dragonfly.c 70.7 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
#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

159 160 161 162 163 164 165 166 167
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
  T_SEND,
  T_BUFFER,
  R_SEND,
  R_ARRIVE,
168 169 170 171
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194
};
/* 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
{
195 196
    MINIMAL = 0,
    NON_MINIMAL,
197 198
    ADAPTIVE,
    PROG_ADAPTIVE
199 200 201 202 203 204 205 206
};

struct router_state
{
   unsigned int router_id;
   unsigned int group_id;
  
   int* global_channel; 
207
   
208 209
   tw_stime* next_output_available_time;
   tw_stime* next_credit_available_time;
210 211
   tw_stime* cur_hist_start_time;
   
212 213
   int* vc_occupancy;
   int* output_vc_state;
214
   int max_router_vc_occupancy;
215 216 217

   const char * anno;
   const dragonfly_param *params;
218 219 220

   int* prev_hist_num;
   int* cur_hist_num;
221 222 223 224 225 226
};

static short routing = MINIMAL;

static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
227
static tw_stime         max_collective = 0;
228 229 230 231 232 233 234 235 236 237 238 239 240 241


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);
}

242 243 244
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
245

246 247 248 249 250 251 252 253 254 255
    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);
256
    if(!p->num_vcs) {
257 258 259 260 261
        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
262
    if(p->local_vc_size <= 0) {
263 264 265 266 267
        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
268
    if(p->global_vc_size <= 0) {
269 270 271 272 273
        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
274
    if(p->cn_vc_size <= 0) {
275 276 277 278 279
        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
280
    if(p->chunk_size <= 0) {
281
        p->chunk_size = 64;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
282
        fprintf(stderr, "Chunk size for packets is not specified, setting to %d\n", p->chunk_size);
283 284 285
    }

    configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
286
    if(p->local_bandwidth <= 0) {
287 288 289 290 291
        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
292
    if(p->global_bandwidth <= 0) {
293 294 295 296 297
        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
298
    if(p->cn_bandwidth <= 0) {
299 300 301 302 303
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }


304

305 306
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
307
            MAX_NAME_LENGTH);
308 309 310 311 312 313 314 315
    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;
316 317 318 319
    else
    {
        fprintf(stderr, 
                "No routing protocol specified, setting to minimal routing\n");
320
        routing = -1;
321 322 323 324 325 326 327 328 329
    }

    // 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;
330 331


332
    p->total_terminals = p->total_routers * p->num_cn;
333
    printf("\n Total nodes %d routers %d groups %d radix %d num_vc %d ", p->num_cn * p->total_routers,
334 335
								p->total_routers,
								p->num_groups,
336 337
								p->radix,
                                                                p->num_vcs);
338 339
}

340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355
/* 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 */
356 357 358 359
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
360
    all_params = calloc(num_params, sizeof(*all_params));
361 362

    for (uint64_t i = 0; i < anno_map->num_annos; i++){
363
        const char * anno = anno_map->annotations[i].ptr;
364 365 366 367 368
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
369 370 371 372 373 374 375
}

/* 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;
376
   long total_term_events, total_router_events;
377
   int total_minimal_packets, total_nonmin_packets, total_completed_packets;
378 379 380 381 382

   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);
383 384 385
   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);

386
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
387 388 389
    {
	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);
390
 	MPI_Reduce(&completed_packets, &total_completed_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
391
    }
392 393
   /* print statistics */
   if(!g_tw_mynode)
394
   {	
395 396
      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);
397
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
398
              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);
399
 
400 401
     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);
402
  }
403 404
   return;
}
405

406 407 408
void dragonfly_collective_init(terminal_state * s,
           		   tw_lp * lp)
{
409 410 411 412 413
    // 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);
414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
    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
}

464
/* dragonfly packet event , generates a dragonfly packet on the compute node */
465
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)
466 467 468 469 470 471
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

472
    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*/
473 474
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
475 476
    strcpy(msg->category, category);
    msg->final_dest_gid = final_dest_lp;
477
    msg->dest_terminal_id = dest_mn_lp;
478
    msg->sender_lp=src_lp;
479
    msg->sender_mn_lp = sender->gid;
480 481 482 483
    msg->packet_size = packet_size;
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
484
    msg->magic = terminal_magic_num;
485 486
    msg->is_pull = is_pull;
    msg->pull_size = pull_size;
487
    msg->chunk_id = 0;
488
    msg->packet_ID = 0;
489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505

    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);
506
    return xfer_to_nic_time;
507 508 509 510 511 512 513 514 515
}

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

516 517 518 519
/* 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)
520
{
521 522 523
  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;
524
  
525 526
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
527
  
528 529
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
530
  
531
  offset=(offset - 1) % (half_channel * num_routers);
532 533 534 535 536 537 538 539 540 541 542 543 544 545 546

  // 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)
{
547
  router_ecount++;
548 549 550 551
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

552
  int dest=0, type = R_BUFFER;
553
  int is_terminal = 0;
554
  int found_magic = router_magic_num;
555
  tw_stime credit_delay = 0.0;
556

557
  const dragonfly_param *p = s->params;
558
  int sender_radix;
559 560 561 562
 // Notify sender terminal about available buffer space
  if(msg->last_hop == TERMINAL)
  {
   dest = msg->src_terminal_id;
563
   sender_radix = msg->local_id % p->num_cn;  
564
   //determine the time in ns to transfer the credit
565
   credit_delay = bytes_to_ns(CREDIT_SIZE, p->cn_bandwidth);
566
   type = T_BUFFER;
567
   is_terminal = 1;
568
   found_magic = terminal_magic_num;
569 570 571 572
  }
   else if(msg->last_hop == GLOBAL)
   {
     dest = msg->intm_lp_id;
573
     sender_radix = p->num_cn + (msg->local_id % p->num_global_channels);
574
     credit_delay = bytes_to_ns(CREDIT_SIZE, p->global_bandwidth);
575 576 577 578
   }
    else if(msg->last_hop == LOCAL)
     {
        dest = msg->intm_lp_id;
579
        sender_radix = p->num_cn + p->num_global_channels + (msg->local_id % p->num_routers);
580
     	credit_delay = bytes_to_ns(CREDIT_SIZE, p->local_bandwidth) * CREDIT_SIZE;
581 582 583 584 585
     }
    else
      printf("\n Invalid message type");

   // Assume it takes 0.1 ns of serialization latency for processing the credits in the queue
586
    assert(sender_radix < s->params->radix );
587
    int output_port = msg->saved_vc / p->num_vcs;
588 589 590

    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]);
591
    ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, (double)credit_delay/1000);
592
	
593
    s->next_credit_available_time[sender_radix]+=ts;
594 595
    if (is_terminal){
        buf_e = model_net_method_event_new(dest, 
596
                s->next_credit_available_time[sender_radix] - tw_now(lp), lp,
597 598 599
                DRAGONFLY, (void**)&buf_msg, NULL);
    }
    else{
600
        buf_e = tw_event_new(dest, s->next_credit_available_time[sender_radix] - tw_now(lp) , lp);
601 602
        buf_msg = tw_event_data(buf_e);
    }
603
    buf_msg->origin_router_id = s->router_id;
604 605
    buf_msg->vc_index = msg->saved_vc;
    buf_msg->type=type;
606
    buf_msg->magic = found_magic;
607 608 609 610 611 612 613 614
    buf_msg->last_hop = msg->last_hop;
    buf_msg->packet_ID=msg->packet_ID;

    tw_event_send(buf_e);

    return;
}

615
static void packet_generate_send_rc(terminal_state * s, 
616 617 618
			    tw_bf * bf, 
			    terminal_message * msg, 
			    tw_lp * lp)
619
{
620 621
    term_pending_ecount++;

622 623 624
    int i;
    tw_rand_reverse_unif(lp->rng);
	 
625 626 627 628 629 630 631 632 633 634
   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;

   if (msg->chunk_id == (msg->num_chunks-1)){
     codes_local_latency_reverse(lp);
   }
635 636 637 638

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

639
     struct mn_stats* stat;
640 641 642 643 644
     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;
}
645

646
/* generates packet at the current dragonfly compute node */
647
static void packet_generate_send(terminal_state * s, 
648 649 650 651
			    tw_bf * bf, 
			    terminal_message * msg, 
			    tw_lp * lp)
{
652 653 654
	term_ecount++;
	term_pending_ecount++;

655
        const dragonfly_param *p = s->params;
656 657 658

	tw_stime ts;
	tw_event *e;
659
        tw_lpid router_id;
660 661 662 663 664 665
	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++;
666

667 668
	if(!num_chunks)
	   num_chunks = 1;
669

670
	msg->num_chunks = num_chunks;
671 672 673
	
	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);
674

675 676
	//  Each packet is broken into chunks and then sent over the channel
	msg->saved_available_time = s->terminal_available_time;
677 678 679
	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));
680 681
	s->terminal_available_time = maxd(s->terminal_available_time, tw_now(lp));
	s->terminal_available_time += ts;
682 683 684

        if(msg->chunk_id == 0)
            msg->travel_start_time = tw_now(lp);
685
	
686 687 688
	int chan = -1, j;
	for(j = 0; j < p->num_vcs; j++)
	 {
689
	     if(s->vc_occupancy[j] < p->cn_vc_size)
690 691 692 693 694
	      {
	       chan=j;
	       break;
	      }
	 }
695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724
	  
	/* 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));
	m->magic = router_magic_num;
	m->origin_router_id = s->router_id;
        m->type = R_ARRIVE;
        m->src_terminal_id = lp->gid;
        m->chunk_id = msg->chunk_id;
   	m->saved_vc = chan;
        m->last_hop = TERMINAL;
        m->intm_group_id = -1;
        m->path_type = -1;
        m->local_event_size_bytes = 0;
        m->local_id = s->terminal_id;
725

726 727 728 729 730 731
	 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);
732 733 734 735 736 737 738 739

	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) 
740
	{
741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759
		// 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);
		}
760
	}
761 762 763
	s->packet_counter++;
	s->vc_occupancy[chan]++;

764 765 766
	if(s->vc_occupancy[chan] > s->max_term_vc_occupancy)
		s->max_term_vc_occupancy = s->vc_occupancy[chan];

767 768 769 770 771 772 773 774 775 776
	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;
777
	stat->send_time += (s->terminal_available_time - tw_now(lp));
778 779
	if(stat->max_event_size < total_event_size)
	  stat->max_event_size = total_event_size;
780 781

        /* Now schedule another packet generate event */
782
	if(msg->chunk_id < num_chunks - 1)
783
	{
784
	     bf->c1 = 1;
785 786 787 788 789 790
	     /* 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 */
791
	     ts = s->terminal_available_time - tw_now(lp) + codes_local_latency(lp);
792 793 794 795 796
	     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));

797
	     m_gen->chunk_id = ++(msg->chunk_id);
798 799 800 801 802 803 804 805 806 807 808 809 810
             m_gen->type = T_GENERATE;

	      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); 
       }
811
	
812 813 814
  return;
}

815 816 817 818 819
static void packet_arrive_rc(terminal_state * s, 
		   tw_bf * bf, 
                   terminal_message * msg, 
                   tw_lp * lp)
{
820
    term_ecount++; 
821 822 823 824 825 826
    uint64_t num_chunks = msg->packet_size / s->params->chunk_size;
    if (msg->packet_size % s->params->chunk_size)
        num_chunks++;
   
    if(msg->chunk_id == num_chunks - 1)
      completed_packets--;
827

828
   if(msg->path_type == MINIMAL && msg->chunk_id == num_chunks - 1)
829 830
	minimal_count--;

831
   if(msg->path_type == NON_MINIMAL && msg->chunk_id == num_chunks - 1)
832 833 834 835
	nonmin_count--;

   tw_rand_reverse_unif(lp->rng);
   s->next_credit_available_time = msg->saved_credit_time;
836
   if(msg->chunk_id == num_chunks-1)
837 838 839 840 841 842 843 844 845 846 847 848 849 850
   {
    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;
   }
    
851
   if (msg->chunk_id == (msg->num_chunks)-1 && 
852 853 854 855 856 857 858 859
		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);
    }

}
860
/* packet arrives at the destination terminal */
861 862 863 864
static void packet_arrive(terminal_state * s, 
		   tw_bf * bf, 
                   terminal_message * msg, 
                   tw_lp * lp)
865
{
866
    ++term_ecount;
867 868 869
    uint64_t num_chunks = msg->packet_size / s->params->chunk_size;
    if (msg->packet_size % s->params->chunk_size)
        num_chunks++;
870

871 872
   if(msg->chunk_id == num_chunks - 1) 
     completed_packets++;
873

874
    if(msg->path_type == MINIMAL && msg->chunk_id == num_chunks - 1)
875 876
	minimal_count++;

877
    if(msg->path_type == NON_MINIMAL && msg->chunk_id == num_chunks - 1)
878
	nonmin_count++;
879 880 881
    
    if(msg->path_type != MINIMAL && msg->path_type != NON_MINIMAL)
	printf("\n Wrong message path type %d ", msg->path_type);
882
#if DEBUG == 1
883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904
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);

    }
#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;
905 906 907 908 909
	 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;

910 911 912 913 914 915 916 917 918 919 920 921
	 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)
	{
922 923
	    if(msg->packet_ID == TRACK)
	    	printf("\n completed at %lf ", tw_now(lp));
924
            void * tmp_ptr = model_net_method_get_edata(DRAGONFLY, msg);
925
            ts = g_tw_lookahead + bytes_to_ns(msg->remote_event_size_bytes, (1/s->params->cn_bandwidth));
926
            if (msg->is_pull){
927 928 929 930
                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);
931
                int net_id = model_net_get_id(LP_METHOD_NM);
932 933 934
                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);
935 936 937 938 939 940 941
            }
            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); 
            }
942 943 944
	}
  }

945 946
  tw_stime credit_delay = bytes_to_ns(CREDIT_SIZE, s->params->cn_bandwidth);
  ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, noise);
947
  
948
  msg->saved_credit_time = s->next_credit_available_time;
949
  s->next_credit_available_time = maxd(s->next_credit_available_time, tw_now(lp));
950 951
  s->next_credit_available_time += ts;

952 953
  if(msg->intm_lp_id == TRACK)
	printf("\n terminal sending credit at chan %d ", msg->saved_vc);
954
  //TODO: be annotation-aware
955
  // no method_event here - message going to router
956 957
  
//  printf("\n current time %lf scheduling in time %lf ", tw_now(lp), s->next_credit_available_time - tw_now(lp));
958
  buf_e = tw_event_new(msg->intm_lp_id, s->next_credit_available_time - tw_now(lp), lp);
959
  buf_msg = tw_event_data(buf_e);
960
  buf_msg->magic = router_magic_num;
961 962 963 964 965 966 967 968 969 970 971 972 973 974
  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 )
{
975
//    printf("\n terminal ID %ld ", lp->gid);
976 977 978 979
    uint32_t h1 = 0, h2 = 0; 
    bj_hashlittle2(LP_METHOD_NM, strlen(LP_METHOD_NM), &h1, &h2);
    terminal_magic_num = h1 + h2;
    
980
    int i;
981 982
    char anno[MAX_NAME_LENGTH];

983
    // Assign the global router ID
984
    // TODO: be annotation-aware
985 986 987 988 989 990 991
    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{
992
        s->anno = strdup(anno);
993 994 995 996
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }

997
   int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
998
           s->anno, 0);
999 1000

   s->terminal_id = (mapping_rep_id * num_lps) + mapping_offset;  
Misbah Mubarak's avatar
Misbah Mubarak committed
1001
   s->router_id=(int)s->terminal_id / (s->params->num_routers/2);
1002 1003 1004
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;

1005 1006
   s->vc_occupancy = (int*)malloc(s->params->num_vcs * sizeof(int));
   s->output_vc_state = (int*)malloc(s->params->num_vcs * sizeof(int));
1007
   s->max_term_vc_occupancy = 0;
1008

1009
   for( i = 0; i < s->params->num_vcs; i++ )
1010 1011 1012 1013
    {
      s->vc_occupancy[i]=0;
      s->output_vc_state[i]=VC_IDLE;
    }
1014
//   printf("\n Terminal ID %d Router ID %d ", s->terminal_id, s->router_id);
1015
   dragonfly_collective_init(s, lp);
1016 1017 1018
   return;
}

1019
/* collective operation for the torus network */
1020
void dragonfly_collective(char const * category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
1021 1022 1023 1024 1025 1026 1027
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    tw_lpid local_nic_id;
    char* tmp_ptr;

1028 1029 1030 1031
    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);
1032

1033
    xfer_to_nic_time = codes_local_latency(sender);
1034 1035 1036 1037 1038 1039 1040 1041 1042 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
    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;
1073
            ts = (1/s->params->cn_bandwidth) * msg->remote_event_size_bytes;
1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101
            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 */
1102 1103 1104 1105 1106
            // 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,
1107 1108
                    s->anno, 0);
            codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, s->anno, 0,
1109 1110
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
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

           /* 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++;

1141 1142 1143
        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,
1144
                s->anno, 0);
1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162

        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 */
1163
            codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, s->anno, 0,
1164 1165
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
1166 1167 1168 1169 1170 1171 1172 1173 1174 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

           /* 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 */
1202 1203 1204
                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);
1205 1206 1207 1208 1209 1210 1211 1212 1213 1214
                //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){
1215
	                memcpy(m_data, model_net_method_get_edata(DRAGONFLY, msg),
1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232
        	               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;
1233 1234
        int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
                NULL, 1);
1235 1236 1237 1238 1239 1240 1241 1242 1243
        bf->c1 = 0;
        bf->c2 = 0;

        send_remote_event(s, bf, msg, lp);

        if(!s->is_leaf)
        {
            bf->c1 = 1;
            tw_event* e_new;
1244
            terminal_message * msg_new;
1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255
            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 */
1256
                        codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM,
1257
                                s->anno, 0, s->children[i]/num_lps,
1258
                                (s->children[i] % num_lps), &child_nic_id);
1259 1260 1261 1262 1263 1264 1265
                        //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);
1266
		        memcpy(msg_new, msg, sizeof(terminal_message));
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285
		        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;
          }
}
1286 1287 1288 1289 1290 1291 1292

static void 
terminal_buf_update_rc(terminal_state * s, 
		    tw_bf * bf, 
		    terminal_message * msg, 
		    tw_lp * lp)
{
1293
    term_ecount++;
1294 1295 1296 1297 1298 1299
    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;
}

1300
/* update the compute node-router channel buffer */
1301
static void 
1302 1303 1304 1305 1306
terminal_buf_update(terminal_state * s, 
		    tw_bf * bf, 
		    terminal_message * msg, 
		    tw_lp * lp)
{
1307
     term_ecount++;
1308 1309 1310 1311
  // 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;
1312 1313 1314 1315 1316 1317 1318
    
    if(s->vc_occupancy[msg_indx] < 0)
    {
        char buf[64];
        int written = sprintf(buf,
                "terminal %d: error vc occupancy \n",
                lp->gid);
1319

1320
        lp_io_write(lp->gid, "dragonfly-vc-errors", written, buf);
1321
    }
1322 1323 1324 1325 1326 1327 1328 1329 1330
    return;
}

void 
terminal_event( terminal_state * s, 
		tw_bf * bf, 
		terminal_message * msg, 
		tw_lp * lp )
{
1331
  assert(msg->magic == terminal_magic_num);
1332 1333 1334 1335
  *(int *)bf = (int)0;
  switch(msg->type)
    {
    case T_GENERATE:
1336
       packet_generate_send(s,bf,msg,lp);
1337 1338 1339 1340 1341 1342 1343 1344 1345
    break;
    
    case T_ARRIVE:
        packet_arrive(s,bf,msg,lp);
    break;
    
    case T_BUFFER:
       terminal_buf_update(s, bf, msg, lp);
     break;
1346 1347 1348 1349
    
    case D_COLLECTIVE_INIT:
      node_collective_init(s, bf, msg, lp);
    break;
1350

1351 1352 1353 1354 1355 1356 1357 1358
    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;
    
1359 1360 1361 1362 1363 1364
    default:
       printf("\n LP %d Terminal message type not supported %d ", (int)lp->gid, msg->type);
    }
}

void