torus.c 59.5 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 19 20

#define DEBUG 1
#define MEAN_INTERVAL 100
21
#define TRACE -1
22

23
#define STATICQ 0
24 25 26
/* collective specific parameters */
#define TREE_DEGREE 4
#define LEVEL_DELAY 1000
27
#define TORUS_COLLECTIVE_DEBUG 0
28 29 30 31
#define NUM_COLLECTIVES  1
#define COLLECTIVE_COMPUTATION_DELAY 5700
#define TORUS_FAN_OUT_DELAY 20.0

32 33 34
#define LP_CONFIG_NM (model_net_lp_config_names[TORUS])
#define LP_METHOD_NM (model_net_method_names[TORUS])

35 36
static double maxd(double a, double b) { return a < b ? b : a; }

37 38 39 40 41
enum routing_algo
{
    STATIC=0,
    ADAPTIVE
};
42 43
/* Torus network model implementation of codes, implements the modelnet API */

44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63
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);
}

64 65 66 67 68 69
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 */
70
    double cn_bandwidth; /* injection bandwidth */
71 72 73 74 75 76 77 78 79 80 81 82 83 84
    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;
85 86 87 88
    double cn_delay;

    double router_delay;
    int routing;
89
};
90 91

/* codes mapping group name, lp type name */
92
static char grp_name[MAX_NAME_LENGTH];
93
/* codes mapping group id, lp type id, repetition id and offset */
94
int mapping_grp_id, mapping_rep_id, mapping_type_id, mapping_offset;
95 96 97 98

/* 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;
99
static tw_stime         max_collective = 0;
100 101 102 103 104 105

/* 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;

106 107 108 109 110
/* annotation-specific parameters (unannotated entry occurs at the 
 * last index) */
static uint64_t                  num_params = 0;
static torus_param             * all_params = NULL;
static const config_anno_map_t * anno_map   = NULL;
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126

typedef struct nodes_state nodes_state;

/* state of a torus node */
struct nodes_state
{
  /* counts the number of packets sent from this compute node */
  unsigned long long packet_counter;            
  /* availability time of each torus link */
  tw_stime** next_link_available_time; 
  /* 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;
127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142
  /* Head and tail of the terminal messages list */
  nodes_message_list **terminal_msgs;
  nodes_message_list **terminal_msgs_tail;
  int all_term_length;
  int issueIdle;
  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;
143 144 145 146 147 148 149 150
  /* 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];
151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176
   /* for collective operations */

  /* collective init time */
  tw_stime collective_init_time;

  /* node ID in the tree */ 
   tw_lpid node_id;

   /* messages sent & received in collectives may get interchanged several times so we have to save the 
     origin server information in the node's state */
   tw_lpid origin_svr; 
  
  /* 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;
177 178 179 180 181

   /* LPs annotation */
   const char * anno;
   /* LPs configuration */
   const torus_param * params;
182 183
};

184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
static void append_to_node_message_list(  
        nodes_state *ns, 
        nodes_message_list ** thisq,
        nodes_message_list ** thistail,
        int index, 
        nodes_message_list *msg) {
    if(thisq[index] == NULL) {
        thisq[index] = msg;
    } else {
        thistail[index]->next = msg;
        msg->prev = thistail[index];
    } 
    thistail[index] = msg;
}
static void prepend_to_node_message_list(  
        nodes_state *ns, 
        nodes_message_list ** thisq,
        nodes_message_list ** thistail,
        int index, 
        nodes_message_list *msg) {
    if(thisq[index] == NULL) {
        thistail[index] = msg;
    } else {
        thisq[index]->prev = msg;
        msg->next = thisq[index];
    } 
    thisq[index] = msg;
}
static void create_prepend_to_node_message_list(
        nodes_state *ns, 
        nodes_message_list ** thisq,
        nodes_message_list ** thistail,
        int index, 
        nodes_message *msg) {
    nodes_message_list* new_entry = (nodes_message_list*)malloc(
        sizeof(nodes_message_list));
    init_nodes_message_list(new_entry, msg);
    if(msg->remote_event_size_bytes) {
        void *m_data = model_net_method_get_edata(TORUS, msg);
        new_entry->event_data = (void*)malloc(msg->remote_event_size_bytes);
        memcpy(new_entry->event_data, m_data, msg->remote_event_size_bytes);
    }
    prepend_to_node_message_list(ns, thisq, thistail, index, new_entry);
}

static nodes_message_list* return_head(
        nodes_state *ns,
        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_state *ns,
        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;
}
static void copy_nodes_list_entry( nodes_message_list *cur_entry,
    nodes_message *msg) {
    nodes_message *cur_msg = &cur_entry->msg;
    strcpy(msg->category, cur_msg->category);
    msg->travel_start_time = cur_msg->travel_start_time;
    msg->packet_ID = cur_msg->packet_ID;    
    msg->final_dest_gid = cur_msg->final_dest_gid;
    msg->dest_lp = cur_msg->dest_lp;
    msg->sender_svr = cur_msg->sender_svr;
    msg->sender_node = cur_msg->sender_node;
    msg->my_N_hop = cur_msg->my_N_hop;
    msg->next_stop = cur_msg->next_stop;
    msg->packet_size = cur_msg->packet_size;
    msg->chunk_id = cur_msg->chunk_id;
    msg->is_pull = cur_msg->is_pull;
    msg->pull_size = cur_msg->pull_size;
    msg->local_event_size_bytes = cur_msg->local_event_size_bytes;
    msg->remote_event_size_bytes = cur_msg->remote_event_size_bytes;

    if(msg->local_event_size_bytes +  msg->remote_event_size_bytes > 0) {
        void *m_data = model_net_method_get_edata(TORUS, msg);
        memcpy(m_data, cur_entry->event_data, 
            msg->local_event_size_bytes +  msg->remote_event_size_bytes);
    }
}

289 290 291 292
static void torus_read_config(
        const char         * anno,
        torus_param        * params){
    char dim_length_str[MAX_NAME_LENGTH];
293
    int i;
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320

    // shorthand
    torus_param *p = params;

    configuration_get_value_int(&config, "PARAMS", "n_dims", anno, &p->n_dims);
    if(!p->n_dims) {
        p->n_dims = 4; /* a 4-D torus */
        fprintf(stderr, 
                "Warning: Number of dimensions not specified, setting to %d\n",
                p->n_dims);
    }

    configuration_get_value_double(&config, "PARAMS", "link_bandwidth", anno,
            &p->link_bandwidth);
    if(!p->link_bandwidth) {
        p->link_bandwidth = 2.0; /*default bg/q configuration */
        fprintf(stderr, "Link bandwidth not specified, setting to %lf\n",
                p->link_bandwidth);
    }

    configuration_get_value_int(&config, "PARAMS", "buffer_size", anno, &p->buffer_size);
    if(!p->buffer_size) {
        p->buffer_size = 2048;
        fprintf(stderr, "Buffer size not specified, setting to %d",
                p->buffer_size);
    }

321

322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346
    configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(!p->chunk_size) {
        p->chunk_size = 32;
        fprintf(stderr, "Warning: Chunk size not specified, setting to %d\n",
                p->chunk_size);
    }
    configuration_get_value_int(&config, "PARAMS", "num_vc", anno, &p->num_vc);
    if(!p->num_vc) {
        /* by default, we have one for taking packets,
         * another for taking credit*/
        p->num_vc = 1;
        fprintf(stderr, "Warning: num_vc not specified, setting to %d\n",
                p->num_vc);
    }

    int rc = configuration_get_value(&config, "PARAMS", "dim_length", anno,
            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)
347
    {
348 349 350 351 352 353 354 355 356 357 358 359
        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
   
    // factor is an exclusive prefix product
360 361
    p->router_delay = 50;
    p->routing = STATIC;
362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382
    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
    p->head_delay = (1.0 / p->link_bandwidth) * p->chunk_size;
    p->credit_delay = (1.0 / p->link_bandwidth) * p->chunk_size;
}

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));

    for (uint64_t i = 0; i < anno_map->num_annos; i++){
383
        const char * anno = anno_map->annotations[i].ptr;
384 385 386 387
        torus_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        torus_read_config(NULL, &all_params[anno_map->num_annos]);
388
    }
389 390
}

391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417
/* 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;
}

418 419 420
void torus_collective_init(nodes_state * s,
           		   tw_lp * lp)
{
421 422
    // TODO: be annotation-aware somehow 
    codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
423
    int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM, s->anno, 0);
424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472
    int num_reps = codes_mapping_get_group_reps(grp_name);
    s->node_id = (mapping_rep_id * num_lps) + mapping_offset;

    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;
        if(next_child < (num_lps * num_reps))
        {
            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
}
473
/* torus packet reverse event */
474 475 476 477 478 479
static void torus_packet_event_rc(tw_lp *sender)
{
  codes_local_latency_reverse(sender);
  return;
}

480 481 482 483 484 485
/* returns the torus message size */
static int torus_get_msg_sz(void)
{
   return sizeof(nodes_message);
}

486
/* torus packet event , generates a torus packet on the compute node */
487 488 489 490 491 492 493 494 495 496
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)
497 498 499 500 501
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    nodes_message * msg;
    char* tmp_ptr;
502

503
    xfer_to_nic_time = g_tw_lookahead + codes_local_latency(sender); /* Throws an error of found last KP time > current event time otherwise */
504 505 506 507
    //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);
508 509 510 511
    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;
512
    msg->sender_node = sender->gid;
513
    msg->packet_size = packet_size;
514
    msg->travel_start_time = tw_now(sender);
515 516
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
517
    msg->chunk_id = 0;
518
    msg->type = GENERATE;
519 520 521
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;

522
    if(is_last_pckt) /* Its the last packet so pass in remote event information*/
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537
    {
        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);
    }
538
    tw_event_send(e_new);
539
    return xfer_to_nic_time;
540
}
541

542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567
/*Sends a 8-byte credit back to the torus node LP that sent the message */
static void credit_send( nodes_state * s, 
	    tw_bf * bf, 
	    tw_lp * lp, 
	    nodes_message * msg,
            int sq)
{
    tw_event * e;
    nodes_message *m;
    tw_stime ts;

    ts = g_tw_lookahead + s->params->credit_delay + tw_rand_unif(lp->rng);
    e = model_net_method_event_new(msg->sender_node, ts, lp, TORUS, 
        (void**)&m, NULL);
    if(sq == -1) {
        m->source_direction = msg->source_direction;
        m->source_dim = msg->source_dim;
        m->source_channel = msg->source_channel;
    } else {
        m->source_direction = msg->saved_queue % 2;
        m->source_dim = msg->saved_queue / 2;
        m->source_channel = msg->saved_channel;
    }
    m->type = CREDIT;
    tw_event_send(e);
}
568 569 570 571 572
/*Initialize the torus model, this initialization part is borrowed from Ning's torus model */
static void torus_init( nodes_state * s, 
	   tw_lp * lp )
{
    int i, j;
573
    char anno[MAX_NAME_LENGTH];
574

575 576 577 578 579 580
    codes_mapping_get_lp_info(lp->gid, grp_name, &mapping_grp_id, NULL, &mapping_type_id, anno, &mapping_rep_id, &mapping_offset);
    if (anno[0] == '\0'){
        s->anno = NULL;
        s->params = &all_params[num_params-1];
    }
    else{
581
        s->anno = strdup(anno);
582 583 584
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }
585

586 587
    // shorthand
    const torus_param *p = s->params;
588

589 590 591 592 593 594 595 596 597 598 599
    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*));
    s->next_link_available_time = 
        (tw_stime**)malloc(2*p->n_dims * sizeof(tw_stime*));
    s->next_credit_available_time = 
        (tw_stime**)malloc(2*p->n_dims * sizeof(tw_stime*));
    s->next_flit_generate_time = 
        (tw_stime**)malloc(2*p->n_dims*sizeof(tw_stime*));

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
    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));
        s->next_credit_available_time[i] = 
                (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
        s->next_flit_generate_time[i] = 
                (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));
    s->terminal_msgs = 
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
    s->terminal_msgs_tail = 
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
    s->pending_msgs = 
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
    s->pending_msgs_tail = 
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
    s->queued_msgs = 
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
    s->queued_msgs_tail = 
        (nodes_message_list***)malloc(2*p->n_dims*sizeof(nodes_message_list**));
    for(i = 0; i < 2*p->n_dims; i++) {
        s->pending_msgs[i] = 
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
        s->pending_msgs_tail[i] = 
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
        s->queued_msgs[i] = 
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
        s->queued_msgs_tail[i] = 
            (nodes_message_list**)malloc(p->num_vc*sizeof(nodes_message_list*));
    }
    s->other_msgs = 
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
    s->other_msgs_tail = 
        (nodes_message_list**)malloc(2*p->n_dims*sizeof(nodes_message_list*));
    s->in_send_loop = 
        (int *)malloc(2*p->n_dims*sizeof(int));

    s->link_traffic = (int64_t *)malloc(2*p->n_dims*sizeof(int64_t));
642
    for(i=0; i < 2*p->n_dims; i++)
643
    {
644 645 646 647 648 649 650
	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));
	s->next_credit_available_time[i] = 
            (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
	s->next_flit_generate_time[i] = 
            (tw_stime*)malloc(p->num_vc * sizeof(tw_stime));
651 652 653 654 655 656 657 658 659
        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;
        s->issueIdle = 0;
660 661
    }

662 663 664
    // 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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
665 666 667 668 669 670
    /* DEBUG
    printf("%lu: my coords:", lp->gid);
    for (i = 0; i < p->n_dims; i++)
        printf(" %d", s->dim_position[i]);
    printf("\n");
    */
671

672 673
  int temp_dim_pos[ p->n_dims ];
  for ( i = 0; i < p->n_dims; i++ )
674 675 676
    temp_dim_pos[ i ] = s->dim_position[ i ];

  // calculate minus neighbour's lpID
677
  for ( j = 0; j < p->n_dims; j++ )
678
    {
679 680
      temp_dim_pos[ j ] = (s->dim_position[ j ] -1 + p->dim_length[ j ]) %
          p->dim_length[ j ];
681

682 683 684 685 686 687 688 689 690 691
      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));
      */

692 693 694
      temp_dim_pos[ j ] = s->dim_position[ j ];
    }
  // calculate plus neighbour's lpID
695
  for ( j = 0; j < p->n_dims; j++ )
696
    {
697 698
      temp_dim_pos[ j ] = ( s->dim_position[ j ] + 1 + p->dim_length[ j ]) %
          p->dim_length[ j ];
699

700 701 702 703 704 705 706 707 708
      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));
      */
709 710 711 712 713

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

  //printf("\n");
714
  for( j=0; j < 2 * p->n_dims; j++ )
715
   {
716
    for( i = 0; i < p->num_vc; i++ )
717 718 719 720
     {
       s->buffer[ j ][ i ] = 0; 
       s->next_link_available_time[ j ][ i ] = 0.0;
       s->next_credit_available_time[j][i] = 0.0; 
721 722 723 724
       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;
725 726 727 728
     }
   }
  // record LP time
    s->packet_counter = 0;
729 730 731 732 733
    torus_collective_init(s, lp);
}


/* collective operation for the torus network */
734
void torus_collective(char const * category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
735 736 737 738 739 740 741
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    nodes_message * msg;
    tw_lpid local_nic_id;
    char* tmp_ptr;

742 743 744 745 746
    // 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);
747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787

    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)
{
     codes_local_latency_reverse(sender);
     return;
}

static void send_remote_event(nodes_state * s,
                        tw_bf * bf,
                        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;
788
            ts = (1/s->params->link_bandwidth) * msg->remote_event_size_bytes;
789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816
            e = codes_event_new(s->origin_svr, ts, lp);
            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,
                        tw_bf * bf,
                        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;
	
        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 */
817 818 819 820 821 822 823 824
            // 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);
825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854

           /* 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);
	    	
            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_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++;

855 856 857 858 859
        // 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);
860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877

        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 */
878 879 880
            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);
881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916

           /* 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);
	    
            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;
           send_remote_event(s, bf, msg, lp);

           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 */
917 918 919
                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);
920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939
                //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);
      		}
		
                msg_new->type = T_COLLECTIVE_FAN_OUT;
                msg_new->sender_node = s->node_id;

                tw_event_send(e_new);
           }
      }
940
}
941 942 943 944 945 946 947
	     
static void node_collective_fan_out(nodes_state * s,
                        tw_bf * bf,
                        nodes_message * msg,
                        tw_lp * lp)
{
        int i;
948 949 950
        //TODO: be annotation-aware
        int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM,
                NULL, 1);
951 952 953 954
        bf->c1 = 0;
        bf->c2 = 0;

        send_remote_event(s, bf, msg, lp);
955

956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971
        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 */
972 973 974
                        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);
975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002
                        //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;
          }
}
    
1003 1004 1005 1006 1007 1008
/*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,
			     tw_lpid * dst_lp, 
			     int * dim, 
			     int * dir )
{
1009 1010
     int dest[s->params->n_dims];
     int dest_id;
1011

1012 1013 1014 1015
  /* dummys - check later */
  *dim = -1;
  *dir = -1;

1016 1017
  to_dim_id(codes_mapping_get_lp_relative_id(*dst_lp, 0, 1),
          s->params->n_dims, s->params->dim_length, dest);
1018

1019
  for(int i = 0; i < s->params->n_dims; i++ )
1020
    {
1021
      if ( s->dim_position[ i ] - dest[ i ] > s->params->half_length[ i ] )
1022 1023 1024 1025 1026 1027
	{
	  dest_id = s->neighbour_plus_lpID[ i ];
	  *dim = i;
	  *dir = 1;
	  break;
	}
1028
      if ( s->dim_position[ i ] - dest[ i ] < -s->params->half_length[ i ] )
1029 1030 1031 1032 1033 1034
	{
	  dest_id = s->neighbour_minus_lpID[ i ];
	  *dim = i;
	  *dir = 0;
	  break;
	}
1035 1036
      if ( ( s->dim_position[i] - dest[i] <= s->params->half_length[i] ) &&
              ( s->dim_position[ i ] - dest[ i ] > 0 ) )
1037 1038 1039 1040 1041 1042
	{
	  dest_id = s->neighbour_minus_lpID[ i ];
	  *dim = i;
	  *dir = 0;
	  break;
	}
1043 1044
      if (( s->dim_position[i] - dest[i] >= -s->params->half_length[i] ) &&
              ( s->dim_position[ i ] - dest[ i ] < 0) )
1045 1046 1047 1048 1049 1050 1051
	{
	  dest_id = s->neighbour_plus_lpID[ i ];
	  *dim = i;
	  *dir = 1;
	  break;
	}
    }
1052 1053

  assert(*dim != -1 && *dir != -1);
1054 1055
  *dst_lp = codes_mapping_get_lpid_from_relative(dest_id, NULL, LP_CONFIG_NM,
          s->anno, 1);
1056
}
1057 1058 1059 1060 1061 1062 1063 1064
static void enqueue_msg( nodes_state * ns, 
        tw_bf * bf, 
        nodes_message * msg, 
        tw_lp * lp,
        tw_lpid dst_lp,
        tw_stime nic_ts) {
    
    int j, tmp_dir=-1, tmp_dim=-1, queue, total_event_size;
1065
    tw_stime ts;
1066
    tw_event * e;
1067
    nodes_message *m;
1068 1069 1070
    
    dimension_order_routing(ns, &dst_lp, &tmp_dim, &tmp_dir);
    queue = tmp_dir + ( tmp_dim * 2 );
1071

1072
    msg->packet_ID = lp->gid + g_tw_nlp * ns->packet_counter;
1073
    msg->my_N_hop = 0;
1074

1075 1076
    int num_chunks = msg->packet_size/ns->params->chunk_size;
    if(msg->packet_size % ns->params->chunk_size)
1077
        num_chunks++;
1078

1079
    ns->packet_counter++;
1080

1081 1082 1083
    if(msg->packet_ID == TRACE)
        printf("\n packet generated %lld at lp %d dest %d final dest %d", 
        msg->packet_ID, (int)lp->gid, (int)dst_lp, (int)msg->dest_lp);
1084
       
1085 1086 1087
    msg->source_direction = tmp_dir;
    msg->source_dim = tmp_dim;
    msg->next_stop = dst_lp;
1088

1089 1090 1091
    for(j = 0; j < num_chunks; j++) { 
        nodes_message_list * cur_chunk = (nodes_message_list *)malloc( 
                sizeof(nodes_message_list));
1092

1093
        init_nodes_message_list(cur_chunk, msg);
1094

1095 1096 1097 1098
        if(msg->remote_event_size_bytes + msg->local_event_size_bytes > 0) {
            cur_chunk->event_data = (char*)malloc(
                msg->remote_event_size_bytes + msg->local_event_size_bytes);
        }
1099

1100 1101 1102 1103 1104 1105 1106 1107 1108 1109
        void *m_data_src = model_net_method_get_edata(TORUS, msg);
        if (msg->remote_event_size_bytes){
            memcpy(cur_chunk->event_data, m_data_src, msg->remote_event_size_bytes);
        }
        if (msg->local_event_size_bytes){
            m_data_src = (char*)m_data_src + msg->remote_event_size_bytes;
            memcpy((char*)cur_chunk->event_data + msg->remote_event_size_bytes, 
                m_data_src, msg->local_event_size_bytes);
        }
        cur_chunk->msg.chunk_id = j;
1110

1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124
        append_to_node_message_list(ns, ns->terminal_msgs,
            ns->terminal_msgs_tail, queue, cur_chunk);
        ns->terminal_length[queue] += ns->params->chunk_size;
        ns->all_term_length += ns->params->chunk_size;
    }
  
  if(ns->all_term_length < 2 * ns->params->buffer_size) 
  {
    model_net_method_idle_event(nic_ts, 0, lp);
  } 
  else 
  {
    bf->c11 = 1;
    ns->issueIdle = 1;
1125
  }
1126

1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137
   if(ns->in_send_loop[queue] == 0) {
       bf->c8 = 1;
       ts = g_tw_lookahead + ns->params->cn_delay * msg->packet_size + tw_rand_unif(lp->rng);
       e = model_net_method_event_new(lp->gid, ts, lp, TORUS, (void**)&m, NULL);
       m->type = SEND;
       m->source_direction = tmp_dir;
       m->source_dim = tmp_dim;
       ns->in_send_loop[queue] = 1;
       tw_event_send(e);
   }

1138
   /* record the statistics of the generated packets */
1139
   total_event_size = model_net_get_msg_sz(TORUS) + msg->remote_event_size_bytes + msg->local_event_size_bytes;   
1140
   mn_stats* stat;
1141
   stat = model_net_find_stats(msg->category, ns->torus_stats_array);
1142 1143
   stat->send_count++;  
   stat->send_bytes += msg->packet_size;
1144
   stat->send_time += (1/ns->params->link_bandwidth) * msg->packet_size;
1145 1146
   /* record the maximum ROSS event size */
   if(stat->max_event_size < total_event_size)
1147
	   stat->max_event_size = total_event_size;
1148
}
1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159
   /* record the maximum ROSS event size */

/*Generates a packet. If there is a buffer slot available, then the packet is 
injected in the network. Else, a buffer overflow exception is thrown.
TODO: We might want to modify this so that if the buffer is full, the packet
injection is delayed in turn slowing down the injection rate. The average achieved
injection rate can be reported at the end of the simulation. */
static void packet_generate( nodes_state * s, 
		tw_bf * bf, 
		nodes_message * msg, 
		tw_lp * lp )
1160 1161
{
    tw_stime ts;
1162 1163
    
    ts = g_tw_lookahead + s->params->cn_delay * msg->packet_size + tw_rand_unif(lp->rng);
1164

1165 1166
    tw_lpid dst_lp = msg->dest_lp;
    enqueue_msg( s, bf, msg, lp, dst_lp, ts );
1167
}
1168 1169
/* send a packet from one torus node to another torus node
 A packet can be up to 256 bytes on BG/L and BG/P and up to 512 bytes on BG/Q */
1170 1171 1172 1173 1174 1175 1176 1177 1178
static void packet_send( nodes_state * s, 
	         tw_bf * bf, 
		 nodes_message * msg, 
		 tw_lp * lp )
{ 
    int tmp_dir, tmp_dim;
    tw_stime ts;
    tw_event *e;
    nodes_message *m;
1179
    int isT = 0;
1180

1181 1182
    tmp_dim = msg->source_dim; 
    tmp_dir = msg->source_direction;
1183

1184
    int queue = tmp_dir + ( tmp_dim * 2 );
1185

1186 1187 1188 1189 1190 1191 1192
    if(s->pending_msgs[queue][STATICQ] == NULL 
        && s->terminal_msgs[queue] == NULL) {
        bf->c1 = 1;
        s->in_send_loop[queue] = 0;
        //printf("[%d] Empty queue; return\n", lp->gid);
        return;
    }
1193

1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210
    nodes_message_list *cur_entry = s->pending_msgs[queue][STATICQ];
            
    if(s->params->routing == STATIC && cur_entry == NULL) {
                if((s->buffer[queue][STATICQ] + (2 * s->params->chunk_size) <= s->params->buffer_size)) {
                    bf->c3 = 1;
                    s->buffer[queue][STATICQ] += s->params->chunk_size;
                    cur_entry = s->terminal_msgs[queue];
                    isT = 1;
                }
            if (cur_entry == NULL) {
                bf->c4 = 1;
                s->in_send_loop[queue] = 0;
                return;
            }
        } else {
            bf->c5 = 1;
        }
1211

1212 1213 1214
    int num_chunks = cur_entry->msg.packet_size/s->params->chunk_size;
    if(cur_entry->msg.packet_size % s->params->chunk_size)
        num_chunks++;
1215

1216 1217 1218 1219 1220 1221 1222
    double bytetime;
    if((cur_entry->msg.packet_size % s->params->chunk_size) && (cur_entry->msg.chunk_id == num_chunks - 1)) 
    {
        bytetime = s->params->head_delay * (cur_entry->msg.packet_size % s->params->chunk_size);
    } 
    else 
        bytetime = s->params->head_delay * s->params->chunk_size;
1223
    
1224
    ts = g_tw_lookahead + tw_rand_unif( lp->rng) + bytetime + s->params->router_delay;
1225

1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238
    //For reverse computation 
    msg->saved_available_time = s->next_link_available_time[queue][0];
    s->next_link_available_time[queue][0] = maxd( s->next_link_available_time[queue][0], tw_now(lp) );
    s->next_link_available_time[queue][0] += ts;
    
    void * m_data;
    ts = s->next_link_available_time[queue][0] - tw_now(lp);
    e = model_net_method_event_new(cur_entry->msg.next_stop, ts,
            lp, TORUS, (void**)&m, &m_data);
    memcpy(m, &cur_entry->msg, sizeof(nodes_message));
    if (m->remote_event_size_bytes){
        memcpy(m_data, cur_entry->event_data, m->remote_event_size_bytes);
    }
1239

1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263
    m->type = ARRIVAL;
    m->source_channel = STATICQ;

    if(msg->packet_ID == TRACE)
        printf("\n lp %d packet %lld flit id %d being sent to %d after time %lf ", 
            (int) lp->gid, m->packet_ID, m->chunk_id, m->next_stop, 
            s->next_link_available_time[queue][0] - tw_now(lp)); 

    m->sender_node = lp->gid;
    m->local_event_size_bytes = 0; /* We just deliver the local event here */

    //printf("[%d] Packet sent: source %d dest %d chunk %d\n", lp->gid, 
    //    m->sender_svr, m->final_dest_gid, m->chunk_id);
    tw_event_send( e );
    if((cur_entry->msg.packet_size % s->params->chunk_size) && (cur_entry->msg.chunk_id == num_chunks - 1)) {
        bf->c20 = 1;
        s->link_traffic[queue] += cur_entry->msg.packet_size %
            s->params->chunk_size;
    } else {
        bf->c21 = 1;
        s->link_traffic[queue] += s->params->chunk_size;
    }
    
    if(cur_entry->msg.chunk_id == num_chunks - 1)
1264
    {
1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323
        /* Invoke an event on the sending server */
        if(cur_entry->msg.local_event_size_bytes > 0)
        {
            bf->c6 = 1;
            tw_event* e_new;
            nodes_message* m_new;
            void* local_event;
            double tsT = codes_local_latency(lp);
            e_new = tw_event_new(cur_entry->msg.sender_svr, tsT, lp);
            m_new = tw_event_data(e_new);
            local_event = (char*)cur_entry->event_data + 
                cur_entry->msg.remote_event_size_bytes;
            memcpy(m_new, local_event, cur_entry->msg.local_event_size_bytes);
            tw_event_send(e_new);
        }
    }

    if(isT) {
        cur_entry = return_head(s, s->terminal_msgs, s->terminal_msgs_tail, 
            queue);
        s->terminal_length[queue] -= s->params->chunk_size; 
        s->all_term_length -= s->params->chunk_size;
    } else {
        bf->c8 = 1;
        cur_entry = return_head(s, s->pending_msgs[queue],
            s->pending_msgs_tail[queue], STATICQ);
    }

    /* store data for reverse handler */
    copy_nodes_list_entry(cur_entry, msg);
    delete_nodes_message_list(cur_entry);
  
    if(s->issueIdle && isT) {
        bf->c22 = 1;
        s->issueIdle = 0;
        model_net_method_idle_event(codes_local_latency(lp), 0, lp);
    }

    if(isT) {
        cur_entry = s->terminal_msgs[queue];
    } else {
        cur_entry = s->pending_msgs[queue][STATICQ];
            if(cur_entry == NULL) {
                cur_entry = s->terminal_msgs[queue];
            }
    }

    if(cur_entry != NULL) {
        bf->c9 = 1;
        ts = ts + tw_rand_unif( lp->rng) * .1;
        e = model_net_method_event_new(lp->gid, ts, lp, TORUS, (void**)&m, NULL);
        m->type = SEND;
        m->source_direction = tmp_dir;
        m->source_dim = tmp_dim;
        tw_event_send(e);
    } else {
        bf->c10 = 1;
        s->in_send_loop[queue] = 0;
        return;
1324
    }
1325 1326
}

1327 1328 1329
/*Processes the packet after it arrives from the neighboring torus node 
 * routes it to the next compute node if this is not the destination
 * OR if this is the destination then a remote event at the server is issued. */
1330 1331 1332 1333 1334 1335 1336 1337 1338
static void packet_arrive( nodes_state * s, 
		    tw_bf * bf, 
		    nodes_message * msg, 
		    tw_lp * lp )
{
  bf->c2 = 0;
  tw_event *e;
  tw_stime ts;
  nodes_message *m;
1339
  mn_stats* stat;
1340 1341 1342 1343 1344

  msg->my_N_hop++;
  ts = 0.1 + tw_rand_exponential(lp->rng, MEAN_INTERVAL/200);
  if(msg->packet_ID == TRACE)
	  printf("\n packet arrived at lp %d final dest %d ", (int)lp->gid, (int)msg->dest_lp);
1345
       
1346
  if( lp->gid == msg->dest_lp )
1347 1348 1349 1350
    {
        bf->c1 = 1;
        credit_send( s, bf, lp, msg, -1);

1351 1352 1353 1354
        uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
        if(msg->packet_size % s->params->chunk_size)
            num_chunks++;

1355 1356 1357
        if( msg->chunk_id == num_chunks - 1 )    
        {
	    bf->c2 = 1;
1358 1359 1360 1361 1362 1363
	    stat = model_net_find_stats(msg->category, s->torus_stats_array);
	    stat->recv_count++;
	    stat->recv_bytes += msg->packet_size;
	    stat->recv_time += tw_now( lp ) - msg->travel_start_time;

	    /*count the number of packets completed overall*/
1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375
	    N_finished_packets++;
	    total_time += tw_now( lp ) - msg->travel_start_time;
	    total_hops += msg->my_N_hop;

	    if (max_latency < tw_now( lp ) - msg->travel_start_time) {
		  bf->c3 = 1;
		  msg->saved_available_time = max_latency;
	          max_latency=tw_now( lp ) - msg->travel_start_time;
     		}
	    // Trigger an event on receiving server
	    if(msg->remote_event_size_bytes)
	    {
1376
               void *tmp_ptr = model_net_method_get_edata(TORUS, msg);
1377
               if (msg->is_pull){
1378
                   int net_id = model_net_get_id(LP_METHOD_NM);
1379 1380 1381 1382
                   struct codes_mctx mc_dst =
                       codes_mctx_set_global_direct(msg->sender_node);
                   struct codes_mctx mc_src =
                       codes_mctx_set_global_direct(lp->gid);
1383
                   msg->event_rc = model_net_event_mctx(net_id, &mc_src, &mc_dst,
1384 1385 1386
                           msg->category, msg->sender_svr, msg->pull_size,
                           0.0, msg->remote_event_size_bytes, tmp_ptr, 0,
                           NULL, lp);
1387 1388 1389 1390 1391 1392 1393
               }
               else{
                   e = tw_event_new(msg->final_dest_gid, ts, lp);
                   m = tw_event_data(e);
                   memcpy(m, tmp_ptr, msg->remote_event_size_bytes);
                   tw_event_send(e);
               }
1394 1395 1396 1397 1398
	    }
       }
    }
  else
    {
1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420
        bf->c6 = 1;
        int tmp_dir, tmp_dim, queue;
        int tmp_dirD, tmp_dimD, queueD;
        tw_lpid dst_lp = msg->dest_lp;
        tw_lpid dst_lpD = msg->dest_lp;
        dimension_order_routing(s, &dst_lp, &tmp_dim, &tmp_dir);
        
        queue = tmp_dir + (tmp_dim * 2);
        queueD =  tmp_dir + ( tmp_dim * 2 );
            
        tmp_dimD = tmp_dim;
        tmp_dirD = tmp_dir;
        dst_lpD = dst_lp;
        
        nodes_message_list * cur_chunk = (nodes_message_list *)malloc( 
                sizeof(nodes_message_list));
        init_nodes_message_list(cur_chunk, msg);

        if(msg->remote_event_size_bytes > 0) {
            void *m_data_src = model_net_method_get_edata(TORUS, msg);
            cur_chunk->event_data = (char*)malloc(msg->remote_event_size_bytes);
            memcpy(cur_chunk->event_data, m_data_src, 
1421
                msg->remote_event_size_bytes);
1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485
        }
        int multfactor = 1;
        if(msg->source_dim != tmp_dim) {
            multfactor = 2;
        }
        /* Message is traveling in the same dimension*/
        if(multfactor == 1) {
            cur_chunk->msg.next_stop = dst_lp;
            cur_chunk->msg.source_dim = tmp_dim;
            cur_chunk->msg.source_direction = tmp_dir;
            msg->saved_queue = queue;
            msg->saved_channel = STATICQ;
            if(s->buffer[queue][STATICQ] + multfactor * s->params->chunk_size 
                > s->params->buffer_size) {
                cur_chunk->msg.saved_queue =  
                    msg->source_direction + ( msg->source_dim * 2 );
                cur_chunk->msg.saved_channel = msg->source_channel;
                append_to_node_message_list(s, s->queued_msgs[queue], 
                        s->queued_msgs_tail[queue], STATICQ, cur_chunk);
                    s->queued_length[queue] += s->params->chunk_size;
            } else {
                bf->c9 = 1;
                s->buffer[queue][STATICQ] += s->params->chunk_size;
                credit_send( s, bf, lp, msg, -1 ); 
                append_to_node_message_list(s, s->pending_msgs[queue], 
                    s->pending_msgs_tail[queue], STATICQ, cur_chunk);
            }
        }
        else
        {
            /* Message is travelling in different dimension so two buffer
             * spaces are required. */
                if(s->buffer[queue][STATICQ] + 2 * s->params->chunk_size 
                    <= s->params->buffer_size) {
                    bf->c11 = 1;
                    cur_chunk->msg.next_stop = dst_lp;
                    cur_chunk->msg.source_dim = tmp_dim;
                    cur_chunk->msg.source_direction = tmp_dir;
                    msg->saved_queue = queue;
                    msg->saved_channel = STATICQ;
                    s->buffer[queue][STATICQ] += s->params->chunk_size;
                    credit_send( s, bf, lp, msg, -1 ); 
                    append_to_node_message_list(s, s->pending_msgs[queue], 
                        s->pending_msgs_tail[queue], STATICQ, cur_chunk);
                }
                else
                {
                    bf->c8 = 1;
                    append_to_node_message_list(s, s->other_msgs, 
                            s->other_msgs_tail, queue, cur_chunk);  
                }
        }
        if(s->in_send_loop[msg->saved_queue] == 0) {
            bf->c13 = 1;
            ts = codes_local_latency(lp);
            e = model_net_method_event_new(lp->gid, ts, lp, TORUS, (void**)&m, 
                NULL);
            m->type = SEND;
            m->source_direction = cur_chunk->msg.source_direction; 
            m->source_dim = cur_chunk->msg.source_dim;
            s->in_send_loop[msg->saved_queue] = 1;
            tw_event_send(e);
        }
    }
1486 1487
}

1488 1489
/* reports torus statistics like average packet latency, maximum packet latency and average
 * number of torus hops traversed by the packet */
1490 1491
static void torus_report_stats()
{
1492
    long long avg_hops, total_finished_packets;
1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504
    tw_stime avg_time, max_time;

    MPI_Reduce( &total_hops, &avg_hops, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
    MPI_Reduce( &N_finished_packets, &total_finished_packets, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
    MPI_Reduce( &total_time, &avg_time, 1,MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
    MPI_Reduce( &max_latency, &max_time, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);

    if(!g_tw_mynode)
     {
       printf(" Average number of hops traversed %f average message latency %lf us maximum message latency %lf us \n", (float)avg_hops/total_finished_packets, avg_time/(total_finished_packets*1000), max_time/1000);
     }
}
1505
/* finalize the torus node and free all event buffers available */
1506 1507 1508
void
final( nodes_state * s, tw_lp * lp )
{
1509
  model_net_print_stats(lp->gid, &s->torus_stats_array); 
1510 1511 1512
  free(s->next_link_available_time);
  free(s->next_credit_available_time);
  free(s->next_flit_generate_time);
1513 1514 1515 1516 1517 1518 1519 1520
  free(s->link_traffic);
  free(s->terminal_msgs);
  free(s->terminal_msgs_tail);
  free(s->pending_msgs);
  free(s->pending_msgs_tail);
  free(s->queued_msgs);
  free(s->queued_msgs_tail);
  free(s->other_msgs);
1521 1522 1523 1524 1525 1526
  // since all LPs are sharing params, just let them leak for now
  // TODO: add a post-sim "cleanup" function?
  //free(s->buffer); 
  //free(s->params->dim_length);
  //free(s->params->factor);
  //free(s->params->half_length);
1527 1528
}

1529
/* increments the buffer count after a credit arrives from the remote compute node */
1530
static void packet_buffer_process( nodes_state * ns, tw_bf * bf, nodes_message * msg, tw_lp * lp )
1531
{
1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567
    int queue = msg->source_direction + ( msg->source_dim * 2 );
    ns->buffer[queue][msg->source_channel] -= ns->params->chunk_size;
        
    if(ns->queued_msgs[queue][STATICQ] != NULL) {
            bf->c2 = 1;
            nodes_message_list *head = return_head(ns, ns->queued_msgs[queue], 
                ns->queued_msgs_tail[queue], STATICQ);
            ns->queued_length[queue] -= ns->params->chunk_size;
            credit_send( ns, bf, lp, &head->msg, 1); 
            append_to_node_message_list(ns, ns->pending_msgs[queue], 
                ns->pending_msgs_tail[queue], STATICQ, head);
            ns->buffer[queue][STATICQ] += ns->params->chunk_size;
        } else if(ns->buffer[queue][STATICQ] + 2 * ns->params->chunk_size 
            <= ns->params->buffer_size) {
            if(ns->other_msgs[queue] != NULL) {
                bf->c3 = 1;
                nodes_message_list *head = return_head(ns, ns->other_msgs, 
                        ns->other_msgs_tail, queue);
                credit_send( ns, bf, lp, &head->msg, 1); 
                append_to_node_message_list(ns, ns->pending_msgs[queue], 
                        ns->pending_msgs_tail[queue], STATICQ, head);
                ns->buffer[queue][STATIC] += ns->params->chunk_size;
            }
           } 
    if(ns->in_send_loop[queue] == 0) {
        bf->c5 = 1;
        tw_stime ts = codes_local_latency(lp);
        nodes_message *m;
        tw_event *e = model_net_method_event_new(lp->gid, ts, lp, TORUS, 
            (void**)&m, NULL);
        m->type = SEND;
        m->source_direction = msg->source_direction;
        m->source_dim = msg->source_dim;
        ns->in_send_loop[queue] = 1;
        tw_event_send(e);
    }
1568
}
1569

1570
/* reverse handler for torus node */
1571 1572 1573 1574 1575 1576 1577 1578 1579 1580
static void node_rc_handler(nodes_state * s, tw_bf * bf, nodes_message * msg, tw_lp * lp)
{
  switch(msg->type)
    {
       case GENERATE:
		   {
		     s->packet_counter--;
	 	     //saved_dim = msg->saved_src_dim;
		     //saved_dir = msg->saved_src_dir;

1581 1582 1583 1584
                     uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
                     if(msg->packet_size % s->params->chunk_size)
                         num_chunks++;

1585 1586
		     if(!num_chunks)
			num_chunks = 1;
1587
		     //s->next_flit_generate_time[(saved_dim * 2) + saved_dir][0] = msg->saved_available_time;
1588 1589 1590 1591 1592 1593
	     	     codes_local_latency_reverse(lp);
			
		     if(bf->c1)
			codes_local_latency_reverse(lp);

		     mn_stats* stat;
1594 1595 1596
		     stat = model_net_find_stats(msg->category, s->torus_stats_array);
		     stat->send_count--; 
		     stat->send_bytes -= msg->packet_size;
1597
		     stat->send_time -= (1/s->params->link_bandwidth) * msg->packet_size;
1598 1599 1600 1601 1602 1603 1604 1605 1606
		   }
	break;
	
	case ARRIVAL:
		   {
  		    tw_rand_reverse_unif(lp->rng);
		    tw_rand_reverse_unif(lp->rng);
		    int next_dim = msg->source_dim;
		    int next_dir = msg->source_direction;
1607 1608 1609
                    uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
                    if(msg->packet_size % s->params->chunk_size)
                        num_chunks++;
1610 1611 1612 1613

		    s->next_credit_available_time[next_dir + ( next_dim * 2 )][0] = msg->saved_available_time;
		    if(bf->c2)
		    {
1614 1615 1616 1617 1618
		       struct mn_stats* stat;
		       stat = model_net_find_stats(msg->category, s->torus_stats_array);
		       stat->recv_count--;
		       stat->recv_bytes -= msg->packet_size;
		        stat->recv_time -= tw_now(lp) - msg->travel_start_time;	    
1619 1620 1621 1622
		       N_finished_packets--;
		       total_time -= tw_now( lp ) - msg->travel_start_time;
		       total_hops -= msg->my_N_hop;
		    }
1623
 		    msg->my_N_hop--;
1624 1625 1626
                    if (lp->gid == msg->dest_lp && 
                            msg->chunk_id == num_chunks-1 &&
                            msg->remote_event_size_bytes && msg->is_pull){
1627
                        model_net_event_rc2(lp, &msg->event_rc);
1628
                    }
1629 1630 1631 1632 1633 1634 1635 1636