torus.c 46.6 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 21

#define CHUNK_SIZE 32
#define DEBUG 1
#define MEAN_INTERVAL 100
22
#define TRACE -1
23

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
/* Torus network model implementation of codes, implements the modelnet API */

39 40 41 42 43 44 45
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 */
    int buffer_size; /* number of buffer slots for each vc in flits*/
46
    //int num_net_traces; /* number of network traces to be mapped on torus */
47 48 49 50 51 52 53 54 55 56 57 58 59 60
    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;
};
61 62

/* codes mapping group name, lp type name */
63
static char grp_name[MAX_NAME_LENGTH];
64
/* codes mapping group id, lp type id, repetition id and offset */
65
int mapping_grp_id, mapping_rep_id, mapping_type_id, mapping_offset;
66 67 68 69

/* 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;
70
static tw_stime         max_collective = 0;
71 72 73 74 75 76

/* 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;

77 78 79 80 81
/* 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;
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105

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;
  /* 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];
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
   /* 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;
132 133 134 135 136

   /* LPs annotation */
   const char * anno;
   /* LPs configuration */
   const torus_param * params;
137 138
};

139 140 141 142
static void torus_read_config(
        const char         * anno,
        torus_param        * params){
    char dim_length_str[MAX_NAME_LENGTH];
143
    int i;
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170

    // 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);
    }

171

172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
    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)
197
    {
198 199 200 201 202 203 204 205 206
        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,",");
    }
207 208 209 210 211 212 213
    /*int num_nodes = 1;

    for( i = 0; i < p->n_dims; i++)
	   num_nodes *= p->dim_length[i];
    
    configuration_get_value_int(&config, "PARAMS", "num_net_traces", anno, &p->num_net_traces);
    if(!p->num_net_traces) {
214

215 216 217 218 219 220
	p->num_net_traces = num_nodes;
        fprintf(stderr, "Number of network traces not specified, setting to %d",
                p->num_net_traces);
    }
   // Number of network traces should be <= number of torus network nodes `
   assert(p->num_net_traces <= num_nodes);*/
221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244
    // create derived parameters
   
    // factor is an exclusive prefix product
    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++){
245
        const char * anno = anno_map->annotations[i].ptr;
246 247 248 249
        torus_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        torus_read_config(NULL, &all_params[anno_map->num_annos]);
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
/* 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;
}

280 281 282
void torus_collective_init(nodes_state * s,
           		   tw_lp * lp)
{
283 284
    // 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);
285
    int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM, s->anno, 0);
286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334
    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
}
335
/* torus packet reverse event */
336 337 338 339 340 341
static void torus_packet_event_rc(tw_lp *sender)
{
  codes_local_latency_reverse(sender);
  return;
}

342 343 344 345 346 347
/* returns the torus message size */
static int torus_get_msg_sz(void)
{
   return sizeof(nodes_message);
}

348
/* torus packet event , generates a torus packet on the compute node */
349
static tw_stime torus_packet_event(char const * category, tw_lpid final_dest_lp, tw_lpid dest_mn_lp, uint64_t packet_size, int is_pull, uint64_t pull_size, tw_stime offset, const mn_sched_params *sched_params, int remote_event_size, const void* remote_event, int self_event_size, const void* self_event, tw_lpid src_lp, tw_lp *sender, int is_last_pckt)
350 351 352 353 354 355
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    nodes_message * msg;
    char* tmp_ptr;
   
356
    xfer_to_nic_time = g_tw_lookahead + codes_local_latency(sender); /* Throws an error of found last KP time > current event time otherwise */
357 358 359 360
    //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);
361 362
    strcpy(msg->category, category);
    msg->final_dest_gid = final_dest_lp;
363
    msg->dest_lp = dest_mn_lp;
364
    msg->sender_svr= src_lp;
365
    msg->sender_node = sender->gid;
366 367 368 369
    msg->packet_size = packet_size;
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = GENERATE;
370 371
    msg->is_pull = is_pull;
    msg->pull_size = pull_size;
372
    
373 374 375 376 377 378 379 380 381 382 383 384 385 386
    if(is_last_pckt) /* Its the last packet so pass in remote event information*/
     {
	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;
         }
	if(self_event_size > 0)
	{
	   msg->local_event_size_bytes = self_event_size;
	   memcpy(tmp_ptr, self_event, self_event_size);
	   tmp_ptr += self_event_size;
	}
387
      // 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);
388 389
     }
    tw_event_send(e_new);
390
    return xfer_to_nic_time;
391
}
392

393 394 395 396 397
/*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;
398
    char anno[MAX_NAME_LENGTH];
399

400 401 402 403 404 405
    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{
406
        s->anno = strdup(anno);
407 408 409
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }
410

411 412
    // shorthand
    const torus_param *p = s->params;
413

414 415 416 417 418 419 420 421 422 423 424 425
    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*));

    for(i=0; i < 2*p->n_dims; i++)
426
    {
427 428 429 430 431 432 433
	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));
434 435
    }

436 437 438
    // 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
439 440 441 442 443 444
    /* DEBUG
    printf("%lu: my coords:", lp->gid);
    for (i = 0; i < p->n_dims; i++)
        printf(" %d", s->dim_position[i]);
    printf("\n");
    */
445

446 447
  int temp_dim_pos[ p->n_dims ];
  for ( i = 0; i < p->n_dims; i++ )
448 449 450
    temp_dim_pos[ i ] = s->dim_position[ i ];

  // calculate minus neighbour's lpID
451
  for ( j = 0; j < p->n_dims; j++ )
452
    {
453 454
      temp_dim_pos[ j ] = (s->dim_position[ j ] -1 + p->dim_length[ j ]) %
          p->dim_length[ j ];
455

456 457 458 459 460 461 462 463 464 465
      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));
      */

466 467 468
      temp_dim_pos[ j ] = s->dim_position[ j ];
    }
  // calculate plus neighbour's lpID
469
  for ( j = 0; j < p->n_dims; j++ )
470
    {
471 472
      temp_dim_pos[ j ] = ( s->dim_position[ j ] + 1 + p->dim_length[ j ]) %
          p->dim_length[ j ];
473

474 475 476 477 478 479 480 481 482
      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));
      */
483 484 485 486 487

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

  //printf("\n");
488
  for( j=0; j < 2 * p->n_dims; j++ )
489
   {
490
    for( i = 0; i < p->num_vc; i++ )
491 492 493 494 495 496 497 498
     {
       s->buffer[ j ][ i ] = 0; 
       s->next_link_available_time[ j ][ i ] = 0.0;
       s->next_credit_available_time[j][i] = 0.0; 
     }
   }
  // record LP time
    s->packet_counter = 0;
499 500 501 502 503
    torus_collective_init(s, lp);
}


/* collective operation for the torus network */
504
void torus_collective(char const * category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
505 506 507 508 509 510 511
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    nodes_message * msg;
    tw_lpid local_nic_id;
    char* tmp_ptr;

512 513 514 515 516
    // 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);
517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557

    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;
558
            ts = (1/s->params->link_bandwidth) * msg->remote_event_size_bytes;
559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586
            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 */
587 588 589 590 591 592 593 594
            // 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);
595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624

           /* 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++;

625 626 627 628 629
        // 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);
630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647

        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 */
648 649 650
            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);
651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686

           /* 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 */
687 688 689
                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);
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
                //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);
           }
      }
710
}
711 712 713 714 715 716 717
	     
static void node_collective_fan_out(nodes_state * s,
                        tw_bf * bf,
                        nodes_message * msg,
                        tw_lp * lp)
{
        int i;
718 719 720
        //TODO: be annotation-aware
        int num_lps = codes_mapping_get_lp_count(grp_name, 1, LP_CONFIG_NM,
                NULL, 1);
721 722 723 724
        bf->c1 = 0;
        bf->c2 = 0;

        send_remote_event(s, bf, msg, lp);
725

726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741
        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 */
742 743 744
                        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);
745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772
                        //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;
          }
}
    
773 774 775 776 777 778
/*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 )
{
779 780
     int dest[s->params->n_dims];
     int dest_id;
781

782 783 784 785
  /* dummys - check later */
  *dim = -1;
  *dir = -1;

786 787
  to_dim_id(codes_mapping_get_lp_relative_id(*dst_lp, 0, 1),
          s->params->n_dims, s->params->dim_length, dest);
788

789
  for(int i = 0; i < s->params->n_dims; i++ )
790
    {
791
      if ( s->dim_position[ i ] - dest[ i ] > s->params->half_length[ i ] )
792 793 794 795 796 797
	{
	  dest_id = s->neighbour_plus_lpID[ i ];
	  *dim = i;
	  *dir = 1;
	  break;
	}
798
      if ( s->dim_position[ i ] - dest[ i ] < -s->params->half_length[ i ] )
799 800 801 802 803 804
	{
	  dest_id = s->neighbour_minus_lpID[ i ];
	  *dim = i;
	  *dir = 0;
	  break;
	}
805 806
      if ( ( s->dim_position[i] - dest[i] <= s->params->half_length[i] ) &&
              ( s->dim_position[ i ] - dest[ i ] > 0 ) )
807 808 809 810 811 812
	{
	  dest_id = s->neighbour_minus_lpID[ i ];
	  *dim = i;
	  *dir = 0;
	  break;
	}
813 814
      if (( s->dim_position[i] - dest[i] >= -s->params->half_length[i] ) &&
              ( s->dim_position[ i ] - dest[ i ] < 0) )
815 816 817 818 819 820 821
	{
	  dest_id = s->neighbour_plus_lpID[ i ];
	  *dim = i;
	  *dir = 1;
	  break;
	}
    }
822 823

  assert(*dim != -1 && *dir != -1);
824 825
  *dst_lp = codes_mapping_get_lpid_from_relative(dest_id, NULL, LP_CONFIG_NM,
          s->anno, 1);
826
}
827 828 829 830 831 832

/*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. */
833 834 835 836 837 838
static void packet_generate( nodes_state * s, 
		tw_bf * bf, 
		nodes_message * msg, 
		tw_lp * lp )
{
//    printf("\n msg local event size %d remote event size %d ", msg->local_event_size_bytes, msg->remote_event_size_bytes);
839
    int j, tmp_dir=-1, tmp_dim=-1, total_event_size;
840 841 842 843 844 845
    tw_stime ts;

//    event triggered when packet head is sent
    tw_event * e_h;
    nodes_message *m;

846
    tw_lpid dst_lp = msg->dest_lp;
847

848 849 850 851 852 853 854 855
    dimension_order_routing( s, &dst_lp, &tmp_dim, &tmp_dir );

    msg->saved_src_dim = tmp_dim;
    msg->saved_src_dir = tmp_dir;

    //msg->saved_available_time = s->next_flit_generate_time[(2*tmp_dim) + tmp_dir][0];
    msg->travel_start_time = tw_now(lp);
    msg->packet_ID = lp->gid + g_tw_nlp * s->packet_counter;
856
    msg->my_N_hop = 0;
857

858 859 860 861
    uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
    if(msg->packet_size % s->params->chunk_size)
        num_chunks++;

862

863 864 865 866 867 868
    s->packet_counter++;

    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);
    for(j = 0; j < num_chunks; j++)
    { 
869
     if(s->buffer[ tmp_dir + ( tmp_dim * 2 ) ][ 0 ] < s->params->buffer_size)
870 871 872 873 874
      {
       ts = j + tw_rand_exponential(lp->rng, MEAN_INTERVAL/200);
       //s->next_flit_generate_time[(2*tmp_dim) + tmp_dir][0] = max(s->next_flit_generate_time[(2*tmp_dim) + tmp_dir][0], tw_now(lp));
       //s->next_flit_generate_time[(2*tmp_dim) + tmp_dir][0] += ts;
       //e_h = tw_event_new( lp->gid, s->next_flit_generate_time[(2*tmp_dim) + tmp_dir][0] - tw_now(lp), lp);
875
       //e_h = tw_event_new(lp->gid, ts, lp);
876 877 878
       msg->source_direction = tmp_dir;
       msg->source_dim = tmp_dim;

879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895
       void *m_data;
       e_h = model_net_method_event_new(lp->gid, ts, lp, TORUS, (void**)&m,
               (void**)&m_data);

       //m = tw_event_data( e_h );
       //memcpy(m, msg, torus_get_msg_sz() + msg->local_event_size_bytes + msg->remote_event_size_bytes);
       void *m_data_src = model_net_method_get_edata(TORUS, msg);
       memcpy(m, msg, sizeof(nodes_message));
       if (msg->remote_event_size_bytes){
           memcpy(m_data, m_data_src,
                   msg->remote_event_size_bytes);
           m_data = (char*)m_data + msg->remote_event_size_bytes;
           m_data_src = (char*)m_data_src + msg->remote_event_size_bytes;
       }
       if (msg->local_event_size_bytes){
           memcpy(m_data, m_data_src, msg->local_event_size_bytes);
       }
896 897 898 899 900 901 902 903 904 905 906 907
       m->next_stop = dst_lp;
       m->chunk_id = j;

      // find destination dimensions using destination LP ID 
       m->type = SEND;
       m->source_direction = tmp_dir;
       m->source_dim = tmp_dim;
       tw_event_send(e_h);
      }
      else 
       {
   printf("\n %d Packet queued in line increase buffer space, dir %d dim %d buffer space %d dest LP %d ", (int)lp->gid, tmp_dir, tmp_dim, s->buffer[ tmp_dir + ( tmp_dim * 2 ) ][ 0 ], (int)msg->dest_lp);
908 909
       MPI_Finalize();
       exit(-1); 
910 911
       }
   }
912

913
   total_event_size = model_net_get_msg_sz(TORUS) + msg->remote_event_size_bytes + msg->local_event_size_bytes;   
914 915 916 917 918
   /* record the statistics of the generated packets */
   mn_stats* stat;
   stat = model_net_find_stats(msg->category, s->torus_stats_array);
   stat->send_count++;  
   stat->send_bytes += msg->packet_size;
919
   stat->send_time += (1/s->params->link_bandwidth) * msg->packet_size;
920 921 922
   /* record the maximum ROSS event size */
   if(stat->max_event_size < total_event_size)
	   stat->max_event_size = total_event_size;
923 924 925 926 927 928 929 930
}
/*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)
{
#if DEBUG
931
    //printf("\n (%lf) sending credit tmp_dir %d tmp_dim %d %lf ", tw_now(lp), msg->source_direction, msg->source_dim, s->params->credit_delay );
932 933 934 935 936 937 938 939 940
#endif
    bf->c1 = 0;
    tw_event * buf_e;
    nodes_message *m;
    tw_stime ts;
    int src_dir = msg->source_direction;
    int src_dim = msg->source_dim;

    msg->saved_available_time = s->next_credit_available_time[(2 * src_dim) + src_dir][0];
941
    s->next_credit_available_time[(2 * src_dim) + src_dir][0] = maxd(s->next_credit_available_time[(2 * src_dim) + src_dir][0], tw_now(lp));
942 943
    ts =  s->params->credit_delay + 
        tw_rand_exponential(lp->rng, s->params->credit_delay/1000);
944 945
    s->next_credit_available_time[(2 * src_dim) + src_dir][0] += ts;

946 947
    //buf_e = tw_event_new( msg->sender_lp, s->next_credit_available_time[(2 * src_dim) + src_dir][0] - tw_now(lp), lp);
    //m = tw_event_data(buf_e);
948
    buf_e = model_net_method_event_new(msg->sender_node,
949 950
            s->next_credit_available_time[(2*src_dim) + src_dir][0] - tw_now(lp),
            lp, TORUS, (void**)&m, NULL);
951 952 953 954 955 956
    m->source_direction = msg->source_direction;
    m->source_dim = msg->source_dim;

    m->type = CREDIT;
    tw_event_send( buf_e );
}
957 958
/* 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 */
959 960 961 962 963 964 965 966 967 968 969 970 971 972
static void packet_send( nodes_state * s, 
	         tw_bf * bf, 
		 nodes_message * msg, 
		 tw_lp * lp )
{ 
    bf->c2 = 0;
    bf->c1 = 0;
    int tmp_dir, tmp_dim;
    tw_stime ts;
    tw_event *e;
    nodes_message *m;
    tw_lpid dst_lp = msg->dest_lp;
    dimension_order_routing( s, &dst_lp, &tmp_dim, &tmp_dir );     

973
    if(s->buffer[ tmp_dir + ( tmp_dim * 2 ) ][ 0 ] < s->params->buffer_size)
974 975 976 977
    {
       bf->c2 = 1;
       msg->saved_src_dir = tmp_dir;
       msg->saved_src_dim = tmp_dim;
978 979
       ts = tw_rand_exponential( lp->rng, s->params->head_delay/200.0 ) + 
           s->params->head_delay;
980 981 982 983

//    For reverse computation 
      msg->saved_available_time = s->next_link_available_time[tmp_dir + ( tmp_dim * 2 )][0];

984
      s->next_link_available_time[tmp_dir + ( tmp_dim * 2 )][0] = maxd( s->next_link_available_time[ tmp_dir + ( tmp_dim * 2 )][0], tw_now(lp) );
985 986
      s->next_link_available_time[tmp_dir + ( tmp_dim * 2 )][0] += ts;
    
987 988 989 990 991 992 993 994 995 996 997 998
      //e = tw_event_new( dst_lp, s->next_link_available_time[tmp_dir + ( tmp_dim * 2 )][0] - tw_now(lp), lp );
      //m = tw_event_data( e );
      //memcpy(m, msg, torus_get_msg_sz() + msg->remote_event_size_bytes);
      void * m_data;
      e = model_net_method_event_new(dst_lp, 
              s->next_link_available_time[tmp_dir+(tmp_dim*2)][0] - tw_now(lp),
              lp, TORUS, (void**)&m, &m_data);
      memcpy(m, 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);
      }
999 1000 1001 1002 1003 1004 1005 1006
      m->type = ARRIVAL;

      if(msg->packet_ID == TRACE)
        printf("\n lp %d packet %lld flit id %d being sent to %d after time %lf ", (int) lp->gid, msg->packet_ID, msg->chunk_id, (int)dst_lp, s->next_link_available_time[tmp_dir + ( tmp_dim * 2 )][0] - tw_now(lp)); 
      //Carry on the message info
      m->source_dim = tmp_dim;
      m->source_direction = tmp_dir;
      m->next_stop = dst_lp;
1007
      m->sender_node = lp->gid;
1008 1009 1010 1011 1012 1013
      m->local_event_size_bytes = 0; /* We just deliver the local event here */

      tw_event_send( e );

      s->buffer[ tmp_dir + ( tmp_dim * 2 ) ][ 0 ]++;
    
1014 1015 1016 1017 1018
      uint64_t num_chunks = msg->packet_size/s->params->chunk_size;

      if(msg->packet_size % s->params->chunk_size)
          num_chunks++;

1019 1020 1021
      if(msg->chunk_id == num_chunks - 1)
      {
        bf->c1 = 1;
1022
	/* Invoke an event on the sending server */
1023 1024 1025 1026
	if(msg->local_event_size_bytes > 0)
	{
          tw_event* e_new;
	  nodes_message* m_new;
1027
	  void* local_event;
1028
	  ts = (1/s->params->link_bandwidth) * msg->local_event_size_bytes;
1029
	  e_new = tw_event_new(msg->sender_svr, ts, lp);
1030
	  m_new = tw_event_data(e_new);
1031 1032 1033 1034
	  //local_event = (char*)msg;
	  //local_event += torus_get_msg_sz() + msg->remote_event_size_bytes;
          local_event = (char*)model_net_method_get_edata(TORUS, msg) +
              msg->remote_event_size_bytes;
1035 1036 1037 1038 1039 1040
	  memcpy(m_new, local_event, msg->local_event_size_bytes);
	  tw_event_send(e_new);
	}
     }
  } // end if
    else
1041
    {
1042
	    printf("\n buffer overflown ");
1043 1044 1045
	    MPI_Finalize();
	    exit(-1);
    }
1046 1047
}

1048 1049 1050
/*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. */
1051 1052 1053 1054 1055 1056 1057 1058 1059
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;
1060
  mn_stats* stat;
1061

1062
  credit_send( s, bf, lp, msg); 
1063 1064 1065 1066 1067 1068 1069
  
  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);
  if( lp->gid == msg->dest_lp )
    {   
1070 1071 1072 1073
        uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
        if(msg->packet_size % s->params->chunk_size)
            num_chunks++;

1074 1075 1076
        if( msg->chunk_id == num_chunks - 1 )    
        {
	    bf->c2 = 1;
1077 1078 1079 1080 1081 1082
	    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*/
1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094
	    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)
	    {
1095
               void *tmp_ptr = model_net_method_get_edata(TORUS, msg);
1096
               if (msg->is_pull){
1097
                   int net_id = model_net_get_id(LP_METHOD_NM);
1098 1099 1100 1101 1102 1103 1104 1105
                   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);
                   model_net_event_mctx(net_id, &mc_src, &mc_dst,
                           msg->category, msg->sender_svr, msg->pull_size,
                           0.0, msg->remote_event_size_bytes, tmp_ptr, 0,
                           NULL, lp);
1106 1107 1108 1109 1110 1111 1112
               }
               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);
               }
1113 1114 1115 1116 1117
	    }
       }
    }
  else
    {
1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128
      //e = tw_event_new(lp->gid, ts , lp);
      //m = tw_event_data( e );
      //memcpy(m, msg, torus_get_msg_sz() + msg->remote_event_size_bytes);
      void *m_data;
      e = model_net_method_event_new(lp->gid, ts, lp, TORUS, (void**)&m,
              &m_data);
      memcpy(m, 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);
      }
1129 1130 1131 1132 1133 1134
      m->type = SEND;
      m->next_stop = -1;
      tw_event_send(e);
   }
}

1135 1136
/* reports torus statistics like average packet latency, maximum packet latency and average
 * number of torus hops traversed by the packet */
1137 1138
static void torus_report_stats()
{
1139
    long long avg_hops, total_finished_packets;
1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151
    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);
     }
}
1152
/* finalize the torus node and free all event buffers available */
1153 1154 1155
void
final( nodes_state * s, tw_lp * lp )
{
1156
  model_net_print_stats(lp->gid, &s->torus_stats_array[0]); 
1157 1158 1159
  free(s->next_link_available_time);
  free(s->next_credit_available_time);
  free(s->next_flit_generate_time);
1160 1161 1162 1163 1164 1165
  // 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);
1166 1167
}

1168
/* increments the buffer count after a credit arrives from the remote compute node */
1169 1170 1171 1172
static void packet_buffer_process( nodes_state * s, tw_bf * bf, nodes_message * msg, tw_lp * lp )
{
   s->buffer[ msg->source_direction + ( msg->source_dim * 2 ) ][  0 ]--;
}
1173

1174
/* reverse handler for torus node */
1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185
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--;
		     int i;//, saved_dim, saved_dir;
	 	     //saved_dim = msg->saved_src_dim;
		     //saved_dir = msg->saved_src_dir;

1186 1187 1188 1189
                     uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
                     if(msg->packet_size % s->params->chunk_size)
                         num_chunks++;

1190 1191
		     //s->next_flit_generate_time[(saved_dim * 2) + saved_dir][0] = msg->saved_available_time;
		     for(i=0; i < num_chunks; i++)
1192 1193 1194 1195 1196
  		        tw_rand_reverse_unif(lp->rng);
	     	     mn_stats* stat;
		     stat = model_net_find_stats(msg->category, s->torus_stats_array);
		     stat->send_count--; 
		     stat->send_bytes -= msg->packet_size;
1197
		     stat->send_time -= (1/s->params->link_bandwidth) * msg->packet_size;
1198 1199 1200 1201 1202 1203 1204 1205 1206
		   }
	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;
1207 1208 1209
                    uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
                    if(msg->packet_size % s->params->chunk_size)
                        num_chunks++;
1210 1211 1212 1213

		    s->next_credit_available_time[next_dir + ( next_dim * 2 )][0] = msg->saved_available_time;
		    if(bf->c2)
		    {
1214 1215 1216 1217 1218
		       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;	    
1219 1220 1221 1222
		       N_finished_packets--;
		       total_time -= tw_now( lp ) - msg->travel_start_time;
		       total_hops -= msg->my_N_hop;
		    }
1223
 		    msg->my_N_hop--;
1224 1225 1226
                    if (lp->gid == msg->dest_lp && 
                            msg->chunk_id == num_chunks-1 &&
                            msg->remote_event_size_bytes && msg->is_pull){