index) {
    nodes_message_list *head = thisq[index];
    if(head != NULL) {
        thisq[index] = head->next;
        if(head->next != NULL) {
            head->next->prev = NULL;
            head->next = NULL;
        } else {
            thistail[index] = NULL;
        }
    }
    return head;
}

static nodes_message_list* return_tail(
        nodes_message_list ** thisq,
        nodes_message_list ** thistail,
        int index) {
    nodes_message_list *tail = thistail[index];
    if(tail->prev != NULL) {
        tail->prev->next = NULL;
        thistail[index] = tail->prev;
        tail->prev = NULL;
    } else {
        thistail[index] = NULL;
        thisq[index] = NULL;
    }
    return tail;
}

278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
/* convert GiB/s and bytes to ns */
static tw_stime bytes_to_ns(uint64_t bytes, double GB_p_s)
{
    tw_stime time;

    /* bytes to GB */
    time = ((double)bytes)/(1024.0*1024.0*1024.0);
    /* GiB to s */
    time = time / GB_p_s;
    /* s to ns */
    time = time * 1000.0 * 1000.0 * 1000.0;

    return(time);
}

293 294 295 296
static void torus_read_config(
        const char         * anno,
        torus_param        * params){
    char dim_length_str[MAX_NAME_LENGTH];
297
    int i;
298 299 300 301

    // shorthand
    torus_param *p = params;

302 303
    int rc = configuration_get_value_int(&config, "PARAMS", "n_dims", anno, &p->n_dims);
    if(rc) {
304
        p->n_dims = 4; /* a 4-D torus */
305
        fprintf(stderr,
306 307 308 309
                "Warning: Number of dimensions not specified, setting to %d\n",
                p->n_dims);
    }

310
   rc = configuration_get_value_double(&config, "PARAMS", "link_bandwidth", anno,
311
            &p->link_bandwidth);
312
    if(rc) {
313 314 315 316 317
        p->link_bandwidth = 2.0; /*default bg/q configuration */
        fprintf(stderr, "Link bandwidth not specified, setting to %lf\n",
                p->link_bandwidth);
    }

318 319
    rc = configuration_get_value_int(&config, "PARAMS", "buffer_size", anno, &p->buffer_size);
    if(rc) {
320 321 322 323 324
        p->buffer_size = 2048;
        fprintf(stderr, "Buffer size not specified, setting to %d",
                p->buffer_size);
    }

325

326 327 328
    rc = configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(rc) {
        p->chunk_size = 128;
329 330 331 332 333 334 335
        fprintf(stderr, "Warning: Chunk size not specified, setting to %d\n",
                p->chunk_size);
    }
        /* by default, we have one for taking packets,
         * another for taking credit*/
        p->num_vc = 1;

336
    rc = configuration_get_value(&config, "PARAMS", "dim_length", anno,
337 338 339 340 341 342 343 344 345
            dim_length_str, MAX_NAME_LENGTH);
    if (rc == 0){
        tw_error(TW_LOC, "couldn't read PARAMS:dim_length");
    }
    char* token;
    p->dim_length=malloc(p->n_dims*sizeof(*p->dim_length));
    token = strtok(dim_length_str, ",");
    i = 0;
    while(token != NULL)
346
    {
347 348 349 350 351 352 353 354 355 356
        sscanf(token, "%d", &p->dim_length[i]);
        if(p->dim_length[i] <= 0)
        {
            tw_error(TW_LOC, "Invalid torus dimension specified "
                    "(%d at pos %d), exiting... ", p->dim_length[i], i);
        }
        i++;
        token = strtok(NULL,",");
    }
    // create derived parameters
357

358
    // factor is an exclusive prefix product
359
    p->router_delay = 50;
360
    p->cn_delay = 10;
361 362 363 364 365 366 367 368 369 370
    p->factor = malloc(p->n_dims * sizeof(int));
    p->factor[0] = 1;
    for(i = 1; i < p->n_dims; i++)
        p->factor[i] = p->factor[i-1] * p->dim_length[i-1];

    p->half_length = malloc(p->n_dims * sizeof(int));
    for (i = 0; i < p->n_dims; i++)
        p->half_length[i] = p->dim_length[i] / 2;

    // some latency numbers
371 372
    p->head_delay = bytes_to_ns(p->chunk_size, p->link_bandwidth);
    p->credit_delay = bytes_to_ns(8, p->link_bandwidth);
373 374 375 376 377 378 379 380
}

static void torus_configure(){
    anno_map = codes_mapping_get_lp_anno_map(LP_CONFIG_NM);
    assert(anno_map);
    num_params = anno_map->num_annos + (anno_map->has_unanno_lp > 0);
    all_params = malloc(num_params * sizeof(*all_params));

381
    for (int i = 0; i < anno_map->num_annos; i++){
382
        const char * anno = anno_map->annotations[i].ptr;
383 384 385 386
        torus_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        torus_read_config(NULL, &all_params[anno_map->num_annos]);
387
    }
388 389
}

390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416
/* helper functions - convert between flat ids and torus n-dimensional ids */
static void to_dim_id(
        int flat_id,
        int ndims,
        const int *dim_lens,
        int *out_dim_ids)
{
    for (int i = 0; i < ndims; i++) {
        out_dim_ids[i] = flat_id % dim_lens[i];
        flat_id /= dim_lens[i];
    }
}

static int to_flat_id(
        int ndims,
        const int *dim_lens,
        const int *dim_ids)
{
    int flat_id = dim_ids[0];
    int mult = dim_lens[0];
    for (int i = 1; i < ndims; i++) {
        flat_id += dim_ids[i] * mult;
        mult *= dim_lens[i];
    }
    return flat_id;
}

417 418 419
void torus_collective_init(nodes_state * s,
           		   tw_lp * lp)
{
420
    // TODO: be annotation-aware somehow
421
    codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
422
    tw_lpid num_lps = codes_mapping_get_lp_count(grp_name, 0, LP_CONFIG_NM, s->anno, 0);
423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448

    int i;
   /* handle collective operations by forming a tree of all the LPs */
   /* special condition for root of the tree */
   if( s->node_id == 0)
    {
        s->parent_node_id = -1;
        s->is_root = 1;
   }
   else
   {
       s->parent_node_id = (s->node_id - ((s->node_id - 1) % TREE_DEGREE)) / TREE_DEGREE;
       s->is_root = 0;
   }
   s->children = (tw_lpid*)malloc(TREE_DEGREE * sizeof(tw_lpid));

   /* set the isleaf to zero by default */
   s->is_leaf = 1;
   s->num_children = 0;

   /* calculate the children of the current node. If its a leaf, no need to set children,
      only set isleaf and break the loop*/

   for( i = 0; i < TREE_DEGREE; i++ )
    {
        tw_lpid next_child = (TREE_DEGREE * s->node_id) + i + 1;
449
        if(next_child < num_lps)
450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469
        {
            s->num_children++;
            s->is_leaf = 0;
            s->children[i] = next_child;
        }
        else
           s->children[i] = -1;
    }

#if TORUS_COLLECTIVE_DEBUG == 1
   printf("\n LP %ld parent node id ", s->node_id);

   for( i = 0; i < TREE_DEGREE; i++ )
        printf(" child node ID %ld ", s->children[i]);
   printf("\n");

   if(s->is_leaf)
        printf("\n LP %ld is leaf ", s->node_id);
#endif
}
470
/* torus packet reverse event */
471 472 473 474 475 476
static void torus_packet_event_rc(tw_lp *sender)
{
  codes_local_latency_reverse(sender);
  return;
}

477 478 479 480 481 482
/* returns the torus message size */
static int torus_get_msg_sz(void)
{
   return sizeof(nodes_message);
}

483
/* torus packet event , generates a torus packet on the compute node */
484 485 486 487 488 489 490 491 492 493
static tw_stime torus_packet_event(
        model_net_request const * req,
        uint64_t message_offset,
        uint64_t packet_size,
        tw_stime offset,
        mn_sched_params const * sched_params,
        void const * remote_event,
        void const * self_event,
        tw_lp *sender,
        int is_last_pckt)
494
{
495 496
    (void)message_offset; // not using atm...
    (void)sched_params; // not using atm...
497 498 499 500
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    nodes_message * msg;
    char* tmp_ptr;
501

502
    xfer_to_nic_time = codes_local_latency(sender); /* Throws an error of found last KP time > current event time otherwise */
503 504 505 506
    //e_new = tw_event_new(local_nic_id, xfer_to_nic_time+offset, sender);
    //msg = tw_event_data(e_new);
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, TORUS, (void**)&msg, (void**)&tmp_ptr);
507 508 509 510
    strcpy(msg->category, req->category);
    msg->final_dest_gid = req->final_dest_lp;
    msg->dest_lp = req->dest_mn_lp;
    msg->sender_svr= req->src_lp;
511
    msg->sender_node = sender->gid;
512
    msg->packet_size = packet_size;
513
    msg->travel_start_time = tw_now(sender);
514 515
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
516
    msg->chunk_id = 0;
517
    msg->type = GENERATE;
518 519 520
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;

521
    if(is_last_pckt) /* Its the last packet so pass in remote event information*/
522 523 524 525 526 527 528 529 530 531 532 533 534 535 536
    {
        if(req->remote_event_size > 0)
        {
            msg->remote_event_size_bytes = req->remote_event_size;
            memcpy(tmp_ptr, remote_event, req->remote_event_size);
            tmp_ptr += req->remote_event_size;
        }
        if(req->self_event_size > 0)
        {
            msg->local_event_size_bytes = req->self_event_size;
            memcpy(tmp_ptr, self_event, req->self_event_size);
            tmp_ptr += req->self_event_size;
        }
        // printf("\n torus remote event %d local event %d last packet %d %lf ", msg->remote_event_size_bytes, msg->local_event_size_bytes, is_last_pckt, xfer_to_nic_time);
    }
537
    tw_event_send(e_new);
538
    return xfer_to_nic_time;
539
}
540

541
/*Sends a 8-byte credit back to the torus node LP that sent the message */
542 543
static void credit_send( nodes_state * s,
	    tw_lp * lp,
544 545 546 547 548 549 550
	    nodes_message * msg,
            int sq)
{
    tw_event * e;
    nodes_message *m;
    tw_stime ts;

551
    ts = (1.1 * g_tw_lookahead) + s->params->credit_delay + tw_rand_unif(lp->rng);
552
    e = model_net_method_event_new(msg->sender_node, ts, lp, TORUS,
553 554 555 556 557 558 559 560 561 562 563
        (void**)&m, NULL);
    if(sq == -1) {
        m->source_direction = msg->source_direction;
        m->source_dim = msg->source_dim;
    } else {
        m->source_direction = msg->saved_queue % 2;
        m->source_dim = msg->saved_queue / 2;
    }
    m->type = CREDIT;
    tw_event_send(e);
}
564
/*Initialize the torus model, this initialization part is borrowed from Ning's torus model */
565
static void torus_init( nodes_state * s,
566 567 568
	   tw_lp * lp )
{
    int i, j;
569
    char anno[MAX_NAME_LENGTH];
570

571 572
    rc_stack_create(&s->st);

573
    codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL, &mapping_type_id, anno, &mapping_rep_id, &mapping_offset);
574

575 576
    s->node_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);

577 578 579 580 581
    if (anno[0] == '\0'){
        s->anno = NULL;
        s->params = &all_params[num_params-1];
    }
    else{
582
        s->anno = strdup(anno);
583 584 585
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }
586

587 588
    // shorthand
    const torus_param *p = s->params;
589

590 591 592
    s->finished_chunks = 0;
    s->finished_packets = 0;

593 594 595 596
    s->neighbour_minus_lpID = (int*)malloc(p->n_dims * sizeof(int));
    s->neighbour_plus_lpID = (int*)malloc(p->n_dims * sizeof(int));
    s->dim_position = (int*)malloc(p->n_dims * sizeof(int));
    s->buffer = (int**)malloc(2*p->n_dims * sizeof(int*));
597
    s->next_link_available_time =
598
        (tw_stime**)malloc(2*p->n_dims * sizeof(tw_stime*));
599
    s->next_credit_available_time =
600
        (tw_stime**)malloc(2*p->n_dims * sizeof(tw_stime*));
601
    s->next_flit_generate_time =
602 603
        (tw_stime**)malloc(2*p->n_dims*sizeof(tw_stime*));

604 605 606 607 608
    for(i=0; i < 2*p->n_dims; i++)
    {
        s->buffer[i] = (int*)malloc(p->num_vc * sizeof(int));
        s->next_link_available_time[i] =
                (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
609
        s->next_credit_available_time[i] =
610
                (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
611
        s->next_flit_generate_time[i] =
612 613 614 615
                (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
    }
    s->terminal_length = (int*)malloc(2*p->n_dims*sizeof(int));
    s->queued_length = (int*)malloc(2*p->n_dims*sizeof(int));
616
    s->terminal_msgs =
617
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
618
    s->terminal_msgs_tail =
619
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
620
    s->pending_msgs =
621
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
622
    s->pending_msgs_tail =
623
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
624
    s->queued_msgs =
625
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
626
    s->queued_msgs_tail =
627
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
628

629
    s->busy_time =
630
        (tw_stime*)malloc(2*p->n_dims*sizeof(tw_stime));
631
    s->last_buf_full =
632 633
        (tw_stime*)malloc(2*p->n_dims*sizeof(tw_stime));

634
    for(i = 0; i < 2*p->n_dims; i++) {
635
        s->pending_msgs[i] =
636
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
637
        s->pending_msgs_tail[i] =
638
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
639
        s->queued_msgs[i] =
640
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
641
        s->queued_msgs_tail[i] =
642
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
643

644
    }
645
    s->other_msgs =
646
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
647
    s->other_msgs_tail =
648
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
649
    s->in_send_loop =
650 651 652
        (int *)malloc(2*p->n_dims*sizeof(int));

    s->link_traffic = (int64_t *)malloc(2*p->n_dims*sizeof(int64_t));
653 654
    s->total_data_sz = 0;

655
    for(i=0; i < 2*p->n_dims; i++)
656
    {
657 658 659
	s->buffer[i] = (int*)malloc(p->num_vc * sizeof(int));
	s->next_link_available_time[i] =
            (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
660
	s->next_credit_available_time[i] =
661
            (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
662
	s->next_flit_generate_time[i] =
663
            (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
664 665 666 667 668 669 670 671
        s->terminal_msgs[i] = NULL;
        s->terminal_msgs_tail[i] = NULL;
        s->other_msgs[i] = NULL;
        s->other_msgs_tail[i] = NULL;
        s->in_send_loop[i] = 0;
        s->terminal_length[i] = 0;
        s->queued_length[i] = 0;
        s->all_term_length = 0;
672 673
        s->busy_time[i] = 0;
        s->last_buf_full[i] = 0;
674 675
    }

676 677 678
    // calculate my torus coords
    to_dim_id(codes_mapping_get_lp_relative_id(lp->gid, 0, 1),
            s->params->n_dims, s->params->dim_length, s->dim_position);
679 680
    /* DEBUG */
    /*printf("%lu: my coords:", lp->gid);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
681 682 683 684
    for (i = 0; i < p->n_dims; i++)
        printf(" %d", s->dim_position[i]);
    printf("\n");
    */
685

686 687
  int temp_dim_pos[ p->n_dims ];
  for ( i = 0; i < p->n_dims; i++ )
688 689 690
    temp_dim_pos[ i ] = s->dim_position[ i ];

  // calculate minus neighbour's lpID
691
  for ( j = 0; j < p->n_dims; j++ )
692
    {
693 694
      temp_dim_pos[ j ] = (s->dim_position[ j ] -1 + p->dim_length[ j ]) %
          p->dim_length[ j ];
695

696 697 698 699 700 701 702 703 704 705
      s->neighbour_minus_lpID[j] =
          to_flat_id(p->n_dims, p->dim_length, temp_dim_pos);

      /* DEBUG
      printf(" minus neighbor: flat:%d lpid:%lu\n",
              s->neighbour_minus_lpID[j],
              codes_mapping_get_lpid_from_relative(s->neighbour_minus_lpID[j],
                      NULL, LP_CONFIG_NM, s->anno, 1));
      */

706 707 708
      temp_dim_pos[ j ] = s->dim_position[ j ];
    }
  // calculate plus neighbour's lpID
709
  for ( j = 0; j < p->n_dims; j++ )
710
    {
711 712
      temp_dim_pos[ j ] = ( s->dim_position[ j ] + 1 + p->dim_length[ j ]) %
          p->dim_length[ j ];
713

714 715 716 717 718 719 720 721 722
      s->neighbour_plus_lpID[j] =
          to_flat_id(p->n_dims, p->dim_length, temp_dim_pos);

      /* DEBUG
      printf(" plus neighbor: flat:%d lpid:%lu\n",
              s->neighbour_plus_lpID[j],
              codes_mapping_get_lpid_from_relative(s->neighbour_plus_lpID[j],
                      NULL, LP_CONFIG_NM, s->anno, 1));
      */
723 724 725 726 727

      temp_dim_pos[ j ] = s->dim_position[ j ];
    }

  //printf("\n");
728
  for( j=0; j < 2 * p->n_dims; j++ )
729
   {
730
    for( i = 0; i < p->num_vc; i++ )
731
     {
732
       s->buffer[ j ][ i ] = 0;
733
       s->next_link_available_time[ j ][ i ] = 0.0;
734
       s->next_credit_available_time[j][i] = 0.0;
735 736 737 738
       s->pending_msgs[j][i] = NULL;
       s->pending_msgs_tail[j][i] = NULL;
       s->queued_msgs[j][i] = NULL;
       s->queued_msgs_tail[j][i] = NULL;
739 740 741 742
     }
   }
  // record LP time
    s->packet_counter = 0;
743 744 745 746 747
    torus_collective_init(s, lp);
}


/* collective operation for the torus network */
748
void torus_collective(char const * category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
749 750 751 752 753 754 755
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    nodes_message * msg;
    tw_lpid local_nic_id;
    char* tmp_ptr;

756 757 758 759 760
    // TODO: be annotation-aware
    codes_mapping_get_lp_info(sender->gid, grp_name, &mapping_grp_id, NULL,
            &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
    codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
            mapping_rep_id, mapping_offset, &local_nic_id);
761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786

    xfer_to_nic_time = g_tw_lookahead + codes_local_latency(sender);
    e_new = model_net_method_event_new(local_nic_id, xfer_to_nic_time,
            sender, TORUS, (void**)&msg, (void**)&tmp_ptr);

    msg->remote_event_size_bytes = message_size;
    strcpy(msg->category, category);
    msg->sender_svr=sender->gid;
    msg->type = T_COLLECTIVE_INIT;

    tmp_ptr = (char*)msg;
    tmp_ptr += torus_get_msg_sz();
    if(remote_event_size > 0)
     {
            msg->remote_event_size_bytes = remote_event_size;
            memcpy(tmp_ptr, remote_event, remote_event_size);
            tmp_ptr += remote_event_size;
     }

    tw_event_send(e_new);
    return;
}

/* reverse for collective operation of the dragonfly network */
void torus_collective_rc(int message_size, tw_lp* sender)
{
787
    (void)message_size; // unneeded
788 789 790 791 792 793 794 795 796 797 798 799 800 801
     codes_local_latency_reverse(sender);
     return;
}

static void send_remote_event(nodes_state * s,
                        nodes_message * msg,
                        tw_lp * lp)
{
    // Trigger an event on receiving server
    if(msg->remote_event_size_bytes)
     {
            tw_event* e;
            tw_stime ts;
            nodes_message * m;
802
            ts = (1/s->params->link_bandwidth) * msg->remote_event_size_bytes;
803
            e = tw_event_new(s->origin_svr, ts, lp);
804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824
            m = tw_event_data(e);
            char* tmp_ptr = (char*)msg;
            tmp_ptr += torus_get_msg_sz();
            memcpy(m, tmp_ptr, msg->remote_event_size_bytes);
            tw_event_send(e);
     }
}

static void node_collective_init(nodes_state * s,
                        nodes_message * msg,
                        tw_lp * lp)
{
        tw_event * e_new;
        tw_lpid parent_nic_id;
        tw_stime xfer_to_nic_time;
        nodes_message * msg_new;
        int num_lps;

        msg->saved_collective_init_time = s->collective_init_time;
        s->collective_init_time = tw_now(lp);
	s->origin_svr = msg->sender_svr;
825

826 827 828 829
        if(s->is_leaf)
        {
            //printf("\n LP %ld sending message to parent %ld ", s->node_id, s->parent_node_id);
            /* get the global LP ID of the parent node */
830 831 832 833 834 835 836 837
            // TODO: be annotation-aware
            codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL,
                    &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
            num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM,
                    NULL, 1);
            codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
838 839 840 841 842 843 844

           /* send a message to the parent that the LP has entered the collective operation */
            xfer_to_nic_time = g_tw_lookahead + LEVEL_DELAY;
            //e_new = codes_event_new(parent_nic_id, xfer_to_nic_time, lp);
	    void* m_data;
	    e_new = model_net_method_event_new(parent_nic_id, xfer_to_nic_time,
            	lp, TORUS, (void**)&msg_new, (void**)&m_data);
845

846 847 848 849 850
            memcpy(msg_new, msg, sizeof(nodes_message));
	    if (msg->remote_event_size_bytes){
        	memcpy(m_data, model_net_method_get_edata(TORUS, msg),
                	msg->remote_event_size_bytes);
      	    }
851

852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867
            msg_new->type = T_COLLECTIVE_FAN_IN;
            msg_new->sender_node = s->node_id;

            tw_event_send(e_new);
        }
        return;
}

static void node_collective_fan_in(nodes_state * s,
                        tw_bf * bf,
                        nodes_message * msg,
                        tw_lp * lp)
{
        int i;
        s->num_fan_nodes++;

868 869 870 871 872
        // TODO: be annotation-aware
        codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL,
                &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
        int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM,
                NULL, 1);
873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890

        tw_event* e_new;
        nodes_message * msg_new;
        tw_stime xfer_to_nic_time;

        bf->c1 = 0;
        bf->c2 = 0;

        /* if the number of fanned in nodes have completed at the current node then signal the parent */
        if((s->num_fan_nodes == s->num_children) && !s->is_root)
        {
            bf->c1 = 1;
            msg->saved_fan_nodes = s->num_fan_nodes-1;
            s->num_fan_nodes = 0;
            tw_lpid parent_nic_id;
            xfer_to_nic_time = g_tw_lookahead + LEVEL_DELAY;

            /* get the global LP ID of the parent node */
891 892 893
            codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
894 895 896 897 898 899 900 901

           /* send a message to the parent that the LP has entered the collective operation */
            //e_new = codes_event_new(parent_nic_id, xfer_to_nic_time, lp);
            //msg_new = tw_event_data(e_new);
	    void * m_data;
      	    e_new = model_net_method_event_new(parent_nic_id,
              xfer_to_nic_time,
              lp, TORUS, (void**)&msg_new, &m_data);
902

903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919
            memcpy(msg_new, msg, sizeof(nodes_message));
            msg_new->type = T_COLLECTIVE_FAN_IN;
            msg_new->sender_node = s->node_id;

            if (msg->remote_event_size_bytes){
	        memcpy(m_data, model_net_method_get_edata(TORUS, msg),
        	        msg->remote_event_size_bytes);
      	   }
            tw_event_send(e_new);
      }

      /* root node starts off with the fan-out phase */
      if(s->is_root && (s->num_fan_nodes == s->num_children))
      {
           bf->c2 = 1;
           msg->saved_fan_nodes = s->num_fan_nodes-1;
           s->num_fan_nodes = 0;
920
           send_remote_event(s, msg, lp);
921 922 923 924 925 926 927 928

           for( i = 0; i < s->num_children; i++ )
           {
                tw_lpid child_nic_id;
                /* Do some computation and fan out immediate child nodes from the collective */
                xfer_to_nic_time = g_tw_lookahead + COLLECTIVE_COMPUTATION_DELAY + LEVEL_DELAY + tw_rand_exponential(lp->rng, (double)LEVEL_DELAY/50);

                /* get global LP ID of the child node */
929 930 931
                codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
                        s->children[i]/num_lps, (s->children[i] % num_lps),
                        &child_nic_id);
932 933 934 935 936 937 938 939 940 941 942 943 944
                //e_new = codes_event_new(child_nic_id, xfer_to_nic_time, lp);

                //msg_new = tw_event_data(e_new);
                void * m_data;
	        e_new = model_net_method_event_new(child_nic_id,
                xfer_to_nic_time,
		lp, TORUS, (void**)&msg_new, &m_data);

		memcpy(msg_new, msg, sizeof(nodes_message));
	        if (msg->remote_event_size_bytes){
	                memcpy(m_data, model_net_method_get_edata(TORUS, msg),
        	               msg->remote_event_size_bytes);
      		}
945

946 947 948 949 950 951
                msg_new->type = T_COLLECTIVE_FAN_OUT;
                msg_new->sender_node = s->node_id;

                tw_event_send(e_new);
           }
      }
952
}
953

954 955 956 957 958 959
static void node_collective_fan_out(nodes_state * s,
                        tw_bf * bf,
                        nodes_message * msg,
                        tw_lp * lp)
{
        int i;
960 961 962
        //TODO: be annotation-aware
        int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM,
                NULL, 1);
963 964 965
        bf->c1 = 0;
        bf->c2 = 0;

966
        send_remote_event(s, msg, lp);
967

968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983
        if(!s->is_leaf)
        {
            bf->c1 = 1;
            tw_event* e_new;
            nodes_message * msg_new;
            tw_stime xfer_to_nic_time;

           for( i = 0; i < s->num_children; i++ )
           {
                xfer_to_nic_time = g_tw_lookahead + TORUS_FAN_OUT_DELAY + tw_rand_exponential(lp->rng, (double)TORUS_FAN_OUT_DELAY/10);

                if(s->children[i] > 0)
                {
                        tw_lpid child_nic_id;

                        /* get global LP ID of the child node */
984 985 986
                        codes_mapping_get_lp_id(grp_name, LP_CONFIG_NM, NULL, 1,
                                s->children[i]/num_lps,
                                (s->children[i] % num_lps), &child_nic_id);
987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013
                        //e_new = codes_event_new(child_nic_id, xfer_to_nic_time, lp);
                        //msg_new = tw_event_data(e_new);
                        //memcpy(msg_new, msg, sizeof(nodes_message) + msg->remote_event_size_bytes);
			void* m_data;
			e_new = model_net_method_event_new(child_nic_id,
							xfer_to_nic_time,
					                lp, TORUS, (void**)&msg_new, &m_data);
		        memcpy(msg_new, msg, sizeof(nodes_message));
		        if (msg->remote_event_size_bytes){
			        memcpy(m_data, model_net_method_get_edata(TORUS, msg),
			                msg->remote_event_size_bytes);
      			}


                        msg_new->type = T_COLLECTIVE_FAN_OUT;
                        msg_new->sender_node = s->node_id;
                        tw_event_send(e_new);
                }
           }
         }
	//printf("\n Fan out phase completed %ld ", lp->gid);
        if(max_collective < tw_now(lp) - s->collective_init_time )
          {
              bf->c2 = 1;
              max_collective = tw_now(lp) - s->collective_init_time;
          }
}
1014

1015 1016
/*Returns the next neighbor to which the packet should be routed by using DOR (Taken from Ning's code of the torus model)*/
static void dimension_order_routing( nodes_state * s,
1017 1018
			     tw_lpid * dst_lp,
			     int * dim,
1019 1020
			     int * dir )
{
1021 1022
     int dest[s->params->n_dims];
     int dest_id;
1023

1024 1025 1026 1027
  /* dummys - check later */
  *dim = -1;
  *dir = -1;

1028 1029
  to_dim_id(codes_mapping_get_lp_relative_id(*dst_lp, 0, 1),
          s->params->n_dims, s->params->dim_length, dest);
1030

1031
  for(int i = 0; i < s->params->n_dims; i++ )
1032
    {
1033
      if ( s->dim_position[ i ] - dest[ i ] > s->params->half_length[ i ] )
1034 1035 1036 1037 1038 1039
	{
	  dest_id = s->neighbour_plus_lpID[ i ];
	  *dim = i;
	  *dir = 1;
	  break;
	}
1040
      if ( s->dim_position[ i ] - dest[ i ] < -s->params->half_length[ i ] )
1041 1042 1043 1044 1045 1046
	{
	  dest_id = s->neighbour_minus_lpID[ i ];
	  *dim = i;
	  *dir = 0;
	  break;
	}
1047 1048
      if ( ( s->dim_position[i] - dest[i] <= s->params->half_length[i] ) &&
              ( s->dim_position[ i ] - dest[ i ] > 0 ) )
1049 1050 1051 1052 1053 1054
	{
	  dest_id = s->neighbour_minus_lpID[ i ];
	  *dim = i;
	  *dir = 0;
	  break;
	}
1055 1056
      if (( s->dim_position[i] - dest[i] >= -s->params->half_length[i] ) &&
              ( s->dim_position[ i ] - dest[ i ] < 0) )
1057 1058 1059 1060 1061 1062 1063
	{
	  dest_id = s->neighbour_plus_lpID[ i ];
	  *dim = i;
	  *dir = 1;
	  break;
	}
    }
1064 1065

  assert(*dim != -1 && *dir != -1);
1066 1067
  *dst_lp = codes_mapping_get_lpid_from_relative(dest_id, NULL, LP_CONFIG_NM,
          s->anno, 1);
1068
}
1069 1070 1071
static void packet_generate( nodes_state * ns,
        tw_bf * bf,
        nodes_message * msg,
1072 1073
        tw_lp * lp)
{
1074
    int tmp_dir=-1, tmp_dim=-1, queue, total_event_size;
1075
    tw_stime ts;
1076
    tw_event * e;