dragonfly.c 86.5 KB
Newer Older
Philip Carns's avatar
Philip Carns committed
1 2 3 4 5 6
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

7 8 9 10
// Local router ID: 0 --- total_router-1
// Router LP ID 
// Terminal LP ID

11 12
#include <ross.h>

13
#include "codes/jenkins-hash.h"
14 15 16 17
#include "codes/codes_mapping.h"
#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
#include "sys/file.h"
21
#include "codes/quickhash.h"
22 23 24 25

#define CREDIT_SIZE 8
#define MEAN_PROCESS 1.0

26 27 28
/* collective specific parameters */
#define TREE_DEGREE 4
#define LEVEL_DELAY 1000
29
#define DRAGONFLY_COLLECTIVE_DEBUG 0
30 31 32
#define NUM_COLLECTIVES  1
#define COLLECTIVE_COMPUTATION_DELAY 5700
#define DRAGONFLY_FAN_OUT_DELAY 20.0
33
#define WINDOW_LENGTH 0
34
#define DFLY_HASH_TABLE_SIZE 65536
35

36
// debugging parameters
37 38
#define TRACK -1
#define TRACK_MSG -1
39
#define PRINT_ROUTER_TABLE 1
Misbah Mubarak's avatar
Misbah Mubarak committed
40
#define DEBUG 0
41
#define USE_DIRECT_SCHEME 1
42

43 44 45
#define LP_CONFIG_NM (model_net_lp_config_names[DRAGONFLY])
#define LP_METHOD_NM (model_net_method_names[DRAGONFLY])

46
long term_ecount, router_ecount, term_rev_ecount, router_rev_ecount;
47

48 49
static double maxd(double a, double b) { return a < b ? b : a; }

50
/* minimal and non-minimal packet counts for adaptive routing*/
51
static unsigned int minimal_count=0, nonmin_count=0, completed_packets = 0;
52

53 54 55 56 57 58
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;
59 60

/* global variables for codes mapping */
61
static char lp_group_name[MAX_NAME_LENGTH];
62 63
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

64 65 66 67 68 69
/* router magic number */
int router_magic_num = 0;

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

70 71 72 73 74 75 76
typedef struct terminal_message_list terminal_message_list;
struct terminal_message_list {
    terminal_message msg;
    char* event_data;
    terminal_message_list *next;
    terminal_message_list *prev;
};
77

78 79 80 81 82 83 84
void init_terminal_message_list(terminal_message_list *this, 
    terminal_message *inmsg) {
    this->msg = *inmsg;
    this->event_data = NULL;
    this->next = NULL;
    this->prev = NULL;
}
85

86 87 88 89
void delete_terminal_message_list(terminal_message_list *this) {
    if(this->event_data != NULL) free(this->event_data);
    free(this);
}
90

91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
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;
108
    int total_terminals;
109
    int num_global_channels;
110 111 112 113
    double cn_delay;
    double local_delay;
    double global_delay;
    double credit_delay;
114 115
};

116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
struct dfly_hash_key
{
    uint64_t message_id;
    tw_lpid sender_id;
};

struct dfly_qhash_entry
{
   struct dfly_hash_key key;
   char * remote_event_data;
   int num_chunks;
   int remote_event_size;
   struct qhash_head hash_link;
};

131 132 133 134 135 136 137 138
/* 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
{
139
   uint64_t packet_counter;
140 141

   // Dragonfly specific parameters
142 143
   unsigned int router_id;
   unsigned int terminal_id;
144 145 146

   // Each terminal will have an input and output channel with the router
   int* vc_occupancy; // NUM_VC
147
   int num_vcs;
148 149
   tw_stime terminal_available_time;
   tw_stime next_credit_available_time;
150 151 152
   terminal_message_list **terminal_msgs;
   terminal_message_list **terminal_msgs_tail;
   int in_send_loop;
153 154 155 156
// 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];
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177
  /* 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;

178 179
   struct rc_stack * st;

180 181 182
   /* to maintain a count of child nodes that have fanned in at the parent during the collective
      fan-in phase*/
   int num_fan_nodes;
183 184 185

   const char * anno;
   const dragonfly_param *params;
186

187 188 189
   struct qhash_table *rank_tbl;
   uint64_t rank_tbl_pop;

190 191 192
   tw_stime   total_msg_time;
   long total_msg_size;
   long finished_msgs;
193
   long finished_packets;
194

195
   char output_buf[512];
196
};
197

198 199 200 201 202
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
203
  T_SEND,
204
  T_BUFFER,
205 206
  R_SEND,
  R_ARRIVE,
207 208 209 210
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
};
/* 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
{
234 235
    MINIMAL = 0,
    NON_MINIMAL,
236 237
    ADAPTIVE,
    PROG_ADAPTIVE
238 239 240 241 242 243
};

struct router_state
{
   unsigned int router_id;
   unsigned int group_id;
244 245
  
   int* global_channel; 
246
   
247 248
   tw_stime* next_output_available_time;
   tw_stime* next_credit_available_time;
249
   tw_stime* cur_hist_start_time;
250 251 252 253 254
   terminal_message_list ***pending_msgs;
   terminal_message_list ***pending_msgs_tail;
   terminal_message_list ***queued_msgs;
   terminal_message_list ***queued_msgs_tail;
   int *in_send_loop;
255
   
256 257
   int** vc_occupancy;
   int* link_traffic;
258 259 260

   const char * anno;
   const dragonfly_param *params;
261 262 263

   int* prev_hist_num;
   int* cur_hist_num;
264 265 266 267
};

static short routing = MINIMAL;

268 269
static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
270
static tw_stime         max_collective = 0;
271

272

273 274 275
static long long       total_hops = 0;
static long long       N_finished_packets = 0;

276 277 278 279 280 281 282 283 284 285 286 287 288 289
static int dragonfly_rank_hash_compare(
        void *key, struct qhash_head *link)
{
    struct dfly_hash_key *message_key = (struct dfly_hash_key *)key;
    struct dfly_qhash_entry *tmp;

    tmp = qhash_entry(link, struct dfly_qhash_entry, hash_link);
    
    if (tmp->key.message_id == message_key->message_id
            && tmp->key.sender_id == message_key->sender_id)
        return 1;

    return 0;
}
290 291 292 293 294 295 296 297 298 299 300 301 302 303
/* 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);
}
304

305 306
/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
307
{
308 309
	   return sizeof(terminal_message);
}
310

311 312 313 314 315 316 317 318 319 320 321 322
static void append_to_terminal_message_list(  
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index, 
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thisq[index] = msg;
    } else {
        thistail[index]->next = msg;
        msg->prev = thistail[index];
    } 
    thistail[index] = msg;
323 324
}

325 326 327 328 329 330 331 332 333 334 335 336 337
static void prepend_to_terminal_message_list(  
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index, 
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thistail[index] = msg;
    } else {
        thisq[index]->prev = msg;
        msg->next = thisq[index];
    } 
    thisq[index] = msg;
}
338

339 340 341 342 343 344 345 346 347 348
static void create_prepend_to_terminal_message_list(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index, 
        terminal_message *msg) {
    terminal_message_list* new_entry = (terminal_message_list*)malloc(
        sizeof(terminal_message_list));
    init_terminal_message_list(new_entry, msg);
    if(msg->remote_event_size_bytes) {
        void *m_data = model_net_method_get_edata(DRAGONFLY, msg);
349 350 351
        size_t s = msg->remote_event_size_bytes + msg->local_event_size_bytes;
        new_entry->event_data = (void*)malloc(s);
        memcpy(new_entry->event_data, m_data, s);
352
    }
353
    prepend_to_terminal_message_list( thisq, thistail, index, new_entry);
354 355
}

356 357 358 359 360 361 362 363 364 365 366 367 368 369 370
static terminal_message_list* return_head(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index) {
    terminal_message_list *head = thisq[index];
    if(head != NULL) {
        thisq[index] = head->next;
        if(head->next != NULL) {
            head->next->prev = NULL;
            head->next = NULL;
        } else {
            thistail[index] = NULL;
        }
    }
    return head;
371 372
}

373 374 375 376 377 378 379 380 381 382 383 384 385 386
static terminal_message_list* return_tail(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index) {
    terminal_message_list *tail = thistail[index];
    if(tail->prev != NULL) {
        tail->prev->next = NULL;
        thistail[index] = tail->prev;
        tail->prev = NULL;
    } else {
        thistail[index] = NULL;
        thisq[index] = NULL;
    }
    return tail;
387 388
}

389 390 391 392 393 394 395
static void copy_terminal_list_entry( terminal_message_list *cur_entry,
    terminal_message *msg) {
    terminal_message *cur_msg = &cur_entry->msg;
    msg->travel_start_time = cur_msg->travel_start_time;
    msg->packet_ID = cur_msg->packet_ID;    
    strcpy(msg->category, cur_msg->category);
    msg->final_dest_gid = cur_msg->final_dest_gid;
396
    msg->msg_start_time = msg->msg_start_time;
397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415
    msg->sender_lp = cur_msg->sender_lp;
    msg->dest_terminal_id = cur_msg->dest_terminal_id;
    msg->src_terminal_id = cur_msg->src_terminal_id;
    msg->local_id = cur_msg->local_id;
    msg->origin_router_id = cur_msg->origin_router_id;
    msg->my_N_hop = cur_msg->my_N_hop;
    msg->my_l_hop = cur_msg->my_l_hop;
    msg->my_g_hop = cur_msg->my_g_hop;
    msg->intm_lp_id = cur_msg->intm_lp_id;
    msg->saved_channel = cur_msg->saved_channel;
    msg->saved_vc = cur_msg->saved_vc;
    msg->last_hop = cur_msg->last_hop;
    msg->path_type = cur_msg->path_type;
    msg->vc_index = cur_msg->vc_index;
    msg->output_chan = cur_msg->output_chan;
    msg->is_pull = cur_msg->is_pull;
    msg->pull_size = cur_msg->pull_size;
    msg->intm_group_id = cur_msg->intm_group_id;
    msg->chunk_id = cur_msg->chunk_id;
416 417
    msg->sender_mn_lp = cur_msg->sender_mn_lp;
    msg->total_size = cur_msg->total_size;
418
    msg->packet_size = cur_msg->packet_size;
419
    msg->message_id = cur_msg->message_id;
420 421 422 423 424 425 426 427 428 429 430 431
    msg->local_event_size_bytes = cur_msg->local_event_size_bytes;
    msg->remote_event_size_bytes = cur_msg->remote_event_size_bytes;
    msg->sender_node = cur_msg->sender_node;
    msg->next_stop = cur_msg->next_stop;
    msg->magic = cur_msg->magic;

    if(msg->local_event_size_bytes +  msg->remote_event_size_bytes > 0) {
        void *m_data = model_net_method_get_edata(DRAGONFLY, msg);
        memcpy(m_data, cur_entry->event_data, 
            msg->local_event_size_bytes +  msg->remote_event_size_bytes);
    }
}
432 433 434
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
435

436 437 438 439 440 441 442 443
    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);
    }

444
    p->num_vcs = 3;
445 446

    configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
447
    if(!p->local_vc_size) {
448 449 450 451 452
        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);
453
    if(!p->global_vc_size) {
454 455 456 457 458
        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);
459
    if(!p->cn_vc_size) {
460 461 462 463 464
        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);
465
    if(!p->chunk_size) {
466
        p->chunk_size = 512;
467
        fprintf(stderr, "Chunk size for packets is specified, setting to %d\n", p->chunk_size);
468 469 470
    }

    configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
471
    if(!p->local_bandwidth) {
472 473 474 475 476
        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);
477
    if(!p->global_bandwidth) {
478 479 480 481 482
        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);
483
    if(!p->cn_bandwidth) {
484 485 486 487
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }

488 489
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
490
            MAX_NAME_LENGTH);
491 492
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
493 494
    else if(strcmp(routing_str, "nonminimal")==0 || 
            strcmp(routing_str,"non-minimal")==0)
495 496 497 498 499
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
500 501 502 503
    else
    {
        fprintf(stderr, 
                "No routing protocol specified, setting to minimal routing\n");
504
        routing = -1;
505 506 507 508 509 510
    }

    // 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;
511
    p->radix = (p->num_cn + p->num_global_channels + p->num_routers);
512
    p->total_routers = p->num_groups * p->num_routers;
513
    p->total_terminals = p->total_routers * p->num_cn;
514 515 516 517 518 519 520
    int rank;
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    if(!rank) {
        printf("\n Total nodes %d routers %d groups %d radix %d \n",
                p->num_cn * p->total_routers, p->total_routers, p->num_groups,
                p->radix);
    }
521
    
522 523 524 525
    p->cn_delay = bytes_to_ns(p->chunk_size, p->cn_bandwidth);
    p->local_delay = bytes_to_ns(p->chunk_size, p->local_bandwidth);
    p->global_delay = bytes_to_ns(p->chunk_size, p->global_bandwidth);
    p->credit_delay = bytes_to_ns(8.0, p->local_bandwidth); //assume 8 bytes packet
526 527 528

}

529 530 531 532
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);
533
    all_params = malloc(num_params * sizeof(*all_params));
534 535

    for (uint64_t i = 0; i < anno_map->num_annos; i++){
536
        const char * anno = anno_map->annotations[i].ptr;
537 538 539 540 541
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
542 543 544 545 546 547 548
}

/* 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;
549
   int total_minimal_packets, total_nonmin_packets, total_completed_packets;
550 551 552 553 554

   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);
555
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
556 557 558
    {
	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);
559
 	MPI_Reduce(&completed_packets, &total_completed_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
560
    }
561

562 563
   /* print statistics */
   if(!g_tw_mynode)
564
   {	
565
      printf(" Average number of hops traversed %f average message latency %lf us maximum message latency %lf us avg time %lf \n", (float)avg_hops/total_finished_packets, avg_time/(total_finished_packets*1000), max_time/1000, avg_time);
566
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
567
              printf("\n ADAPTIVE ROUTING STATS: %d percent packets routed minimally %d percent packets routed non-minimally completed packets %d ", total_minimal_packets, total_nonmin_packets, total_completed_packets);
568 569
 
  }
570 571
   return;
}
572

573 574 575
void dragonfly_collective_init(terminal_state * s,
           		   tw_lp * lp)
{
576 577 578 579 580
    // 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);
581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630
    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
}

631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663
/* initialize a dragonfly compute node terminal */
void 
terminal_init( terminal_state * s, 
	       tw_lp * lp )
{
    uint32_t h1 = 0, h2 = 0; 
    bj_hashlittle2(LP_METHOD_NM, strlen(LP_METHOD_NM), &h1, &h2);
    terminal_magic_num = h1 + h2;
    
    int i;
    char anno[MAX_NAME_LENGTH];

    // Assign the global router ID
    // TODO: be annotation-aware
    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{
        s->anno = strdup(anno);
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }

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

   s->terminal_id = (mapping_rep_id * num_lps) + mapping_offset;  
   s->router_id=(int)s->terminal_id / (s->params->num_routers/2);
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;
664
   
665 666 667
   s->finished_msgs = 0;
   s->total_msg_time = 0.0;
   s->total_msg_size = 0;
668

669
   rc_stack_create(&s->st);
670 671 672 673 674 675 676 677
   s->num_vcs = 1;
   s->vc_occupancy = (int*)malloc(s->num_vcs * sizeof(int));

   for( i = 0; i < s->num_vcs; i++ )
    {
      s->vc_occupancy[i]=0;
    }

678 679 680 681 682
   s->rank_tbl = qhash_init(dragonfly_rank_hash_compare, quickhash_64bit_hash, DFLY_HASH_TABLE_SIZE);

   if(!s->rank_tbl)
       tw_error(TW_LOC, "\n Hash table not initialized! ");

683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698
   s->terminal_msgs = 
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
   s->terminal_msgs_tail = 
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
   s->terminal_msgs[0] = NULL;
   s->terminal_msgs_tail[0] = NULL;
   s->in_send_loop = 0;

   dragonfly_collective_init(s, lp);
   return;
}


/* sets up the router virtual channels, global channels, 
 * local channels, compute node channels */
void router_setup(router_state * r, tw_lp * lp)
699
{
700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741
    uint32_t h1 = 0, h2 = 0; 
    bj_hashlittle2(LP_METHOD_NM, strlen(LP_METHOD_NM), &h1, &h2);
    router_magic_num = h1 + h2;
    
    char anno[MAX_NAME_LENGTH];
    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'){
        r->anno = NULL;
        r->params = &all_params[num_params-1];
    } else{
        r->anno = strdup(anno);
        int id = configuration_get_annotation_index(anno, anno_map);
        r->params = &all_params[id];
    }

    // shorthand
    const dragonfly_param *p = r->params;

   r->router_id=mapping_rep_id + mapping_offset;
   r->group_id=r->router_id/p->num_routers;

   r->global_channel = (int*)malloc(p->num_global_channels * sizeof(int));
   r->next_output_available_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
   r->next_credit_available_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
   r->cur_hist_start_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
   r->link_traffic = (int*)malloc(p->radix * sizeof(int));
   r->cur_hist_num = (int*)malloc(p->radix * sizeof(int));
   r->prev_hist_num = (int*)malloc(p->radix * sizeof(int));
   
   r->vc_occupancy = (int**)malloc(p->radix * sizeof(int*));
   r->in_send_loop = (int*)malloc(p->radix * sizeof(int));
   r->pending_msgs = 
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
   r->pending_msgs_tail = 
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
   r->queued_msgs = 
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
   r->queued_msgs_tail = 
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
  
742
   for(int i=0; i < p->radix; i++)
743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761
    {
       // Set credit & router occupancy
	r->next_output_available_time[i]=0;
        r->next_credit_available_time[i]=0;
	r->cur_hist_start_time[i] = 0;
        r->link_traffic[i]=0;
	r->cur_hist_num[i] = 0;
	r->prev_hist_num[i] = 0;
        
        r->in_send_loop[i] = 0;
        r->vc_occupancy[i] = (int*)malloc(p->num_vcs * sizeof(int));
        r->pending_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
            sizeof(terminal_message_list*));
        r->pending_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
            sizeof(terminal_message_list*));
        r->queued_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
            sizeof(terminal_message_list*));
        r->queued_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
            sizeof(terminal_message_list*));
762
        for(int j = 0; j < p->num_vcs; j++) {
763 764 765 766 767 768 769 770 771
            r->vc_occupancy[i][j] = 0;
            r->pending_msgs[i][j] = NULL;
            r->pending_msgs_tail[i][j] = NULL;
            r->queued_msgs[i][j] = NULL;
            r->queued_msgs_tail[i][j] = NULL;
        }
    }

#if DEBUG == 1
772
//   printf("\n LP ID %d VC occupancy radix %d Router %d is connected to ", lp->gid, p->radix, r->router_id);
773 774 775 776
#endif 
   //round the number of global channels to the nearest even number
#if USE_DIRECT_SCHEME
       int first = r->router_id % p->num_routers;
777
       for(int i=0; i < p->num_global_channels; i++)
778 779 780 781 782 783 784 785 786 787 788 789 790
        {
            int target_grp = first;
            if(target_grp == r->group_id) {
                target_grp = p->num_groups - 1;
            }
            int my_pos = r->group_id % p->num_routers;
            if(r->group_id == p->num_groups - 1) {
                my_pos = target_grp % p->num_routers;
            }
            r->global_channel[i] = target_grp * p->num_routers + my_pos;
            first += p->num_routers;
        }
#else
791 792 793
   int router_offset = (r->router_id % p->num_routers) * 
    (p->num_global_channels / 2) + 1;
   for(int i=0; i < p->num_global_channels; i++)
794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821
    {
      if(i % 2 != 0)
          {
             r->global_channel[i]=(r->router_id + (router_offset * p->num_routers))%p->total_routers;
             router_offset++;
          }
          else
           {
             r->global_channel[i]=r->router_id - ((router_offset) * p->num_routers);
           }
        if(r->global_channel[i]<0)
         {
           r->global_channel[i]=p->total_routers+r->global_channel[i]; 
	 }
#if DEBUG == 1
    printf("\n channel %d ", r->global_channel[i]);
#endif 
    }
#endif

#if DEBUG == 1
   printf("\n");
#endif
   return;
}	


/* dragonfly packet event , generates a dragonfly packet on the compute node */
822 823 824 825 826 827 828 829 830 831
static tw_stime dragonfly_packet_event(
        model_net_request const * req,
        uint64_t message_offset,
        uint64_t packet_size,
        tw_stime offset,
        mn_sched_params const * sched_params,
        void const * remote_event,
        void const * self_event,
        tw_lp *sender,
        int is_last_pckt)
832
{
833 834 835 836 837
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

838 839 840
    xfer_to_nic_time = codes_local_latency(sender); 
    //e_new = tw_event_new(sender->gid, xfer_to_nic_time+offset, sender);
    //msg = tw_event_data(e_new);
841 842
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
843 844
    strcpy(msg->category, req->category);
    msg->final_dest_gid = req->final_dest_lp;
845
    msg->total_size = req->msg_size;
846
    msg->sender_lp=req->src_lp;
847
    msg->sender_mn_lp = sender->gid;
848
    msg->packet_size = packet_size;
849
    msg->travel_start_time = tw_now(sender);
850 851 852
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
853
    msg->dest_terminal_id = req->dest_mn_lp;
854
    msg->message_id = req->msg_id;
855 856
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;
857
    msg->magic = terminal_magic_num; 
858 859
    msg->msg_start_time = req->msg_start_time;

860 861
    if(is_last_pckt) /* Its the last packet so pass in remote and local event information*/
      {
862
	if(req->remote_event_size > 0)
863
	 {
864 865 866
		msg->remote_event_size_bytes = req->remote_event_size;
		memcpy(tmp_ptr, remote_event, req->remote_event_size);
		tmp_ptr += req->remote_event_size;
867
	}
868
	if(req->self_event_size > 0)
869
	{
870 871 872
		msg->local_event_size_bytes = req->self_event_size;
		memcpy(tmp_ptr, self_event, req->self_event_size);
		tmp_ptr += req->self_event_size;
873 874
	}
     }
875
	   //printf("\n dragonfly remote event %d local event %d last packet %d %lf ", msg->remote_event_size_bytes, msg->local_event_size_bytes, is_last_pckt, xfer_to_nic_time);
876
    tw_event_send(e_new);
877
    return xfer_to_nic_time;
878 879 880 881 882 883 884 885 886
}

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

887 888 889
/* 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,
890 891
		    int num_routers,
                    int total_groups)
892
{
893 894 895 896 897 898 899
#if USE_DIRECT_SCHEME
  int dest = dest_gid;
  if(dest == total_groups - 1) {
      dest = src_gid;
  }
  return src_gid * num_routers + (dest % num_routers);
#else
900 901 902
  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;
903
  
904 905
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
906
  
907 908
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
909
  
910
  offset=(offset - 1) % (half_channel * num_routers);
911 912 913 914 915 916 917 918 919 920

  // 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;
921
#endif
922 923 924
}	

/*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*/
925 926
void router_credit_send(router_state * s, tw_bf * bf, terminal_message * msg, 
  tw_lp * lp, int sq) {
927 928 929 930
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

931
  int dest = 0,  type = R_BUFFER;
932
  int is_terminal = 0;
933

934
  const dragonfly_param *p = s->params;
935 936 937 938 939 940 941 942 943 944 945 946 947
 
  // Notify sender terminal about available buffer space
  if(msg->last_hop == TERMINAL) {
    dest = msg->src_terminal_id;
    type = T_BUFFER;
    is_terminal = 1;
  } else if(msg->last_hop == GLOBAL) {
    dest = msg->intm_lp_id;
  } else if(msg->last_hop == LOCAL) {
    dest = msg->intm_lp_id;
  } else
    printf("\n Invalid message type");

948
  ts = g_tw_lookahead + p->credit_delay +  tw_rand_unif(lp->rng);
949
	
950 951 952 953 954 955 956 957 958 959 960 961 962 963
  if (is_terminal) {
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY, 
      (void**)&buf_msg, NULL);
    buf_msg->magic = terminal_magic_num;
  } else {
    buf_e = tw_event_new(dest, ts , lp);
    buf_msg = tw_event_data(buf_e);
    buf_msg->magic = router_magic_num;
  }
 
  if(sq == -1) {
    buf_msg->vc_index = msg->vc_index;
    buf_msg->output_chan = msg->output_chan;
  } else {
964
    buf_msg->vc_index = msg->saved_vc;
965 966 967 968
    buf_msg->output_chan = msg->saved_channel;
  }
  
  buf_msg->type = type;
969

970 971
  tw_event_send(buf_e);
  return;
972 973
}

974
void packet_generate_rc(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
975
{
976 977 978
   term_rev_ecount++;
   term_ecount--;

979

980 981 982
   int num_chunks = msg->packet_size/s->params->chunk_size;
   if(msg->packet_size % s->params->chunk_size)
       num_chunks++;
983

984
   if(!num_chunks)
985
       num_chunks = 1;
986

987 988 989 990 991 992
   int i;
   for(i = 0; i < num_chunks; i++) {
        delete_terminal_message_list(return_tail(s->terminal_msgs, 
          s->terminal_msgs_tail, 0));
   }
    if(bf->c5) {
993
        codes_local_latency_reverse(lp);
994 995 996 997
        s->in_send_loop = 0;
    }
     struct mn_stats* stat;
     stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
998 999 1000 1001
     stat->send_count--;
     stat->send_bytes -= msg->packet_size;
     stat->send_time -= (1/s->params->cn_bandwidth) * msg->packet_size;
}
1002

1003
/* generates packet at the current dragonfly compute node */
1004 1005 1006
void packet_generate(terminal_state * s, tw_bf * bf, terminal_message * msg, 
  tw_lp * lp) {
  term_ecount++;
1007

1008
  tw_stime ts;
1009

1010
  assert(lp->gid != msg->dest_terminal_id);
1011
  const dragonfly_param *p = s->params;
1012

1013

1014 1015
  int i, total_event_size;
  int num_chunks = msg->packet_size / p->chunk_size;
1016 1017
  if (msg->packet_size % s->params->chunk_size) 
      num_chunks++;
1018 1019 1020 1021

  if(!num_chunks)
    num_chunks = 1;

1022
  msg->packet_ID = lp->gid + g_tw_nlp * s->packet_counter;
1023 1024 1025 1026
  msg->my_N_hop = 0;
  msg->my_l_hop = 0;
  msg->my_g_hop = 0;
  msg->intm_group_id = -1;
1027

1028 1029 1030
  if(msg->packet_ID == TRACK && msg->message_id == TRACK_MSG)
      printf("\n Packet generated at terminal %d destination %d ", lp->gid, s->router_id);

1031 1032 1033 1034 1035
  for(i = 0; i < num_chunks; i++)
  {
    terminal_message_list *cur_chunk = (terminal_message_list*)malloc(
      sizeof(terminal_message_list));
    init_terminal_message_list(cur_chunk, msg);
1036

1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050
    if(msg->remote_event_size_bytes + msg->local_event_size_bytes > 0) {
      cur_chunk->event_data = (char*)malloc(
          msg->remote_event_size_bytes + msg->local_event_size_bytes);
    }
    
    void * m_data_src = model_net_method_get_edata(DRAGONFLY, msg);
    if (msg->remote_event_size_bytes){
      memcpy(cur_chunk->event_data, m_data_src, msg->remote_event_size_bytes);
    }
    if (msg->local_event_size_bytes){ 
      m_data_src = (char*)m_data_src + msg->remote_event_size_bytes;
      memcpy((char*)cur_chunk->event_data + msg->remote_event_size_bytes, 
          m_data_src, msg->local_event_size_bytes);
    }
1051

1052 1053 1054 1055
    cur_chunk->msg.chunk_id = i;
    append_to_terminal_message_list(s->terminal_msgs, s->terminal_msgs_tail,
      0, cur_chunk);
  }
1056

1057
  
1058 1059
  if(s->in_send_loop == 0) {
    bf->c5 = 1;
1060
    ts = codes_local_latency(lp);
1061 1062 1063 1064 1065 1066 1067 1068
    terminal_message *m;
    tw_event* e = model_net_method_event_new(lp->gid, ts, lp, DRAGONFLY, 
      (void**)&m, NULL);
    m->type = T_SEND;
    m->magic = terminal_magic_num;
    s->in_send_loop = 1;
    tw_event_send(e);
  }
1069

1070 1071 1072 1073 1074 1075 1076 1077
  total_event_size = model_net_get_msg_sz(DRAGONFLY) + 
      msg->remote_event_size_bytes + msg->local_event_size_bytes;
  mn_stats* stat;
  stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
  stat->send_count++;
  stat->send_bytes += msg->packet_size;
  stat->send_time += (1/p->cn_bandwidth) * msg->packet_size;
  if(stat->max_event_size < total_event_size)
1078
	  stat->max_event_size = total_event_size;