dragonfly.c 69.6 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 9 10
// Local router ID: 0 --- total_router-1
// Router LP ID 
// Terminal LP ID

11 12 13
#include <ross.h>

#include "codes/codes_mapping.h"
14
#include "codes/jenkins-hash.h"
15 16 17
#include "codes/codes.h"
#include "codes/model-net.h"
#include "codes/model-net-method.h"
18 19
#include "codes/model-net-lp.h"
#include "codes/net/dragonfly.h"
20 21 22 23

#define CREDIT_SIZE 8
#define MEAN_PROCESS 1.0

24 25 26
/* collective specific parameters */
#define TREE_DEGREE 4
#define LEVEL_DELAY 1000
27
#define DRAGONFLY_COLLECTIVE_DEBUG 0
28 29 30
#define NUM_COLLECTIVES  1
#define COLLECTIVE_COMPUTATION_DELAY 5700
#define DRAGONFLY_FAN_OUT_DELAY 20.0
31
#define WINDOW_LENGTH 0
32

33
// debugging parameters
34
#define TRACK 25
35
#define PRINT_ROUTER_TABLE 1
36
#define DEBUG 0
37

38 39 40
#define LP_CONFIG_NM (model_net_lp_config_names[DRAGONFLY])
#define LP_METHOD_NM (model_net_method_names[DRAGONFLY])

41 42 43 44 45 46
#define DRAGONFLY_DBG 0
#define dprintf(_fmt, ...) \
    do {if (CLIENT_DBG) printf(_fmt, __VA_ARGS__);} while (0)

long term_ecount, router_ecount, term_pending_ecount, router_pending_ecount;

47 48
static double maxd(double a, double b) { return a < b ? b : a; }

49 50
// arrival rate
static double MEAN_INTERVAL=200.0;
51 52
// threshold for adaptive routing
static int adaptive_threshold = 10;
53 54

/* minimal and non-minimal packet counts for adaptive routing*/
55
unsigned int minimal_count=0, nonmin_count=0, completed_packets = 0;
56

57 58 59 60 61 62
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;
63 64

/* global variables for codes mapping */
65
static char lp_group_name[MAX_NAME_LENGTH];
66 67
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

68 69 70 71 72 73 74 75
/* router magic number */
int router_magic_num = 0;

/* terminal magic number */
int terminal_magic_num = 0;

/* number of routers in a mapping group */
static int num_routers_per_mgrp = 0;
76

77 78 79 80 81 82
/* maximum number of terminals and routers */
int max_term_occupancy, max_router_occupancy;

/* noise of 1ns */
double noise = 1.0;

83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
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 chunk_size; /* full-sized packets are broken into smaller chunks.*/

    // derived parameters
    int num_cn;
    int num_groups;
    int radix;
    int total_routers;
101
    int total_terminals;
102 103 104
    int num_global_channels;
};

105 106 107 108 109 110 111 112 113 114 115 116
/* 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
117 118
   tw_lpid router_id;
   tw_lpid terminal_id;
119 120 121 122 123 124 125 126 127 128

   // 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];
129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
  /* 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;
153
   int max_term_vc_occupancy;
154 155 156

   const char * anno;
   const dragonfly_param *params;
157
};
158

159 160 161 162 163 164
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
  T_BUFFER,
165
  R_FORWARD,
166 167 168 169
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192
};
/* 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
{
193 194
    MINIMAL = 0,
    NON_MINIMAL,
195 196
    ADAPTIVE,
    PROG_ADAPTIVE
197 198 199 200 201 202 203 204
};

struct router_state
{
   unsigned int router_id;
   unsigned int group_id;
  
   int* global_channel; 
205
   
206 207
   tw_stime* next_output_available_time;
   tw_stime* next_credit_available_time;
208 209
   tw_stime* cur_hist_start_time;
   
210 211
   int* vc_occupancy;
   int* output_vc_state;
212
   int max_router_vc_occupancy;
213 214 215

   const char * anno;
   const dragonfly_param *params;
216 217 218

   int* prev_hist_num;
   int* cur_hist_num;
219 220 221 222 223 224
};

static short routing = MINIMAL;

static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
225
static tw_stime         max_collective = 0;
226 227 228 229 230 231 232 233 234 235 236 237 238 239


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);
}

240 241 242
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
243

244 245 246 247 248 249 250 251 252 253
    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);
254
    if(!p->num_vcs) {
255 256 257 258 259
        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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
260
    if(p->local_vc_size <= 0) {
261 262 263 264 265
        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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
266
    if(p->global_vc_size <= 0) {
267 268 269 270 271
        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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
272
    if(p->cn_vc_size <= 0) {
273 274 275 276 277
        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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
278
    if(p->chunk_size <= 0) {
279
        p->chunk_size = 64;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
280
        fprintf(stderr, "Chunk size for packets is not specified, setting to %d\n", p->chunk_size);
281 282 283
    }

    configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
284
    if(p->local_bandwidth <= 0) {
285 286 287 288 289
        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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
290
    if(p->global_bandwidth <= 0) {
291 292 293 294 295
        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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
296
    if(p->cn_bandwidth <= 0) {
297 298 299 300 301
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }


302

303 304
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
305
            MAX_NAME_LENGTH);
306 307 308 309 310 311 312 313
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
    else if(strcmp(routing_str, "nonminimal")==0 || strcmp(routing_str,"non-minimal")==0)
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
314 315 316 317
    else
    {
        fprintf(stderr, 
                "No routing protocol specified, setting to minimal routing\n");
318
        routing = -1;
319 320 321 322 323 324 325 326 327
    }

    // 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;
328 329


330
    p->total_terminals = p->total_routers * p->num_cn;
331
    printf("\n Total nodes %d routers %d groups %d radix %d num_vc %d ", p->num_cn * p->total_routers,
332 333
								p->total_routers,
								p->num_groups,
334 335
								p->radix,
                                                                p->num_vcs);
336 337
}

338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353
/* convert GiB/s and bytes to ns */
static tw_stime bytes_to_ns(uint64_t bytes, double GB_p_s)
{
    tw_stime time;

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

    return(time);
}

/* reverse computation for msg ready event */
354 355 356 357
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);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
358
    all_params = calloc(num_params, sizeof(*all_params));
359 360

    for (uint64_t i = 0; i < anno_map->num_annos; i++){
361
        const char * anno = anno_map->annotations[i].ptr;
362 363 364 365 366
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
367 368 369 370 371 372 373
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
   long long avg_hops, total_finished_packets;
   tw_stime avg_time, max_time;
374
   long total_term_events, total_router_events;
375
   int total_minimal_packets, total_nonmin_packets, total_completed_packets;
376 377 378 379 380

   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);
381 382 383
   MPI_Reduce( &term_ecount, &total_term_events, 1, MPI_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
   MPI_Reduce( &router_ecount, &total_router_events, 1, MPI_LONG, MPI_SUM, 0, MPI_COMM_WORLD);

384
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
385 386 387
    {
	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);
388
 	MPI_Reduce(&completed_packets, &total_completed_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
389
    }
390 391
   /* print statistics */
   if(!g_tw_mynode)
392
   {	
393 394
      printf("\n total finished packets %lld ", total_finished_packets);
      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);
395
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
396
              printf("\n ADAPTIVE ROUTING STATS: %d packets routed minimally %d packets routed non-minimally completed packets %d ", total_minimal_packets, total_nonmin_packets, total_completed_packets);
397
 
398 399
     printf("\n Max terminal occupancy %d max router occupancy %d ", max_term_occupancy, max_router_occupancy);
     printf("\n Total terminal events generated %ld router events %ld ", total_term_events, total_router_events);
400
  }
401 402
   return;
}
403

404 405 406
void dragonfly_collective_init(terminal_state * s,
           		   tw_lp * lp)
{
407 408 409 410 411
    // 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);
412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
    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
}

462
/* dragonfly packet event , generates a dragonfly packet on the compute node */
463
static tw_stime dragonfly_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)
464 465 466 467 468 469
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

470
    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*/
471 472
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
473 474
    strcpy(msg->category, category);
    msg->final_dest_gid = final_dest_lp;
475
    msg->dest_terminal_id = dest_mn_lp;
476
    msg->sender_lp=src_lp;
477
    msg->sender_mn_lp = sender->gid;
478 479 480 481
    msg->packet_size = packet_size;
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
482
    msg->magic = terminal_magic_num;
483 484
    msg->is_pull = is_pull;
    msg->pull_size = pull_size;
485
    msg->chunk_id = 0;
486
    msg->packet_ID = 0;
487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503

    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;
	}
     }
    tw_event_send(e_new);
504
    return xfer_to_nic_time;
505 506 507 508 509 510 511 512 513
}

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

514 515 516 517
/* given two group IDs, find the router of the src_gid that connects to the dest_gid*/
tw_lpid getRouterFromGroupID(int dest_gid, 
		    int src_gid,
		    int num_routers)
518
{
519 520 521
  int group_begin = src_gid * num_routers;
  int group_end = (src_gid * num_routers) + num_routers-1;
  int offset = (dest_gid * num_routers - group_begin) / num_routers;
522
  
523 524
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
525
  
526 527
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
528
  
529
  offset=(offset - 1) % (half_channel * num_routers);
530 531 532 533 534 535 536 537 538 539 540 541 542 543 544

  // 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)
{
545
  router_ecount++;
546 547 548 549
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

550
  int dest=0, type = R_BUFFER;
551
  int is_terminal = 0;
552
  int found_magic = router_magic_num;
553
  tw_stime credit_delay = 0.0;
554

555
  const dragonfly_param *p = s->params;
556
  int sender_radix;
557 558 559 560
 // Notify sender terminal about available buffer space
  if(msg->last_hop == TERMINAL)
  {
   dest = msg->src_terminal_id;
561
   sender_radix = msg->local_id % p->num_cn;  
562
   //determine the time in ns to transfer the credit
563
   credit_delay = bytes_to_ns(CREDIT_SIZE, p->cn_bandwidth);
564
   type = T_BUFFER;
565
   is_terminal = 1;
566
   found_magic = terminal_magic_num;
567 568 569 570
  }
   else if(msg->last_hop == GLOBAL)
   {
     dest = msg->intm_lp_id;
571
     sender_radix = p->num_cn + (msg->local_id % p->num_global_channels);
572
     credit_delay = bytes_to_ns(CREDIT_SIZE, p->global_bandwidth);
573 574 575 576
   }
    else if(msg->last_hop == LOCAL)
     {
        dest = msg->intm_lp_id;
577
        sender_radix = p->num_cn + p->num_global_channels + (msg->local_id % p->num_routers);
578
     	credit_delay = bytes_to_ns(CREDIT_SIZE, p->local_bandwidth) * CREDIT_SIZE;
579 580 581 582
     }
    else
      printf("\n Invalid message type");

583 584
    msg->sender_radix = sender_radix;

585
    assert(sender_radix < s->params->radix );
586

587 588
    msg->saved_credit_time = s->next_credit_available_time[sender_radix];
    s->next_credit_available_time[sender_radix] = maxd(tw_now(lp), s->next_credit_available_time[sender_radix]);
589
    ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, (double)credit_delay/1000);
590
	
591
    s->next_credit_available_time[sender_radix]+=ts;
592 593
    if (is_terminal){
        buf_e = model_net_method_event_new(dest, 
594
                s->next_credit_available_time[sender_radix] - tw_now(lp), lp,
595 596 597
                DRAGONFLY, (void**)&buf_msg, NULL);
    }
    else{
598
        buf_e = tw_event_new(dest, s->next_credit_available_time[sender_radix] - tw_now(lp) , lp);
599 600
        buf_msg = tw_event_data(buf_e);
    }
601
    buf_msg->origin_router_id = s->router_id;
602 603
    buf_msg->vc_index = msg->saved_vc;
    buf_msg->type=type;
604
    buf_msg->magic = found_magic;
605 606 607 608 609 610 611 612
    buf_msg->last_hop = msg->last_hop;
    buf_msg->packet_ID=msg->packet_ID;

    tw_event_send(buf_e);

    return;
}

613
static void packet_generate_send_rc(terminal_state * s, 
614 615 616
			    tw_bf * bf, 
			    terminal_message * msg, 
			    tw_lp * lp)
617
{
618 619
   term_ecount++;
   tw_rand_reverse_unif(lp->rng);
620
	 
621 622 623 624 625 626 627
   s->terminal_available_time = msg->saved_available_time;
   tw_rand_reverse_unif(lp->rng);
   int vc = msg->saved_vc;
   s->vc_occupancy[vc]--;
   s->packet_counter--;
   s->output_vc_state[vc] = VC_IDLE;

628 629 630
   if(bf->c2)
     s->max_term_vc_occupancy = msg->saved_occupancy;

631 632 633
   if (msg->chunk_id == (msg->num_chunks-1)){
     codes_local_latency_reverse(lp);
   }
634 635 636 637

    if(bf->c1)
        codes_local_latency_reverse(lp);

638
     struct mn_stats* stat;
639 640 641 642 643
     stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
     stat->send_count--;
     stat->send_bytes -= msg->packet_size;
     stat->send_time -= (1/s->params->cn_bandwidth) * msg->packet_size;
}
644

645
/* generates packet at the current dragonfly compute node */
646
static void packet_generate_send(terminal_state * s, 
647 648 649 650
			    tw_bf * bf, 
			    terminal_message * msg, 
			    tw_lp * lp)
{
651
	bf->c1 = 0;
652 653
	term_ecount++;

654
        const dragonfly_param *p = s->params;
655

656
	tw_stime ts, travel_start_time;
657
	tw_event *e;
658
        tw_lpid router_id;
659 660 661 662 663 664
	terminal_message *m;
	int i, total_event_size;

	uint64_t num_chunks = msg->packet_size / p->chunk_size;
	if (msg->packet_size % s->params->chunk_size)
	  num_chunks++;
665

666 667
	if(!num_chunks)
	   num_chunks = 1;
668

669
	msg->num_chunks = num_chunks;
670 671 672
	
	if(!msg->packet_ID)
	    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);
673

674 675
	//  Each packet is broken into chunks and then sent over the channel
	msg->saved_available_time = s->terminal_available_time;
676 677 678
	tw_stime head_delay = bytes_to_ns(s->params->chunk_size, s->params->cn_bandwidth);
	ts = head_delay + tw_rand_exponential(lp->rng, noise);
	//printf("\n ts %f calculated %f ", ts, s->params->chunk_size * (1/s->params->cn_bandwidth));
679 680
	s->terminal_available_time = maxd(s->terminal_available_time, tw_now(lp));
	s->terminal_available_time += ts;
681

682
	
683 684 685
	int chan = -1, j;
	for(j = 0; j < p->num_vcs; j++)
	 {
686
	     if(s->vc_occupancy[j] < p->cn_vc_size)
687 688 689 690 691
	      {
	       chan=j;
	       break;
	      }
	 }
692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710
	  
	/* for reverse computation */
	msg->saved_vc = chan;
  
        /* simulation should exit */
	if(chan == -1)
	    tw_error(TW_LOC, "\n No terminal buffers available, increase buffer size");
	   
        //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/num_routers_per_mgrp, 
		   s->router_id % num_routers_per_mgrp, &router_id);
   
        e = tw_event_new(router_id, s->terminal_available_time - tw_now(lp), lp);
	m = tw_event_data(e);
	memcpy(m, msg, sizeof(terminal_message));
711 712
        
	if(msg->chunk_id == 0)
713 714 715
            travel_start_time = tw_now(lp);
	else
	    travel_start_time = msg->travel_start_time;
716
	
717 718
	m->magic = router_magic_num;
	m->origin_router_id = s->router_id;
719
        m->type = R_FORWARD;
720 721 722 723
        m->src_terminal_id = lp->gid;
        m->chunk_id = msg->chunk_id;
        m->last_hop = TERMINAL;
        m->intm_group_id = -1;
724 725
        m->travel_start_time = travel_start_time;
	m->path_type = -1;
726 727
        m->local_event_size_bytes = 0;
        m->local_id = s->terminal_id;
728

729 730 731 732 733 734
	 if (msg->remote_event_size_bytes){
		memcpy(m+1, model_net_method_get_edata(DRAGONFLY, msg),
			msg->remote_event_size_bytes);
	   }
   
	tw_event_send(e);
735 736 737 738 739 740 741 742

	if(msg->packet_ID == TRACK && msg->chunk_id == msg->num_chunks - 1)
	   printf("\n packet %d generated chunk id %d reach at time %lf ", 
			msg->packet_ID, 
			msg->chunk_id, 
			s->terminal_available_time - tw_now(lp));	

         if(msg->chunk_id == msg->num_chunks - 1) 
743
	{
744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762
		// 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) +
		      s->terminal_available_time - tw_now(lp), 0, lp);

		/* local completion message */
		if(msg->local_event_size_bytes > 0)
		 {
		   tw_event* e_new;
		   terminal_message* m_new;
		   void* local_event = 
		       (char*)model_net_method_get_edata(DRAGONFLY, msg) + 
		       msg->remote_event_size_bytes;
		   ts = g_tw_lookahead + (1/s->params->cn_bandwidth) * msg->local_event_size_bytes;
		   e_new = tw_event_new(msg->sender_lp, ts, lp);
		   m_new = tw_event_data(e_new);
		   memcpy(m_new, local_event, msg->local_event_size_bytes);
		   tw_event_send(e_new);
		}
763
	}
764 765 766
	s->packet_counter++;
	s->vc_occupancy[chan]++;

767
	if(s->vc_occupancy[chan] > s->max_term_vc_occupancy)
768 769 770
	 {
		bf->c2 = 1;
		msg->saved_occupancy = s->max_term_vc_occupancy;
771
		s->max_term_vc_occupancy = s->vc_occupancy[chan];
772
	 }
773

774 775 776 777 778 779 780 781 782 783
	if(s->vc_occupancy[chan] >= s->params->cn_vc_size)
	    s->output_vc_state[chan] = VC_CREDIT;

	/* calculate statistics */
	total_event_size = model_net_get_msg_sz(DRAGONFLY) + 
	msg->remote_event_size_bytes + msg->local_event_size_bytes;
	struct mn_stats* stat;
	stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
	stat->send_count++;
	stat->send_bytes += msg->packet_size;
784
	stat->send_time += (s->terminal_available_time - tw_now(lp));
785 786
	if(stat->max_event_size < total_event_size)
	  stat->max_event_size = total_event_size;
787 788

        /* Now schedule another packet generate event */
789
	if(msg->chunk_id < num_chunks - 1)
790
	{
791
	     bf->c1 = 1;
792 793 794 795 796 797
	     /* Issue another packet generate event */
	     tw_event * e_gen;
	     terminal_message * m_gen;
	     void * m_gen_data;

	     /* Keep the packet generate event a little behind packet send */
798
	     ts = s->terminal_available_time - tw_now(lp) + codes_local_latency(lp);
799 800 801 802 803
	     e_gen = model_net_method_event_new(lp->gid, ts, lp, DRAGONFLY, (void**)&m_gen,(void**)&m_gen_data); 

	     void *m_gen_data_src = model_net_method_get_edata(DRAGONFLY, msg);
	     memcpy(m_gen, msg, sizeof(terminal_message));

804
	     m_gen->chunk_id = msg->chunk_id + 1;
805
             m_gen->type = T_GENERATE;
806 807
	     m_gen->travel_start_time = travel_start_time;
     		
808 809 810 811 812 813 814 815 816 817 818
	      if (msg->remote_event_size_bytes){
		memcpy(m_gen_data, m_gen_data_src,
			msg->remote_event_size_bytes);
		m_gen_data = (char*)m_gen_data + msg->remote_event_size_bytes;
		m_gen_data_src = (char*)m_gen_data_src + msg->remote_event_size_bytes;
	      }
	    if (msg->local_event_size_bytes)
		memcpy(m_gen_data, m_gen_data_src, msg->local_event_size_bytes);
     
           tw_event_send(e_gen); 
       }
819
	
820 821 822
  return;
}

823 824 825 826 827
static void packet_arrive_rc(terminal_state * s, 
		   tw_bf * bf, 
                   terminal_message * msg, 
                   tw_lp * lp)
{
828
    term_ecount++; 
829
    uint64_t num_chunks = msg->num_chunks;
830 831 832 833 834
    if (msg->packet_size % s->params->chunk_size)
        num_chunks++;
   
    if(msg->chunk_id == num_chunks - 1)
      completed_packets--;
835

836
   if(msg->path_type == MINIMAL && msg->chunk_id == num_chunks - 1)
837 838
	minimal_count--;

839
   if(msg->path_type == NON_MINIMAL && msg->chunk_id == num_chunks - 1)
840 841 842 843
	nonmin_count--;

   tw_rand_reverse_unif(lp->rng);
   s->next_credit_available_time = msg->saved_credit_time;
844
   if(msg->chunk_id == num_chunks-1)
845 846 847 848 849 850 851 852 853 854 855 856 857 858
   {
    mn_stats* stat;
    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;
    
    N_finished_packets--;
    
    dragonfly_total_time -= (tw_now(lp) - msg->travel_start_time);
    if(bf->c3)
	 dragonfly_max_latency = msg->saved_available_time;
   }
    
859
   if (msg->chunk_id == (msg->num_chunks)-1 && 
860 861 862 863 864 865 866 867
		msg->remote_event_size_bytes && 
		msg->is_pull)
    {
	    int net_id = model_net_get_id(LP_METHOD_NM);
	    model_net_event_rc(net_id, lp, msg->pull_size);
    }

}
868
/* packet arrives at the destination terminal */
869 870 871 872
static void packet_arrive(terminal_state * s, 
		   tw_bf * bf, 
                   terminal_message * msg, 
                   tw_lp * lp)
873
{
874
    ++term_ecount;
875
    uint64_t num_chunks = msg->num_chunks;
876 877
    if (msg->packet_size % s->params->chunk_size)
        num_chunks++;
878

879 880
   if(msg->chunk_id == num_chunks - 1) 
     completed_packets++;
881

882
    if(msg->path_type == MINIMAL && msg->chunk_id == num_chunks - 1)
883 884
	minimal_count++;

885
    if(msg->path_type == NON_MINIMAL && msg->chunk_id == num_chunks - 1)
886
	nonmin_count++;
887 888 889
    
    if(msg->path_type != MINIMAL && msg->path_type != NON_MINIMAL)
	printf("\n Wrong message path type %d ", msg->path_type);
890

891
#if DEBUG == 1
892
if(msg->packet_ID == TRACK && msg->chunk_id == num_chunks-1)
893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913
    {
	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);

    }
#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;

  if(msg->chunk_id == num_chunks-1)
  {
	 bf->c2 = 1;
914 915 916 917 918
	 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;

919 920 921 922 923 924 925 926 927 928 929 930
	 N_finished_packets++;
	 dragonfly_total_time += tw_now( lp ) - msg->travel_start_time;

	 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)
	{
931 932
	    if(msg->packet_ID == TRACK)
	    	printf("\n completed at %lf ", tw_now(lp));
933
            void * tmp_ptr = model_net_method_get_edata(DRAGONFLY, msg);
934
            ts = g_tw_lookahead + bytes_to_ns(msg->remote_event_size_bytes, (1/s->params->cn_bandwidth));
935
            if (msg->is_pull){
936 937 938 939
                struct codes_mctx mc_dst =
                    codes_mctx_set_global_direct(msg->sender_mn_lp);
                struct codes_mctx mc_src =
                    codes_mctx_set_global_direct(lp->gid);
940
                int net_id = model_net_get_id(LP_METHOD_NM);
941 942 943
                model_net_event_mctx(net_id, &mc_src, &mc_dst, msg->category,
                        msg->sender_lp, msg->pull_size, ts,
                        msg->remote_event_size_bytes, tmp_ptr, 0, NULL, lp);
944 945 946 947 948 949 950
            }
            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); 
            }
951 952 953
	}
  }

954 955
  tw_stime credit_delay = bytes_to_ns(CREDIT_SIZE, s->params->cn_bandwidth);
  ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, noise);
956
  
957
  msg->saved_credit_time = s->next_credit_available_time;
958
  s->next_credit_available_time = maxd(s->next_credit_available_time, tw_now(lp));
959 960
  s->next_credit_available_time += ts;

961 962
  if(msg->intm_lp_id == TRACK)
	printf("\n terminal sending credit at chan %d ", msg->saved_vc);
963
  //TODO: be annotation-aware
964
  // no method_event here - message going to router
965 966
  
//  printf("\n current time %lf scheduling in time %lf ", tw_now(lp), s->next_credit_available_time - tw_now(lp));
967
  buf_e = tw_event_new(msg->intm_lp_id, s->next_credit_available_time - tw_now(lp), lp);
968
  buf_msg = tw_event_data(buf_e);
969
  buf_msg->magic = router_magic_num;
970 971 972 973 974 975 976 977 978 979 980 981 982 983
  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 )
{
984
//    printf("\n terminal ID %ld ", lp->gid);
985 986 987 988
    uint32_t h1 = 0, h2 = 0; 
    bj_hashlittle2(LP_METHOD_NM, strlen(LP_METHOD_NM), &h1, &h2);
    terminal_magic_num = h1 + h2;
    
989
    int i;
990 991
    char anno[MAX_NAME_LENGTH];

992
    // Assign the global router ID
993
    // TODO: be annotation-aware
994 995 996 997 998 999 1000
    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{
1001
        s->anno = strdup(anno);
1002 1003 1004 1005
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }

1006
   int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
1007
           s->anno, 0);
1008 1009

   s->terminal_id = (mapping_rep_id * num_lps) + mapping_offset;  
Misbah Mubarak's avatar
Misbah Mubarak committed
1010
   s->router_id=(int)s->terminal_id / (s->params->num_routers/2);
1011 1012 1013
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;

1014 1015
   s->vc_occupancy = (int*)malloc(s->params->num_vcs * sizeof(int));
   s->output_vc_state = (int*)malloc(s->params->num_vcs * sizeof(int));
1016
   s->max_term_vc_occupancy = 0;
1017

1018
   for( i = 0; i < s->params->num_vcs; i++ )
1019 1020 1021 1022
    {
      s->vc_occupancy[i]=0;
      s->output_vc_state[i]=VC_IDLE;
    }
1023
//   printf("\n Terminal ID %d Router ID %d ", s->terminal_id, s->router_id);
1024
   dragonfly_collective_init(s, lp);
1025 1026 1027
   return;
}

1028
/* collective operation for the torus network */
1029
void dragonfly_collective(char const * category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
1030 1031 1032 1033 1034 1035 1036
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    tw_lpid local_nic_id;
    char* tmp_ptr;

1037 1038 1039 1040
    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);
1041

1042
    xfer_to_nic_time = codes_local_latency(sender);
1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081
    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;
1082
            ts = (1/s->params->cn_bandwidth) * msg->remote_event_size_bytes;
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 1108 1109 1110
            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 */
1111 1112 1113 1114 1115
            // 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,
1116 1117
                    s->anno, 0);
            codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, s->anno, 0,
1118 1119
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
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 1148 1149

           /* 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++;

1150 1151 1152
        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,
1153
                s->anno, 0);
1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171

        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 */
1172
            codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, s->anno, 0,
1173 1174
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210

           /* 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 */
1211 1212 1213
                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);
1214 1215 1216 1217 1218 1219 1220 1221 1222 1223
                //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){
1224
	                memcpy(m_data, model_net_method_get_edata(DRAGONFLY, msg),
1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241
        	               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;
1242 1243
        int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
                NULL, 1);
1244 1245 1246 1247 1248 1249 1250 1251 1252
        bf->c1 = 0;
        bf->c2 = 0;

        send_remote_event(s, bf, msg, lp);

        if(!s->is_leaf)
        {
            bf->c1 = 1;
            tw_event* e_new;
1253
            terminal_message * msg_new;
1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264
            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 */
1265
                        codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM,
1266
                                s->anno, 0, s->children[i]/num_lps,
1267
                                (s->children[i] % num_lps), &child_nic_id);
1268 1269 1270 1271 1272 1273 1274
                        //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);
1275
		        memcpy(msg_new, msg, sizeof(terminal_message));
1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294
		        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;
          }
}
1295 1296 1297 1298 1299 1300 1301

static void 
terminal_buf_update_rc(terminal_state * s, 
		    tw_bf * bf, 
		    terminal_message * msg, 
		    tw_lp * lp)
{
1302
    term_ecount++;
1303 1304 1305 1306 1307 1308
    int msg_indx = msg->vc_index;
    s->vc_occupancy[msg_indx]++;
    if(s->vc_occupancy[msg_indx] == s->params->cn_vc_size)
       s->output_vc_state[msg_indx] = VC_CREDIT;
}

1309
/* update the compute node-router channel buffer */
1310
static void 
1311 1312 1313 1314 1315
terminal_buf_update(terminal_state * s, 
		    tw_bf * bf, 
		    terminal_message * msg, 
		    tw_lp * lp)
{
1316
     term_ecount++;