dragonfly.c 61.7 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 9 10
// Local router ID: 0 --- total_router-1
// Router LP ID 
// Terminal LP ID

11 12 13 14 15 16
#include <ross.h>

#include "codes/codes_mapping.h"
#include "codes/codes.h"
#include "codes/model-net.h"
#include "codes/model-net-method.h"
17 18
#include "codes/model-net-lp.h"
#include "codes/net/dragonfly.h"
19 20 21 22

#define CREDIT_SIZE 8
#define MEAN_PROCESS 1.0

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

31 32 33 34 35
// debugging parameters
#define TRACK 235221
#define PRINT_ROUTER_TABLE 1
#define DEBUG 1

36 37 38
#define LP_CONFIG_NM (model_net_lp_config_names[DRAGONFLY])
#define LP_METHOD_NM (model_net_method_names[DRAGONFLY])

39 40
static double maxd(double a, double b) { return a < b ? b : a; }

41 42
// arrival rate
static double MEAN_INTERVAL=200.0;
43 44
// threshold for adaptive routing
static int adaptive_threshold = 10;
45 46

/* minimal and non-minimal packet counts for adaptive routing*/
47
int minimal_count=0, nonmin_count=0;
48

49 50 51 52 53 54
typedef struct dragonfly_param dragonfly_param;
/* annotation-specific parameters (unannotated entry occurs at the 
 * last index) */
static uint64_t                  num_params = 0;
static dragonfly_param         * all_params = NULL;
static const config_anno_map_t * anno_map   = NULL;
55 56

/* global variables for codes mapping */
57
static char lp_group_name[MAX_NAME_LENGTH];
58 59
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
struct dragonfly_param
{
    // configuration parameters
    int num_routers; /*Number of routers in a group*/
    double local_bandwidth;/* bandwidth of the router-router channels within a group */
    double global_bandwidth;/* bandwidth of the inter-group router connections */
    double cn_bandwidth;/* bandwidth of the compute node channels connected to routers */
    int num_vcs; /* number of virtual channels */
    int local_vc_size; /* buffer size of the router-router channels */
    int global_vc_size; /* buffer size of the global channels */
    int cn_vc_size; /* buffer size of the compute node channels */
    int routing; /* minimal or non-minimal routing */
    int chunk_size; /* full-sized packets are broken into smaller chunks.*/

    // derived parameters
    int num_cn;
    int num_groups;
    int radix;
    int total_routers;
    int num_global_channels;
};

82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105
/* handles terminal and router events like packet generate/send/receive/buffer */
typedef enum event_t event_t;

typedef struct terminal_state terminal_state;
typedef struct router_state router_state;

/* dragonfly compute node data structure */
struct terminal_state
{
   unsigned long long packet_counter;

   // Dragonfly specific parameters
   unsigned int router_id;
   unsigned int terminal_id;

   // Each terminal will have an input and output channel with the router
   int* vc_occupancy; // NUM_VC
   int* output_vc_state;
   tw_stime terminal_available_time;
   tw_stime next_credit_available_time;
// Terminal generate, sends and arrival T_SEND, T_ARRIVAL, T_GENERATE
// Router-Router Intra-group sends and receives RR_LSEND, RR_LARRIVE
// Router-Router Inter-group sends and receives RR_GSEND, RR_GARRIVE
   struct mn_stats dragonfly_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
  /* 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;
130 131 132

   const char * anno;
   const dragonfly_param *params;
133
};
134

135 136 137 138 139 140 141 142 143
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
  T_SEND,
  T_BUFFER,
  R_SEND,
  R_ARRIVE,
144 145 146 147
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
};
/* status of a virtual channel can be idle, active, allocated or wait for credit */
enum vc_status
{
   VC_IDLE,
   VC_ACTIVE,
   VC_ALLOC,
   VC_CREDIT
};

/* whether the last hop of a packet was global, local or a terminal */
enum last_hop
{
   GLOBAL,
   LOCAL,
   TERMINAL
};

/* three forms of routing algorithms available, adaptive routing is not
 * accurate and fully functional in the current version as the formulas
 * for detecting load on global channels are not very accurate */
enum ROUTING_ALGO
{
171 172 173
    MINIMAL = 0,
    NON_MINIMAL,
    ADAPTIVE
174 175 176 177 178 179 180 181 182 183 184 185
};

struct router_state
{
   unsigned int router_id;
   unsigned int group_id;
  
   int* global_channel; 
   tw_stime* next_output_available_time;
   tw_stime* next_credit_available_time;
   int* vc_occupancy;
   int* output_vc_state;
186 187 188

   const char * anno;
   const dragonfly_param *params;
189 190 191 192 193 194
};

static short routing = MINIMAL;

static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
195
static tw_stime         max_collective = 0;
196 197 198 199 200 201 202 203 204 205 206 207 208 209


static long long       total_hops = 0;
static long long       N_finished_packets = 0;

/* returns the dragonfly router lp type for lp registration */
static const tw_lptype* dragonfly_get_router_lp_type(void);

/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
{
	   return sizeof(terminal_message);
}

210 211 212
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
213

214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
    configuration_get_value_int(&config, "PARAMS", "num_routers", anno,
            &p->num_routers);
    if(p->num_routers <= 0) {
        p->num_routers = 4;
        fprintf(stderr, "Number of dimensions not specified, setting to %d\n",
                p->num_routers);
    }

    configuration_get_value_int(&config, "PARAMS", "num_vcs", anno,
            &p->num_vcs);
    if(p->num_vcs <= 0) {
        p->num_vcs = 1;
        fprintf(stderr, "Number of virtual channels not specified, setting to %d\n", p->num_vcs);
    }

    configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
    if(!p->local_vc_size) {
        p->local_vc_size = 1024;
        fprintf(stderr, "Buffer size of local channels not specified, setting to %d\n", p->local_vc_size);
    }

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

    configuration_get_value_int(&config, "PARAMS", "cn_vc_size", anno, &p->cn_vc_size);
    if(!p->cn_vc_size) {
        p->cn_vc_size = 1024;
        fprintf(stderr, "Buffer size of compute node channels not specified, setting to %d\n", p->cn_vc_size);
    }

    configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(!p->chunk_size) {
        p->chunk_size = 64;
        fprintf(stderr, "Chunk size for packets is specified, setting to %d\n", p->chunk_size);
    }

    configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
    if(!p->local_bandwidth) {
        p->local_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of local channels not specified, setting to %lf\n", p->local_bandwidth);
    }

    configuration_get_value_double(&config, "PARAMS", "global_bandwidth", anno, &p->global_bandwidth);
    if(!p->global_bandwidth) {
        p->global_bandwidth = 4.7;
        fprintf(stderr, "Bandwidth of global channels not specified, setting to %lf\n", p->global_bandwidth);
    }

    configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", anno, &p->cn_bandwidth);
    if(!p->cn_bandwidth) {
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }


    char routing[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing,
            MAX_NAME_LENGTH);
    if(strcmp(routing, "minimal") == 0)
        p->routing = 0;
    else if(strcmp(routing, "nonminimal")==0 || strcmp(routing,"non-minimal")==0)
        p->routing = 1;
    else if (strcmp(routing, "adaptive") == 0)
        p->routing = 2;
    else
    {
        fprintf(stderr, 
                "No routing protocol specified, setting to minimal routing\n");
        p->routing = 0;
    }

    // set the derived parameters
    p->num_cn = p->num_routers/2;
    p->num_global_channels = p->num_routers/2;
    p->num_groups = p->num_routers * p->num_cn + 1;
    p->radix = p->num_vcs *
        (p->num_cn + p->num_global_channels + p->num_routers);
    p->total_routers = p->num_groups * p->num_routers;
}

static void dragonfly_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++){
        const char * anno = anno_map->annotations[i];
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
310 311 312 313 314 315 316 317
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
/* TODO: Add dragonfly packet average, maximum latency and average number of hops traversed */
   long long avg_hops, total_finished_packets;
   tw_stime avg_time, max_time;
318
   int total_minimal_packets, total_nonmin_packets;
319 320 321 322 323

   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( &dragonfly_total_time, &avg_time, 1,MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
   MPI_Reduce( &dragonfly_max_latency, &max_time, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
324 325 326 327 328
   if(routing == ADAPTIVE)
    {
	MPI_Reduce(&minimal_count, &total_minimal_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
 	MPI_Reduce(&nonmin_count, &total_nonmin_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
    }
329 330 331 332 333

   /* print statistics */
   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);
334 335 336 337
     if(routing == ADAPTIVE)
              printf("\n ADAPTIVE ROUTING STATS: %d packets routed minimally %d packets routed non-minimally ", total_minimal_packets, total_nonmin_packets);
 
  }
338 339
   return;
}
340

341 342 343
void dragonfly_collective_init(terminal_state * s,
           		   tw_lp * lp)
{
344 345 346 347 348
    // TODO: be annotation-aware
    codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id, NULL,
            &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
    int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
            NULL, 1);
349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398
    int num_reps = codes_mapping_get_group_reps(lp_group_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 DRAGONFLY_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
}

399
/* dragonfly packet event , generates a dragonfly packet on the compute node */
400
static tw_stime dragonfly_packet_event(char* category, tw_lpid final_dest_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)
401 402 403 404 405 406
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

407
    xfer_to_nic_time = codes_local_latency(sender); /* Throws an error of found last KP time > current event time otherwise when LPs of one type are placed together*/
408 409
    //printf("\n transfer in time %f %f ", xfer_to_nic_time+offset, tw_now(sender));
    //e_new = tw_event_new(sender->gid, xfer_to_nic_time+offset, sender);
410 411 412
    //msg = tw_event_data(e_new);
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
413 414
    strcpy(msg->category, category);
    msg->final_dest_gid = final_dest_lp;
415
    msg->sender_lp=src_lp;
416 417 418 419
    msg->packet_size = packet_size;
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
420 421
    msg->is_pull = is_pull;
    msg->pull_size = pull_size;
422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437

    if(is_last_pckt) /* Its the last packet so pass in remote and local 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;
	}
     }
438
	   //printf("\n dragonfly 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);
439
    tw_event_send(e_new);
440
    return xfer_to_nic_time;
441 442 443 444 445 446 447 448 449 450 451 452 453 454
}

/* dragonfly packet event reverse handler */
static void dragonfly_packet_event_rc(tw_lp *sender)
{
	  codes_local_latency_reverse(sender);
	    return;
}

/* given a group ID gid, find the router in the current group that is attached
 * to a router in the group gid */
tw_lpid getRouterFromGroupID(int gid, 
		    router_state * r)
{
455 456 457 458
    const dragonfly_param *p = r->params;
  int group_begin = r->group_id * p->num_routers;
  int group_end = (r->group_id * p->num_routers) + p->num_routers-1;
  int offset = (gid * p->num_routers - group_begin) / p->num_routers;
459
  
460 461
  if((gid * p->num_routers) < group_begin)
    offset = (group_begin - gid * p->num_routers) / p->num_routers; // take absolute value
462
  
463 464
  int half_channel = p->num_global_channels / 2;
  int index = (offset - 1)/(half_channel * p->num_routers);
465
  
466
  offset=(offset - 1) % (half_channel * p->num_routers);
467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486

  // If the destination router is in the same group
  tw_lpid router_id;

  if(index % 2 != 0)
    router_id = group_end - (offset / half_channel); // start from the end
  else
    router_id = group_begin + (offset / half_channel);

  return router_id;
}	

/*When a packet is sent from the current router and a buffer slot becomes available, a credit is sent back to schedule another packet event*/
void router_credit_send(router_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
{
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

  int dest=0, credit_delay=0, type = R_BUFFER;
487
  int is_terminal = 0;
488

489
  const dragonfly_param *p = s->params;
490
  int sender_radix;
491 492 493 494
 // Notify sender terminal about available buffer space
  if(msg->last_hop == TERMINAL)
  {
   dest = msg->src_terminal_id;
495
   sender_radix = msg->local_id % p->num_cn;  
496
   //determine the time in ns to transfer the credit
497
   credit_delay = (1 / p->cn_bandwidth) * CREDIT_SIZE;
498
   type = T_BUFFER;
499
   is_terminal = 1;
500 501 502 503
  }
   else if(msg->last_hop == GLOBAL)
   {
     dest = msg->intm_lp_id;
504
     sender_radix = p->num_cn + (msg->local_id % p->num_routers);
505
     credit_delay = (1 / p->global_bandwidth) * CREDIT_SIZE;
506 507 508 509
   }
    else if(msg->last_hop == LOCAL)
     {
        dest = msg->intm_lp_id;
510
        sender_radix = p->num_cn + p->num_routers + (msg->local_id % p->num_routers);
511
     	credit_delay = (1/p->local_bandwidth) * CREDIT_SIZE;
512 513 514 515 516
     }
    else
      printf("\n Invalid message type");

   // Assume it takes 0.1 ns of serialization latency for processing the credits in the queue
517
    int output_port = msg->saved_vc / p->num_vcs;
518 519 520

    msg->saved_available_time = s->next_credit_available_time[sender_radix];
    s->next_credit_available_time[sender_radix] = maxd(tw_now(lp), s->next_credit_available_time[output_port]);
521
    ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, (double)credit_delay/1000);
522
	
523
    s->next_credit_available_time[sender_radix]+=ts;
524 525
    if (is_terminal){
        buf_e = model_net_method_event_new(dest, 
526
                s->next_credit_available_time[sender_radix] - tw_now(lp), lp,
527 528 529
                DRAGONFLY, (void**)&buf_msg, NULL);
    }
    else{
530
        buf_e = tw_event_new(dest, s->next_credit_available_time[sender_radix] - tw_now(lp) , lp);
531 532
        buf_msg = tw_event_data(buf_e);
    }
533 534 535 536 537 538 539 540 541 542 543 544 545
    buf_msg->vc_index = msg->saved_vc;
    buf_msg->type=type;
    buf_msg->last_hop = msg->last_hop;
    buf_msg->packet_ID=msg->packet_ID;

    tw_event_send(buf_e);

    return;
}

/* generates packet at the current dragonfly compute node */
void packet_generate(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
{
546
    tw_lpid dest_terminal_id;
547 548
    dest_terminal_id = model_net_find_local_device(DRAGONFLY, s->anno, 0,
            msg->final_dest_gid);
549 550 551 552
    msg->dest_terminal_id = dest_terminal_id;

    const dragonfly_param *p = s->params;

553 554 555
  tw_stime ts;
  tw_event *e;
  terminal_message *m;
556
  int i, total_event_size;
557 558 559
  uint64_t num_chunks = msg->packet_size / p->chunk_size;
  if (msg->packet_size % s->params->chunk_size)
      num_chunks++;
560 561
  msg->packet_ID = lp->gid + g_tw_nlp * s->packet_counter + tw_rand_integer(lp->rng, 0, lp->gid + g_tw_nlp * s->packet_counter);
  msg->travel_start_time = tw_now(lp);
562
  msg->my_N_hop = 0;
563 564
  for(i = 0; i < num_chunks; i++)
  {
565 566
	  // Before
	  // msg->my_N_hop = 0; generating a packet, check if the input queue is available
567
        ts = g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, MEAN_INTERVAL/200);
568
	int chan = -1, j;
569
	for(j = 0; j < p->num_vcs; j++)
570
	 {
571
	     if(s->vc_occupancy[j] < p->cn_vc_size * num_chunks)
572 573 574 575 576 577
	      {
	       chan=j;
	       break;
	      }
         }

578 579 580 581 582 583 584 585
        // this is a terminal event, so use the method-event version
       //e = tw_event_new(lp->gid, i + ts, lp);
       //m = tw_event_data(e);
       //memcpy(m, msg, sizeof(terminal_message) + msg->remote_event_size_bytes + msg->local_event_size_bytes);
       void * m_data;
       e = model_net_method_event_new(lp->gid, i+ts, lp, DRAGONFLY,
               (void**)&m, &m_data);
       memcpy(m, msg, sizeof(terminal_message));
586
       m->dest_terminal_id = dest_terminal_id;
587 588 589 590 591 592 593 594 595
       void * m_data_src = model_net_method_get_edata(DRAGONFLY, msg);
       if (msg->remote_event_size_bytes){
            memcpy(m_data, m_data_src, msg->remote_event_size_bytes);
       }
       if (msg->local_event_size_bytes){ 
            memcpy((char*)m_data + msg->remote_event_size_bytes,
                    (char*)m_data_src + msg->remote_event_size_bytes,
                    msg->local_event_size_bytes);
       }
596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616
       m->intm_group_id = -1;
       m->saved_vc=0;
       m->chunk_id = i;
       
       if(msg->packet_ID == TRACK && msg->chunk_id == num_chunks-1)
         printf("\n packet generated %lld at terminal %d chunk id %d ", msg->packet_ID, (int)lp->gid, i);
       
       m->output_chan = -1;
       if(chan != -1) // If the input queue is available
   	{
	    // Send the packet out
	     m->type = T_SEND;
 	     tw_event_send(e);
        }
      else
         {
	  printf("\n Exceeded queue size, exitting %d", s->vc_occupancy[0]);
	  MPI_Finalize();
	  exit(-1);
        } //else
  } // for
617 618
  total_event_size = model_net_get_msg_sz(DRAGONFLY) + 
      msg->remote_event_size_bytes + msg->local_event_size_bytes;
619 620 621 622
  mn_stats* stat;
  stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
  stat->send_count++;
  stat->send_bytes += msg->packet_size;
623
  stat->send_time += (1/p->cn_bandwidth) * msg->packet_size;
624 625
  if(stat->max_event_size < total_event_size)
	  stat->max_event_size = total_event_size;
626

627 628 629 630 631 632 633 634 635 636 637 638 639 640 641
  return;
}

/* sends the packet from the current dragonfly compute node to the attached router */
void packet_send(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
{
  tw_stime ts;
  tw_event *e;
  terminal_message *m;
  tw_lpid router_id;
  /* Route the packet to its source router */ 
   int vc=msg->saved_vc;

   //  Each packet is broken into chunks and then sent over the channel
   msg->saved_available_time = s->terminal_available_time;
642
   double head_delay = (1/s->params->cn_bandwidth) * s->params->chunk_size;
643
   ts = head_delay + tw_rand_exponential(lp->rng, (double)head_delay/200);
644
   s->terminal_available_time = maxd(s->terminal_available_time, tw_now(lp));
645 646
   s->terminal_available_time += ts;

647 648 649 650 651
   //TODO: be annotation-aware
   codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id, NULL,
           &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
   codes_mapping_get_lp_id(lp_group_name, "dragonfly_router", NULL, 1,
           s->router_id, 0, &router_id);
652
   // we are sending an event to the router, so no method_event here
653 654
   e = tw_event_new(router_id, s->terminal_available_time - tw_now(lp), lp);

655 656 657 658
   uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
   if(msg->packet_size % s->params->chunk_size)
       num_chunks++;

659 660 661
   if(msg->packet_ID == TRACK && msg->chunk_id == num_chunks-1)
     printf("\n terminal %d packet %lld chunk %d being sent to router %d router id %d ", (int)lp->gid, (long long)msg->packet_ID, msg->chunk_id, (int)router_id, s->router_id);
   m = tw_event_data(e);
662 663 664 665 666
   memcpy(m, msg, sizeof(terminal_message));
   if (msg->remote_event_size_bytes){
        memcpy(m+1, model_net_method_get_edata(DRAGONFLY, msg),
                msg->remote_event_size_bytes);
   }
667 668 669 670 671 672
   m->type = R_ARRIVE;
   m->src_terminal_id = lp->gid;
   m->saved_vc = vc;
   m->last_hop = TERMINAL;
   m->intm_group_id = -1;
   m->local_event_size_bytes = 0;
673
   m->local_id = s->terminal_id;
674 675 676 677 678
   tw_event_send(e);
//  Each chunk is 32B and the VC occupancy is in chunks to enable efficient flow control

   if(msg->chunk_id == num_chunks - 1) 
    {
679 680 681
      // now that message is sent, issue an "idle" event to tell the scheduler
      // when I'm next available
      model_net_method_idle_event(codes_local_latency(lp) +
682
              s->terminal_available_time - tw_now(lp), 0, lp);
683

684 685 686 687 688
      /* local completion message */
      if(msg->local_event_size_bytes > 0)
	 {
           tw_event* e_new;
	   terminal_message* m_new;
689 690 691
	   void* local_event = 
               (char*)model_net_method_get_edata(DRAGONFLY, msg) + 
               msg->remote_event_size_bytes;
692
	   ts = g_tw_lookahead + (1/s->params->cn_bandwidth) * msg->local_event_size_bytes;
693
	   e_new = tw_event_new(msg->sender_lp, ts, lp);
694 695 696 697 698 699 700 701 702
	   m_new = tw_event_data(e_new);
	   memcpy(m_new, local_event, msg->local_event_size_bytes);
	   tw_event_send(e_new);
	}
    }
   
   s->packet_counter++;
   s->vc_occupancy[vc]++;

703
   if(s->vc_occupancy[vc] >= (s->params->cn_vc_size * num_chunks))
704 705 706 707 708 709 710
      s->output_vc_state[vc] = VC_CREDIT;
   return;
}

/* packet arrives at the destination terminal */
void packet_arrive(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
{
711 712 713
    uint64_t num_chunks = msg->packet_size / s->params->chunk_size;
    if (msg->packet_size % s->params->chunk_size)
        num_chunks++;
714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738
#if DEBUG
if( msg->packet_ID == TRACK && msg->chunk_id == num_chunks-1)
    {
	printf( "(%lf) [Terminal %d] packet %lld has arrived  \n",
              tw_now(lp), (int)lp->gid, msg->packet_ID);

	printf("travel start time is %f\n",
                msg->travel_start_time);

	printf("My hop now is %d\n",msg->my_N_hop);
    }
#endif

  // Packet arrives and accumulate # queued
  // Find a queue with an empty buffer slot
   tw_event * e, * buf_e;
   terminal_message * m, * buf_msg;
   tw_stime ts;
   bf->c3 = 0;
   bf->c2 = 0;

   msg->my_N_hop++;
  if(msg->chunk_id == num_chunks-1)
  {
	 bf->c2 = 1;
739 740 741 742 743
	 mn_stats* stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
	 stat->recv_count++;
	 stat->recv_bytes += msg->packet_size;
	 stat->recv_time += tw_now(lp) - msg->travel_start_time;

744 745 746 747 748 749 750 751 752 753 754 755 756
	 N_finished_packets++;
	 dragonfly_total_time += tw_now( lp ) - msg->travel_start_time;
	 total_hops += msg->my_N_hop;

	 if (dragonfly_max_latency < tw_now( lp ) - msg->travel_start_time) 
	 {
		bf->c3 = 1;
		msg->saved_available_time = dragonfly_max_latency;
		dragonfly_max_latency=tw_now( lp ) - msg->travel_start_time;
	 }
	// Trigger an event on receiving server
	if(msg->remote_event_size_bytes)
	{
757
            void * tmp_ptr = model_net_method_get_edata(DRAGONFLY, msg);
758
            ts = g_tw_lookahead + 0.1 + (1/s->params->cn_bandwidth) * msg->remote_event_size_bytes;
759
            if (msg->is_pull){
760
                int net_id = model_net_get_id(LP_METHOD_NM);
761 762 763 764 765 766 767 768 769 770
                model_net_event(net_id, msg->category, msg->sender_lp,
                        msg->pull_size, ts, msg->remote_event_size_bytes,
                        tmp_ptr, 0, NULL, lp);
            }
            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); 
            }
771 772 773
	}
  }

774
  int credit_delay = (1 / s->params->cn_bandwidth) * CREDIT_SIZE;
775
  ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, credit_delay/1000);
776
  
777
  msg->saved_credit_time = s->next_credit_available_time;
778
  s->next_credit_available_time = maxd(s->next_credit_available_time, tw_now(lp));
779 780 781
  s->next_credit_available_time += ts;

  tw_lpid router_dest_id;
782 783 784
  //TODO: be annotation-aware
  codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id, NULL,
          &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
785
  codes_mapping_get_lp_id(lp_group_name, "dragonfly_router", s->anno, 0,
786
          s->router_id, 0, &router_dest_id);
787
  // no method_event here - message going to router
788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
  buf_e = tw_event_new(router_dest_id, s->next_credit_available_time - tw_now(lp), lp);
  buf_msg = tw_event_data(buf_e);
  buf_msg->vc_index = msg->saved_vc;
  buf_msg->type=R_BUFFER;
  buf_msg->packet_ID=msg->packet_ID;
  buf_msg->last_hop = TERMINAL;
  tw_event_send(buf_e);

  return;
}

/* initialize a dragonfly compute node terminal */
void 
terminal_init( terminal_state * s, 
	       tw_lp * lp )
{
    int i;
805 806
    char anno[MAX_NAME_LENGTH];

807
    // Assign the global router ID
808
    // TODO: be annotation-aware
809 810 811 812 813 814 815
    codes_mapping_get_lp_info(lp->gid, lp_group_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{
816
        s->anno = strdup(anno);
817 818 819 820
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }

821
   int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
822
           s->anno, 0);
823 824

   s->terminal_id = (mapping_rep_id * num_lps) + mapping_offset;  
Misbah Mubarak's avatar
Misbah Mubarak committed
825
   s->router_id=(int)s->terminal_id / (s->params->num_routers/2);
826 827 828
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;

829 830
   s->vc_occupancy = (int*)malloc(s->params->num_vcs * sizeof(int));
   s->output_vc_state = (int*)malloc(s->params->num_vcs * sizeof(int));
831

832
   for( i = 0; i < s->params->num_vcs; i++ )
833 834 835 836
    {
      s->vc_occupancy[i]=0;
      s->output_vc_state[i]=VC_IDLE;
    }
837
   dragonfly_collective_init(s, lp);
838 839 840
   return;
}

841 842 843 844 845 846 847 848 849
/* collective operation for the torus network */
void dragonfly_collective(char* category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    tw_lpid local_nic_id;
    char* tmp_ptr;

850 851 852 853
    codes_mapping_get_lp_info(sender->gid, lp_group_name, &mapping_grp_id,
            NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
    codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, NULL, 1,
            mapping_rep_id, mapping_offset, &local_nic_id);
854

855
    xfer_to_nic_time = codes_local_latency(sender);
856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894
    e_new = model_net_method_event_new(local_nic_id, xfer_to_nic_time,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);

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

    tmp_ptr = (char*)msg;
    tmp_ptr += dragonfly_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 dragonfly_collective_rc(int message_size, tw_lp* sender)
{
     codes_local_latency_reverse(sender);
     return;
}

static void send_remote_event(terminal_state * s,
                        tw_bf * bf,
                        terminal_message * msg,
                        tw_lp * lp)
{
    // Trigger an event on receiving server
    if(msg->remote_event_size_bytes)
     {
            tw_event* e;
            tw_stime ts;
            terminal_message * m;
895
            ts = (1/s->params->cn_bandwidth) * msg->remote_event_size_bytes;
896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923
            e = codes_event_new(s->origin_svr, ts, lp);
            m = tw_event_data(e);
            char* tmp_ptr = (char*)msg;
            tmp_ptr += dragonfly_get_msg_sz();
            memcpy(m, tmp_ptr, msg->remote_event_size_bytes);
            tw_event_send(e);
     }
}

static void node_collective_init(terminal_state * s,
                        tw_bf * bf,
                        terminal_message * msg,
                        tw_lp * lp)
{
        tw_event * e_new;
        tw_lpid parent_nic_id;
        tw_stime xfer_to_nic_time;
        terminal_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 */
924 925 926 927 928
            // TODO: be annotation-aware
            codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id,
                    NULL, &mapping_type_id, NULL, &mapping_rep_id,
                    &mapping_offset);
            num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
929 930
                    s->anno, 0);
            codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, s->anno, 0,
931 932
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962

           /* 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, DRAGONFLY, (void**)&msg_new, (void**)&m_data);
	    	
            memcpy(msg_new, msg, sizeof(terminal_message));
	    if (msg->remote_event_size_bytes){
        	memcpy(m_data, model_net_method_get_edata(DRAGONFLY, msg),
                	msg->remote_event_size_bytes);
      	    }
	    
            msg_new->type = D_COLLECTIVE_FAN_IN;
            msg_new->sender_node = s->node_id;

            tw_event_send(e_new);
        }
        return;
}

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

963 964 965
        codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id,
                NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
        int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
966
                s->anno, 0);
967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984

        tw_event* e_new;
        terminal_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 */
985
            codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, s->anno, 0,
986 987
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023

           /* 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, DRAGONFLY, (void**)&msg_new, &m_data);
	    
            memcpy(msg_new, msg, sizeof(terminal_message));
            msg_new->type = D_COLLECTIVE_FAN_IN;
            msg_new->sender_node = s->node_id;

            if (msg->remote_event_size_bytes){
	        memcpy(m_data, model_net_method_get_edata(DRAGONFLY, 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 */
1024 1025 1026
                codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, NULL, 1,
                        s->children[i]/num_lps, (s->children[i] % num_lps),
                        &child_nic_id);
1027 1028 1029 1030 1031 1032 1033 1034 1035 1036
                //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, DRAGONFLY, (void**)&msg_new, &m_data);

		memcpy(msg_new, msg, sizeof(terminal_message));
	        if (msg->remote_event_size_bytes){
1037
	                memcpy(m_data, model_net_method_get_edata(DRAGONFLY, msg),
1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054
        	               msg->remote_event_size_bytes);
      		}
		
                msg_new->type = D_COLLECTIVE_FAN_OUT;
                msg_new->sender_node = s->node_id;

                tw_event_send(e_new);
           }
      }
}

static void node_collective_fan_out(terminal_state * s,
                        tw_bf * bf,
                        terminal_message * msg,
                        tw_lp * lp)
{
        int i;
1055 1056
        int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
                NULL, 1);
1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077
        bf->c1 = 0;
        bf->c2 = 0;

        send_remote_event(s, bf, msg, lp);

        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 + DRAGONFLY_FAN_OUT_DELAY + tw_rand_exponential(lp->rng, (double)DRAGONFLY_FAN_OUT_DELAY/10);

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

                        /* get global LP ID of the child node */
1078
                        codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM,
1079
                                s->anno, 0, s->children[i]/num_lps,
1080
                                (s->children[i] % num_lps), &child_nic_id);
1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107
                        //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, DRAGONFLY, (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(DRAGONFLY, msg),
			                msg->remote_event_size_bytes);
      			}


                        msg_new->type = D_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;
          }
}
1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
/* update the compute node-router channel buffer */
void 
terminal_buf_update(terminal_state * s, 
		    tw_bf * bf, 
		    terminal_message * msg, 
		    tw_lp * lp)
{
  // Update the buffer space associated with this router LP 
    int msg_indx = msg->vc_index;
    
    s->vc_occupancy[msg_indx]--;
    s->output_vc_state[msg_indx] = VC_IDLE;

    return;
}

void 
terminal_event( terminal_state * s, 
		tw_bf * bf, 
		terminal_message * msg, 
		tw_lp * lp )
{
  *(int *)bf = (int)0;
  switch(msg->type)
    {
    case T_GENERATE:
       packet_generate(s,bf,msg,lp);
    break;
    
    case T_ARRIVE:
        packet_arrive(s,bf,msg,lp);
    break;
    
    case T_SEND:
      packet_send(s,bf,msg,lp);
    break;
    
    case T_BUFFER:
       terminal_buf_update(s, bf, msg, lp);
     break;
1148 1149 1150 1151
    
    case D_COLLECTIVE_INIT:
      node_collective_init(s, bf, msg, lp);
    break;
1152

1153 1154 1155 1156 1157 1158 1159 1160
    case D_COLLECTIVE_FAN_IN:
      node_collective_fan_in(s, bf, msg, lp);
    break;

    case D_COLLECTIVE_FAN_OUT:
      node_collective_fan_out(s, bf, msg, lp);
    break;
    
1161 1162 1163 1164 1165 1166
    default:
       printf("\n LP %d Terminal message type not supported %d ", (int)lp->gid, msg->type);
    }
}

void 
1167
dragonfly_terminal_final( terminal_state * s, 
1168 1169
      tw_lp * lp )
{
1170
	model_net_print_stats(lp->gid, s->dragonfly_stats_array);
1171 1172
}

1173 1174 1175 1176 1177
void dragonfly_router_final(router_state * s,
		tw_lp * lp)
{
   free(s->global_channel);
}
1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188
/* get the next stop for the current packet
 * determines if it is a router within a group, a router in another group
 * or the destination terminal */
tw_lpid 
get_next_stop(router_state * s, 
		      tw_bf * bf, 
		      terminal_message * msg, 
		      tw_lp * lp, 
		      int path)
{
   int dest_lp;
1189
   tw_lpid router_dest_id = -1;
1190 1191 1192
   int i;
   int dest_group_id;

1193 1194 1195 1196 1197
   //TODO: be annotation-aware
   codes_mapping_get_lp_info(msg->dest_terminal_id, lp_group_name,
           &mapping_grp_id, NULL, &mapping_type_id, NULL, &mapping_rep_id,
           &mapping_offset); 
   int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
1198 1199
           s->anno, 0);
   int dest_router_id = (mapping_offset + (mapping_rep_id * num_lps)) / s->params->num_routers;
1200
   
1201 1202
   codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id, NULL,
           &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
1203 1204 1205 1206
   int local_router_id = (mapping_offset + mapping_rep_id);

   bf->c2 = 0;

1207
  /* If the packet has arrived at the destination router */
1208 1209 1210 1211 1212 1213
   if(dest_router_id == local_router_id)
    {
        dest_lp = msg->dest_terminal_id;

        return dest_lp;
    }
1214
   /* Generate inter-mediate destination for non-minimal routing (selecting a random group) */
1215
   if(msg->last_hop == TERMINAL && msg->path_type == NON_MINIMAL)
1216
    {
1217
      if(dest_router_id / s->params->num_routers != s->group_id)
1218 1219
         {
            bf->c2 = 1;
1220
            int intm_grp_id = tw_rand_integer(lp->rng, 0, s->params->num_groups-1);
1221 1222
            //int intm_grp_id = (s->group_id + s->group_id/2) % num_groups;
	    msg->intm_group_id = intm_grp_id;
1223 1224
          }    
    }
1225
  /* It means that the packet has arrived at the inter-mediate group for non-minimal routing. Reset the group now. */
1226 1227 1228 1229
   if(msg->intm_group_id == s->group_id)
   {  
           msg->intm_group_id = -1;//no inter-mediate group
   } 
1230
  /* Intermediate group ID is set. Divert the packet to an intermediate group. */
1231 1232 1233 1234
  if(msg->intm_group_id >= 0)
   {
      dest_group_id = msg->intm_group_id;
   }
1235
  else /* direct the packet to the destination group */
1236
   {
1237
     dest_group_id = dest_router_id / s->params->num_routers;
1238 1239
   }
  
1240
  /* It means the packet has arrived at the destination group. Now divert it to the destination router. */
1241 1242 1243 1244 1245 1246
  if(s->group_id == dest_group_id)
   {
     dest_lp = dest_router_id;
   }
   else
   {
1247
      /* Packet is at the source or intermediate group. Find a router that has a path to the destination group. */
1248 1249 1250 1251
      dest_lp=getRouterFromGroupID(dest_group_id,s);
  
      if(dest_lp == local_router_id)
      {
1252
        for(i=0; i < s->params->num_global_channels; i++)
1253
           {
1254
            if(s->global_channel[i] / s->params->num_routers == dest_group_id)
1255 1256 1257 1258
                dest_lp=s->global_channel[i];
          }
      }
   }
1259
  codes_mapping_get_lp_id(lp_group_name, "dragonfly_router", s->anno, 0, dest_lp,
1260
          0, &router_dest_id);
1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272
  return router_dest_id;
}

/* gets the output port corresponding to the next stop of the message */
int 
get_output_port( router_state * s, 
		tw_bf * bf, 
		terminal_message * msg, 
		tw_lp * lp, 
		int next_stop )
{
  int output_port = -1, i, terminal_id;
1273 1274 1275
  codes_mapping_get_lp_info(msg->dest_terminal_id, lp_group_name,
          &mapping_grp_id, NULL, &mapping_type_id, NULL, &mapping_rep_id,
          &mapping_offset);
1276
  int num_lps = codes_mapping_get_lp_count(lp_group_name,1,LP_CONFIG_NM,s->anno,0);
1277
  terminal_id = (mapping_rep_id * num_lps) + mapping_offset;
1278 1279 1280

  if(next_stop == msg->dest_terminal_id)
   {
1281 1282
      output_port = s->params->num_routers + s->params->num_global_channels +
          ( terminal_id % s->params->num_cn);
1283 1284
      //if(output_port > 6)
	//      printf("\n incorrect output port %d terminal id %d ", output_port, terminal_id);
1285 1286 1287
    }
    else
    {
1288 1289
     codes_mapping_get_lp_info(next_stop, lp_group_name, &mapping_grp_id,
             NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
1290
     int local_router_id = mapping_rep_id + mapping_offset;
1291
     int intm_grp_id = local_router_id / s->params->num_routers;
1292 1293 1294

     if(intm_grp_id != s->group_id)
      {
1295
        for(i=0; i < s->params->num_global_channels; i++)
1296
         {
1297
           if(s->global_channel[i] == local_router_id)
1298
             output_port = s->params->num_routers + i;
1299 1300 1301 1302
          }
      }
      else
       {
1303
        output_port = local_router_id % s->params->num_routers;
1304
       }
1305
//	      printf("\n output port not found %d next stop %d local router id %d group id %d intm grp id %d %d", output_port, next_stop, local_router_id, s->group_id, intm_grp_id, local_router_id%num_routers);
1306 1307 1308 1309 1310 1311 1312 1313 1314 1315
    }
    return output_port;
}

/* routes the current packet to the next stop */
void 
router_packet_send( router_state * s, 
		    tw_bf * bf, 
		     terminal_message * msg, tw_lp * lp)
{
1316
//   *(int *)bf = (int)0;
1317 1318 1319 1320 1321
   tw_stime ts;
   tw_event *e;
   terminal_message *m;

   int next_stop = -1, output_port = -1, output_chan = -1;
1322 1323
   float bandwidth = s->params->local_bandwidth;
   int path = s->params->routing;
1324
   int minimal_out_port = -1, nonmin_out_port = -1;
1325 1326
   bf->c3 = 0;

1327 1328 1329 1330 1331 1332
   uint64_t num_chunks = msg->packet_size/s->params->chunk_size;
   if(msg->packet_size % s->params->chunk_size)
       num_chunks++;
    

   if(msg->last_hop == TERMINAL && s->params->routing == ADAPTIVE)
1333
   {
1334 1335 1336 1337 1338 1339 1340
  // decide which routing to take
    int minimal_next_stop=get_next_stop(s, bf, msg, lp, MINIMAL);
    minimal_out_port = get_output_port(s, bf, msg, lp, minimal_next_stop);
    int nonmin_next_stop = get_next_stop(s, bf, msg, lp, NON_MINIMAL);
    nonmin_out_port = get_output_port(s, bf, msg, lp, nonmin_next_stop);
    int nonmin_port_count = s->vc_occupancy[nonmin_out_port];
    int min_port_count = s->vc_occupancy[minimal_out_port];
1341 1342
    int nonmin_vc = s->vc_occupancy[nonmin_out_port * s->params->num_vcs + 2];
    int min_vc = s->vc_occupancy[minimal_out_port * s->params->num_vcs + 1];
1343 1344

    // Adaptive routing condition from the dragonfly paper Page 83
1345
   // modified according to booksim adaptive routing condition
1346 1347 1348
   if((min_vc <= (nonmin_vc * 2 + adaptive_threshold) && minimal_out_port == nonmin_out_port)
               || (min_port_count <= (nonmin_port_count * 2 + adaptive_threshold) && minimal_out_port != nonmin_out_port))
        {
1349
	   msg->path_type = MINIMAL;
1350 1351 1352 1353
           next_stop = minimal_next_stop;
           output_port = minimal_out_port;
           minimal_count++;
           msg->intm_group_id = -1;
1354

1355 1356 1357 1358 1359
           if(msg->packet_ID == TRACK)
              printf("\n (%lf) [Router %d] Packet %d routing minimally ", tw_now(lp), (int)lp->gid, (int)msg->packet_ID);
        }
       else
         {
1360
	   msg->path_type = NON_MINIMAL;
1361 1362 1363 1364 1365 1366 1367 1368 1369 1370
           next_stop = nonmin_next_stop;
           output_port = nonmin_out_port;
           nonmin_count++;
           if(msg->packet_ID == TRACK)
                printf("\n (%lf) [Router %d] Packet %d routing non-minimally ", tw_now(lp), (int)lp->gid, (int)msg->packet_ID);

         }
  }
  else
   {
1371
	msg->path_type = routing; /*defaults to the routing algorithm if we don't have adaptive routing here*/
1372 1373 1374
   	next_stop = get_next_stop(s, bf, msg, lp, path);
   	output_port = get_output_port(s, bf, msg, lp, next_stop); 
   }
1375
   output_chan = output_port * s->params->num_vcs;
1376
    // Even numbered channels for minimal routing
1377
   // Odd numbered channels for nonminimal routing
1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388
   // Separate the queue occupancy into minimal and non minimal virtual channels if the min & non min
   // paths start at the same output port
   /*if((routing == ADAPTIVE) && (minimal_out_port == nonmin_out_port))
   {
        if(path == MINIMAL)
          output_chan = output_chan + 1;
        else
          if(path == NON_MINIMAL)
            output_chan = output_chan + 2;
   }*/

1389
   int global=0;
1390
   int buf_size = s->params->local_vc_size;
1391

1392 1393
   assert(output_port != -1);
   assert(output_chan != -1);
1394
   // Allocate output Virtual Channel
1395 1396
  if(output_port >= s->params->num_routers && 
          output_port < s->params->num_routers + s->params->num_global_channels)
1397
  {
1398
	 bandwidth = s->params->global_bandwidth;
1399
	 global = 1;
1400
	 buf_size = s->params->global_vc_size;
1401 1402
  }

1403 1404
  if(output_port >= s->params->num_routers + s->params->num_global_channels)
	buf_size = s->params->cn_vc_size;
1405 1406 1407

   if(s->vc_occupancy[output_chan] >= buf_size)
    {
1408
	    printf("\n %lf Router %ld buffers overflowed from incoming terminals channel %d occupancy %d radix %d next_stop %d ", tw_now(lp),(long int) lp->gid, output_chan, s->vc_occupancy[output_chan], s->params->radix, next_stop);
1409
	    bf->c3 = 1;
1410 1411 1412
	    return;
	    //MPI_Finalize();
	    //exit(-1);
1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424
    }

#if DEBUG
if( msg->packet_ID == TRACK && next_stop != msg->dest_terminal_id && msg->chunk_id == num_chunks-1)
  {
   printf("\n (%lf) [Router %d] Packet %lld being sent to intermediate group router %d Final destination terminal %d Output Channel Index %d Saved vc %d msg_intm_id %d \n", 
              tw_now(lp), (int)lp->gid, msg->packet_ID, next_stop, 
	      msg->dest_terminal_id, output_chan, msg->saved_vc, msg->intm_group_id);
  }
#endif
 // If source router doesn't have global channel and buffer space is available, then assign to appropriate intra-group virtual channel 
  msg->saved_available_time = s->next_output_available_time[output_port];
1425
  ts = g_tw_lookahead + 0.1 + ((1/bandwidth) * s->params->chunk_size) + tw_rand_exponential(lp->rng, (double)s->params->chunk_size/200);
1426

1427
  s->next_output_available_time[output_port] = maxd(s->next_output_available_time[output_port], tw_now(lp));
1428
  s->next_output_available_time[output_port] += ts;
1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444
  // dest can be a router or a terminal, so we must check
  void * m_data;
  if (next_stop == msg->dest_terminal_id){
      e = model_net_method_event_new(next_stop, 
              s->next_output_available_time[output_port] - tw_now(lp), lp,
              DRAGONFLY, (void**)&m, &m_data);
  }
  else{
      e = tw_event_new(next_stop, s->next_output_available_time[output_port] - tw_now(lp), lp);
      m = tw_event_data(e);
      m_data = m+1;
  }
  memcpy(m, msg, sizeof(terminal_message));
  if (msg->remote_event_size_bytes){
      memcpy(m_data, msg+1, msg->remote_event_size_bytes);
  }
1445 1446 1447 1448 1449 1450 1451

  if(global)
    m->last_hop=GLOBAL;
  else
    m->last_hop = LOCAL;

  m->saved_vc = output_chan;
1452
  m->local_id = s->router_id;
1453 1454 1455 1456
  msg->old_vc = output_chan;
  m->intm_lp_id = lp->gid;
  s->vc_occupancy[output_chan]++;

1457 1458
  /* Determine the event type. If the packet has arrived at the final destination
     router then it should arrive at the destination terminal next. */
1459 1460 1461 1462
  if(next_stop == msg->dest_terminal_id)
  {
    m->type = T_ARRIVE;

1463
    if(s->vc_occupancy[output_chan] >= s->params->cn_vc_size * num_chunks)
1464 1465 1466 1467
      s->output_vc_state[output_chan] = VC_CREDIT;
  }
  else
  {
1468
    /* The packet has to be sent to another router */
1469 1470
    m->type = R_ARRIVE;

1471
   /* If this is a global channel then the buffer space is different */
1472 1473
   if