dragonfly.c 94.6 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
#include <ross.h>

13
#define DEBUG_LP 892
14
#include "codes/jenkins-hash.h"
15 16 17 18
#include "codes/codes_mapping.h"
#include "codes/codes.h"
#include "codes/model-net.h"
#include "codes/model-net-method.h"
19 20
#include "codes/model-net-lp.h"
#include "codes/net/dragonfly.h"
21
#include "sys/file.h"
22
#include "codes/quickhash.h"
23
#include "codes/rc-stack.h"
24 25 26 27

#define CREDIT_SIZE 8
#define MEAN_PROCESS 1.0

28 29 30
/* collective specific parameters */
#define TREE_DEGREE 4
#define LEVEL_DELAY 1000
31
#define DRAGONFLY_COLLECTIVE_DEBUG 0
32 33 34
#define NUM_COLLECTIVES  1
#define COLLECTIVE_COMPUTATION_DELAY 5700
#define DRAGONFLY_FAN_OUT_DELAY 20.0
35
#define WINDOW_LENGTH 0
36
#define DFLY_HASH_TABLE_SIZE 262144
37

38
// debugging parameters
39 40
#define TRACK -1
#define TRACK_PKT -1
41
#define TRACK_MSG -1
42
#define PRINT_ROUTER_TABLE 1
Misbah Mubarak's avatar
Misbah Mubarak committed
43
#define DEBUG 0
44
#define USE_DIRECT_SCHEME 1
45
#define MAX_STATS 65536
46

47 48 49
#define LP_CONFIG_NM (model_net_lp_config_names[DRAGONFLY])
#define LP_METHOD_NM (model_net_method_names[DRAGONFLY])

50
int debug_slot_count = 0;
51
long term_ecount, router_ecount, term_rev_ecount, router_rev_ecount;
52
long packet_gen = 0, packet_fin = 0;
53

54 55
static double maxd(double a, double b) { return a < b ? b : a; }

56
/* minimal and non-minimal packet counts for adaptive routing*/
57
static int minimal_count=0, nonmin_count=0;
58

59 60 61 62 63 64
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;
65 66

/* global variables for codes mapping */
67
static char lp_group_name[MAX_NAME_LENGTH];
68 69
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

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

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

76 77
FILE * dragonfly_log = NULL;

78 79
int sample_bytes_written = 0;

80 81 82 83 84 85 86
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;
};
87

88 89 90 91 92 93 94
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;
}
95

96 97 98 99
void delete_terminal_message_list(terminal_message_list *this) {
    if(this->event_data != NULL) free(this->event_data);
    free(this);
}
100

101 102 103 104 105 106 107 108 109 110 111 112 113 114
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;
115
    int num_groups;
116 117
    int radix;
    int total_routers;
118
    int total_terminals;
119
    int num_global_channels;
120 121 122 123
    double cn_delay;
    double local_delay;
    double global_delay;
    double credit_delay;
124
    double router_delay;
125 126
};

127 128 129 130 131 132
struct dfly_hash_key
{
    uint64_t message_id;
    tw_lpid sender_id;
};

133 134 135 136 137 138 139 140 141 142 143
struct dfly_sample_stats
{
   tw_lpid terminal_id;
   long fin_chunks_sample;
   long data_size_sample;
   double fin_hops_sample;
   tw_stime fin_chunks_time;
   tw_stime busy_time_sample;
   tw_stime end_time;
};

144 145 146 147 148 149 150 151 152
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;
};

153 154 155 156 157 158 159 160
/* 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
{
161
   uint64_t packet_counter;
162

163 164 165
   int packet_gen;
   int packet_fin;

166
   // Dragonfly specific parameters
167 168
   unsigned int router_id;
   unsigned int terminal_id;
169 170 171

   // Each terminal will have an input and output channel with the router
   int* vc_occupancy; // NUM_VC
172
   int num_vcs;
173
   tw_stime terminal_available_time;
174 175 176
   terminal_message_list **terminal_msgs;
   terminal_message_list **terminal_msgs_tail;
   int in_send_loop;
177 178 179 180
// 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];
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
  /* 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;

202
   struct rc_stack * st;
203 204
   int issueIdle;
   int terminal_length;
205

206 207 208
   /* to maintain a count of child nodes that have fanned in at the parent during the collective
      fan-in phase*/
   int num_fan_nodes;
209 210 211

   const char * anno;
   const dragonfly_param *params;
212

213 214 215
   struct qhash_table *rank_tbl;
   uint64_t rank_tbl_pop;

Misbah Mubarak's avatar
Misbah Mubarak committed
216
   tw_stime   total_time;
217
   long total_msg_size;
218
   double total_hops;
219
   long finished_msgs;
220
   long finished_chunks;
221
   long finished_packets;
222

223 224
   tw_stime last_buf_full;
   tw_stime busy_time;
225
   char output_buf[4096];
226 227
   /* For LP suspend functionality */
   int error_ct;
228 229 230 231 232 233 234 235 236 237 238 239

   /* For sampling */
   long fin_chunks_sample;
   long data_size_sample;
   double fin_hops_sample;
   tw_stime fin_chunks_time;
   tw_stime busy_time_sample;

   char sample_buf[4096];
   struct dfly_sample_stats * sample_stat;
   int op_arr_size;
   int max_arr_size;
240
};
241

242 243 244 245 246
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
247
  T_SEND,
248
  T_BUFFER,
249 250
  R_SEND,
  R_ARRIVE,
251 252 253 254
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
};
/* 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
{
278 279
    MINIMAL = 0,
    NON_MINIMAL,
280 281
    ADAPTIVE,
    PROG_ADAPTIVE
282 283 284 285 286
};

struct router_state
{
   unsigned int router_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
287
   int group_id;
288 289
  
   int* global_channel; 
290
   
291
   tw_stime* next_output_available_time;
292
   tw_stime* cur_hist_start_time;
293 294 295
   tw_stime* last_buf_full;
   tw_stime* busy_time;

296 297 298 299 300
   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;
301
   int *queued_count;
302
   struct rc_stack * st;
303
   
304
   int** vc_occupancy;
305
   int64_t* link_traffic;
306 307 308

   const char * anno;
   const dragonfly_param *params;
309 310 311

   int* prev_hist_num;
   int* cur_hist_num;
312
   
313
   char output_buf[4096];
314
   char output_buf2[4096];
315 316 317 318
};

static short routing = MINIMAL;

319 320
static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
321
static tw_stime         max_collective = 0;
322

323

324 325
static long long       total_hops = 0;
static long long       N_finished_packets = 0;
326 327 328
static long long       total_msg_sz = 0;
static long long       N_finished_msgs = 0;
static long long       N_finished_chunks = 0;
329

330 331 332 333
static int dragonfly_rank_hash_compare(
        void *key, struct qhash_head *link)
{
    struct dfly_hash_key *message_key = (struct dfly_hash_key *)key;
334
    struct dfly_qhash_entry *tmp = NULL;
335 336 337 338 339 340 341 342 343

    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;
}
344 345
static int dragonfly_hash_func(void *k, int table_size)
{
346
    struct dfly_hash_key *tmp = (struct dfly_hash_key *)k;
347 348 349
    uint32_t pc = 0, pb = 0;	
    bj_hashlittle2(tmp, sizeof(*tmp), &pc, &pb);
    /*uint64_t key = (~tmp->message_id) + (tmp->message_id << 18);
350 351 352
    key = key * 21;
    key = ~key ^ (tmp->sender_id >> 4);
    key = key * tmp->sender_id; 
353 354
    return (int)(key & (table_size - 1));*/
    return (int)(pc % (table_size - 1));
355 356
}

357 358 359 360 361 362 363 364 365 366 367 368 369 370
/* 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);
}
371

372 373
/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
374
{
375 376
	   return sizeof(terminal_message);
}
377

378 379 380 381 382 383
static void free_tmp(void * ptr)
{
    struct dfly_qhash_entry * dfly = ptr; 
    free(dfly->remote_event_data);
    free(dfly);
}
384 385 386 387 388 389 390 391 392 393 394 395
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;
396 397
}

398 399 400 401 402 403 404 405 406 407 408 409 410
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;
}
411

412 413 414 415 416 417 418 419 420 421 422 423 424 425 426
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;
427 428
}

429 430 431 432 433
static terminal_message_list* return_tail(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index) {
    terminal_message_list *tail = thistail[index];
434
    assert(tail);
435 436 437 438 439 440 441 442 443
    if(tail->prev != NULL) {
        tail->prev->next = NULL;
        thistail[index] = tail->prev;
        tail->prev = NULL;
    } else {
        thistail[index] = NULL;
        thisq[index] = NULL;
    }
    return tail;
444 445
}

446 447 448
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
449

450
    int rc = configuration_get_value_int(&config, "PARAMS", "num_routers", anno,
451
            &p->num_routers);
452
    if(rc) {
453 454 455 456 457
        p->num_routers = 4;
        fprintf(stderr, "Number of dimensions not specified, setting to %d\n",
                p->num_routers);
    }

458
    p->num_vcs = 3;
459

460 461
    rc = configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
    if(rc) {
462 463 464 465
        p->local_vc_size = 1024;
        fprintf(stderr, "Buffer size of local channels not specified, setting to %d\n", p->local_vc_size);
    }

466 467
    rc = configuration_get_value_int(&config, "PARAMS", "global_vc_size", anno, &p->global_vc_size);
    if(rc) {
468 469 470 471
        p->global_vc_size = 2048;
        fprintf(stderr, "Buffer size of global channels not specified, setting to %d\n", p->global_vc_size);
    }

472 473
    rc = configuration_get_value_int(&config, "PARAMS", "cn_vc_size", anno, &p->cn_vc_size);
    if(rc) {
474 475 476 477
        p->cn_vc_size = 1024;
        fprintf(stderr, "Buffer size of compute node channels not specified, setting to %d\n", p->cn_vc_size);
    }

478 479
    rc = configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(rc) {
480
        p->chunk_size = 512;
481
        fprintf(stderr, "Chunk size for packets is specified, setting to %d\n", p->chunk_size);
482 483
    }

484 485
    rc = configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
    if(rc) {
486 487 488 489
        p->local_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of local channels not specified, setting to %lf\n", p->local_bandwidth);
    }

490 491
    rc = configuration_get_value_double(&config, "PARAMS", "global_bandwidth", anno, &p->global_bandwidth);
    if(rc) {
492 493 494 495
        p->global_bandwidth = 4.7;
        fprintf(stderr, "Bandwidth of global channels not specified, setting to %lf\n", p->global_bandwidth);
    }

496 497
    rc = configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", anno, &p->cn_bandwidth);
    if(rc) {
498 499 500 501
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }

502 503 504 505
    p->router_delay = 50;
    configuration_get_value_double(&config, "PARAMS", "router_delay", anno,
            &p->router_delay);

506 507
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
508
            MAX_NAME_LENGTH);
509 510
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
511 512
    else if(strcmp(routing_str, "nonminimal")==0 || 
            strcmp(routing_str,"non-minimal")==0)
513 514 515 516 517
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
518 519 520 521
    else
    {
        fprintf(stderr, 
                "No routing protocol specified, setting to minimal routing\n");
522
        routing = -1;
523 524 525 526 527 528
    }

    // 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;
529
    p->radix = (p->num_global_channels + p->num_routers + p->num_cn);
530
    p->total_routers = p->num_groups * p->num_routers;
531
    p->total_terminals = p->total_routers * p->num_cn;
532 533 534 535 536 537 538
    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);
    }
539
    
540 541 542 543
    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
544 545 546

}

547 548 549 550
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);
551
    all_params = malloc(num_params * sizeof(*all_params));
552

Jonathan Jenkins's avatar
Jonathan Jenkins committed
553
    for (int i = 0; i < anno_map->num_annos; i++){
554
        const char * anno = anno_map->annotations[i].ptr;
555 556 557 558 559
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
560 561 562 563 564
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
565 566
   long long avg_hops, total_finished_packets, total_finished_chunks;
   long long total_finished_msgs, final_msg_sz;
567
   tw_stime avg_time, max_time;
568
   int total_minimal_packets, total_nonmin_packets;
569
   long total_gen, total_fin;
570 571 572

   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);
573 574 575
   MPI_Reduce( &N_finished_msgs, &total_finished_msgs, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
   MPI_Reduce( &N_finished_chunks, &total_finished_chunks, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
   MPI_Reduce( &total_msg_sz, &final_msg_sz, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
576 577
   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);
578 579 580
   
   MPI_Reduce( &packet_gen, &total_gen, 1, MPI_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
   MPI_Reduce( &packet_fin, &total_fin, 1, MPI_LONG, MPI_SUM, 0, MPI_COMM_WORLD);
581
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
582 583 584 585
    {
	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);
    }
586

587 588
   /* print statistics */
   if(!g_tw_mynode)
589
   {	
590 591
      printf(" Average number of hops traversed %f average chunk latency %lf us maximum chunk latency %lf us avg message size %lf bytes finished messages %lld finished chunks %lld \n", 
              (float)avg_hops/total_finished_chunks, avg_time/(total_finished_chunks*1000), max_time/1000, (float)final_msg_sz/total_finished_msgs, total_finished_msgs, total_finished_chunks);
592
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
593 594
              printf("\n ADAPTIVE ROUTING STATS: %d chunks routed minimally %d chunks routed non-minimally completed packets %lld \n", 
                      total_minimal_packets, total_nonmin_packets, total_finished_chunks);
595
 
596
      printf("\n Total packets generated %ld finished %ld \n", total_gen, total_fin);
597
   }
598 599
   return;
}
600

601 602 603
void dragonfly_collective_init(terminal_state * s,
           		   tw_lp * lp)
{
604 605 606 607 608
    // 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);
609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636
    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;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
637
        if(next_child < ((tw_lpid)num_lps * (tw_lpid)num_reps))
638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658
        {
            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
}

659 660 661 662 663
/* initialize a dragonfly compute node terminal */
void 
terminal_init( terminal_state * s, 
	       tw_lp * lp )
{
664 665 666
    s->packet_gen = 0;
    s->packet_fin = 0;

667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694
    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;
695
   
696
   s->finished_msgs = 0;
Misbah Mubarak's avatar
Misbah Mubarak committed
697 698 699
   s->finished_chunks = 0;
   s->finished_packets = 0;
   s->total_time = 0.0;
700
   s->total_msg_size = 0;
701

702 703 704
   s->last_buf_full = 0.0;
   s->busy_time = 0.0;

705
   rc_stack_create(&s->st);
706 707 708 709 710 711 712 713
   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;
    }

714
   s->rank_tbl = qhash_init(dragonfly_rank_hash_compare, dragonfly_hash_func, DFLY_HASH_TABLE_SIZE);
715 716 717 718

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

719 720 721 722 723 724
   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;
725
   s->terminal_length = 0;
726
   s->in_send_loop = 0;
727
   s->issueIdle = 0;
728 729 730 731 732 733 734 735 736

   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)
737
{
738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763
    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->cur_hist_start_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
764
   r->link_traffic = (int64_t*)malloc(p->radix * sizeof(int64_t));
765 766 767 768 769 770 771 772 773 774 775 776 777
   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**));
778
   r->queued_count = (int*)malloc(p->radix * sizeof(int));
779 780
   r->last_buf_full = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
   r->busy_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
781

782
   rc_stack_create(&r->st);
783
   for(int i=0; i < p->radix; i++)
784 785
    {
       // Set credit & router occupancy
786 787
    r->last_buf_full[i] = 0.0;
    r->busy_time[i] = 0.0;
788 789
	r->next_output_available_time[i]=0;
	r->cur_hist_start_time[i] = 0;
790
    r->link_traffic[i]=0;
791 792
	r->cur_hist_num[i] = 0;
	r->prev_hist_num[i] = 0;
793
    r->queued_count[i] = 0;    
794 795 796 797 798 799 800 801 802 803
        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*));
804
        for(int j = 0; j < p->num_vcs; j++) {
805 806 807 808 809 810 811 812 813
            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
814
//   printf("\n LP ID %d VC occupancy radix %d Router %d is connected to ", lp->gid, p->radix, r->router_id);
815 816 817
#endif 
   //round the number of global channels to the nearest even number
#if USE_DIRECT_SCHEME
818
       int first = r->router_id % p->num_routers;
819
       for(int i=0; i < p->num_global_channels; i++)
820
        {
821
            int target_grp = first;
822 823 824
            if(target_grp == r->group_id) {
                target_grp = p->num_groups - 1;
            }
825
            int my_pos = r->group_id % p->num_routers;
826 827 828 829 830 831 832
            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
833 834 835
   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++)
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863
    {
      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 */
864 865 866 867 868 869 870 871 872 873
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)
874
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
875 876
    (void)message_offset;
    (void)sched_params;
877 878 879 880 881
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

882 883 884
    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);
885 886
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
887 888
    strcpy(msg->category, req->category);
    msg->final_dest_gid = req->final_dest_lp;
889
    msg->total_size = req->msg_size;
890
    msg->sender_lp=req->src_lp;
891
    msg->sender_mn_lp = sender->gid;
892
    msg->packet_size = packet_size;
893
    msg->travel_start_time = tw_now(sender);
894 895 896
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
897
    msg->dest_terminal_id = req->dest_mn_lp;
898
    msg->message_id = req->msg_id;
899 900
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;
901
    msg->magic = terminal_magic_num; 
902 903
    msg->msg_start_time = req->msg_start_time;

904 905
    if(is_last_pckt) /* Its the last packet so pass in remote and local event information*/
      {
906
	if(req->remote_event_size > 0)
907
	 {
908 909 910
		msg->remote_event_size_bytes = req->remote_event_size;
		memcpy(tmp_ptr, remote_event, req->remote_event_size);
		tmp_ptr += req->remote_event_size;
911
	}
912
	if(req->self_event_size > 0)
913
	{
914 915 916
		msg->local_event_size_bytes = req->self_event_size;
		memcpy(tmp_ptr, self_event, req->self_event_size);
		tmp_ptr += req->self_event_size;
917 918
	}
     }
919
	   //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);
920
    tw_event_send(e_new);
921
    return xfer_to_nic_time;
922 923 924 925 926 927 928 929 930
}

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

931
/* given two group IDs, find the router of the src_gid that connects to the dest_gid*/
932
tw_lpid getRouterFromGroupID(int dest_gid, 
933
		    int src_gid,
934
		    int num_routers,
935
            int total_groups)
936
{
937 938 939 940 941 942 943
#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
944 945 946
  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;
947
  
948 949
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
950
  
951 952
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
953
  
954
  offset=(offset - 1) % (half_channel * num_routers);
955 956

  // If the destination router is in the same group
957
  tw_lpid router_id;
958 959 960 961 962 963 964

  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;
965
#endif
966 967 968
}	

/*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*/
Jonathan Jenkins's avatar
Jonathan Jenkins committed
969
void router_credit_send(router_state * s, terminal_message * msg, 
970
  tw_lp * lp, int sq) {
971 972 973 974
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

975
  int dest = 0,  type = R_BUFFER;
976
  int is_terminal = 0;
977

978
  const dragonfly_param *p = s->params;
979 980 981 982 983 984 985 986 987 988 989 990 991
 
  // 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");

992
  ts = g_tw_lookahead + p->credit_delay +  tw_rand_unif(lp->rng);
993