dragonfly.c 102 KB
Newer Older
1 2 3 4 5 6
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

7
// Local router ID: 0 --- total_router-1
8
// Router LP ID
9 10
// 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
43
#define DEBUG 0
44
#define USE_DIRECT_SCHEME 1
45
#define MAX_STATS 65536
46

47 48 49 50
#define LP_CONFIG_NM_TERM (model_net_lp_config_names[DRAGONFLY])
#define LP_METHOD_NM_TERM (model_net_method_names[DRAGONFLY])
#define LP_CONFIG_NM_ROUT (model_net_lp_config_names[DRAGONFLY_ROUTER])
#define LP_METHOD_NM_ROUT (model_net_method_names[DRAGONFLY_ROUTER])
51

52
int debug_slot_count = 0;
53
long term_ecount, router_ecount, term_rev_ecount, router_rev_ecount;
54
long packet_gen = 0, packet_fin = 0;
55

56 57
static double maxd(double a, double b) { return a < b ? b : a; }

58
/* minimal and non-minimal packet counts for adaptive routing*/
59
static int minimal_count=0, nonmin_count=0;
60
static int num_routers_per_mgrp = 0;
61

62
typedef struct dragonfly_param dragonfly_param;
63
/* annotation-specific parameters (unannotated entry occurs at the
64 65 66 67
 * last index) */
static uint64_t                  num_params = 0;
static dragonfly_param         * all_params = NULL;
static const config_anno_map_t * anno_map   = NULL;
68 69

/* global variables for codes mapping */
70
static char lp_group_name[MAX_NAME_LENGTH];
71 72
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

73 74 75 76 77 78
/* router magic number */
int router_magic_num = 0;

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

79 80
FILE * dragonfly_log = NULL;

81
int sample_bytes_written = 0;
82
int sample_rtr_bytes_written = 0;
83

84 85 86
char cn_sample_file[MAX_NAME_LENGTH];
char router_sample_file[MAX_NAME_LENGTH];

87 88 89 90 91 92 93
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;
};
94

95
void init_terminal_message_list(terminal_message_list *this,
96 97 98 99 100 101
    terminal_message *inmsg) {
    this->msg = *inmsg;
    this->event_data = NULL;
    this->next = NULL;
    this->prev = NULL;
}
102

103 104 105 106
void delete_terminal_message_list(terminal_message_list *this) {
    if(this->event_data != NULL) free(this->event_data);
    free(this);
}
107

108 109 110 111 112 113 114 115 116 117 118 119 120 121
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;
122
    int num_groups;
123 124
    int radix;
    int total_routers;
125
    int total_terminals;
126
    int num_global_channels;
127 128 129 130
    double cn_delay;
    double local_delay;
    double global_delay;
    double credit_delay;
131
    double router_delay;
132 133
};

134 135 136 137 138 139
struct dfly_hash_key
{
    uint64_t message_id;
    tw_lpid sender_id;
};

140 141 142 143
struct dfly_router_sample
{
    tw_lpid router_id;
    tw_stime* busy_time;
144
    int64_t* link_traffic_sample;
145
    tw_stime end_time;
146 147
    long fwd_events;
    long rev_events;
148 149 150
};

struct dfly_cn_sample
151 152 153 154 155 156 157 158
{
   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;
159 160
   long fwd_events;
   long rev_events;
161 162
};

163 164 165 166 167 168 169 170 171
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;
};

172 173 174 175 176 177 178 179
/* 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
{
180
   uint64_t packet_counter;
181

182 183 184
   int packet_gen;
   int packet_fin;

185
   // Dragonfly specific parameters
186 187
   unsigned int router_id;
   unsigned int terminal_id;
188 189 190

   // Each terminal will have an input and output channel with the router
   int* vc_occupancy; // NUM_VC
191
   int num_vcs;
192
   tw_stime terminal_available_time;
193 194 195
   terminal_message_list **terminal_msgs;
   terminal_message_list **terminal_msgs_tail;
   int in_send_loop;
196 197 198 199
// 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];
200 201 202
  /* collective init time */
  tw_stime collective_init_time;

203
  /* node ID in the tree */
204 205
   tw_lpid node_id;

206
   /* messages sent & received in collectives may get interchanged several times so we have to save the
207
     origin server information in the node's state */
208 209
   tw_lpid origin_svr;

210 211 212 213 214 215 216 217 218 219 220
  /* 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;

221
   struct rc_stack * st;
222 223
   int issueIdle;
   int terminal_length;
224

225 226 227
   /* to maintain a count of child nodes that have fanned in at the parent during the collective
      fan-in phase*/
   int num_fan_nodes;
228 229 230

   const char * anno;
   const dragonfly_param *params;
231

232 233 234
   struct qhash_table *rank_tbl;
   uint64_t rank_tbl_pop;

235
   tw_stime   total_time;
236
   uint64_t total_msg_size;
237
   double total_hops;
238
   long finished_msgs;
239
   long finished_chunks;
240
   long finished_packets;
241

242 243
   tw_stime last_buf_full;
   tw_stime busy_time;
244
   char output_buf[4096];
245 246
   /* For LP suspend functionality */
   int error_ct;
247 248 249 250 251 252 253 254 255

   /* 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];
256
   struct dfly_cn_sample * sample_stat;
257 258
   int op_arr_size;
   int max_arr_size;
259

260 261 262
   /* for logging forward and reverse events */
   long fwd_events;
   long rev_events;
263
};
264

265 266 267 268 269
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
270
  T_SEND,
271
  T_BUFFER,
272 273
  R_SEND,
  R_ARRIVE,
274 275 276 277
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300
};
/* 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
{
301 302
    MINIMAL = 0,
    NON_MINIMAL,
303 304
    ADAPTIVE,
    PROG_ADAPTIVE
305 306 307 308 309
};

struct router_state
{
   unsigned int router_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
310
   int group_id;
311 312 313
   int op_arr_size;
   int max_arr_size;

314 315
   int* global_channel;

316
   tw_stime* next_output_available_time;
317
   tw_stime* cur_hist_start_time;
318
   tw_stime* last_buf_full;
319

320
   tw_stime* busy_time;
321
   tw_stime* busy_time_sample;
322

323 324 325 326 327
   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;
328
   int *queued_count;
329
   struct rc_stack * st;
330

331
   int** vc_occupancy;
332
   int64_t* link_traffic;
333
   int64_t * link_traffic_sample;
334 335 336

   const char * anno;
   const dragonfly_param *params;
337 338 339

   int* prev_hist_num;
   int* cur_hist_num;
340

341
   char output_buf[4096];
342
   char output_buf2[4096];
343 344

   struct dfly_router_sample * rsamples;
345

346 347
   long fwd_events;
   long rev_events;
348 349 350 351
};

static short routing = MINIMAL;

352 353
static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
354
static tw_stime         max_collective = 0;
355

356

357 358
static long long       total_hops = 0;
static long long       N_finished_packets = 0;
359 360 361
static long long       total_msg_sz = 0;
static long long       N_finished_msgs = 0;
static long long       N_finished_chunks = 0;
362

363 364 365 366
static int dragonfly_rank_hash_compare(
        void *key, struct qhash_head *link)
{
    struct dfly_hash_key *message_key = (struct dfly_hash_key *)key;
367
    struct dfly_qhash_entry *tmp = NULL;
368 369

    tmp = qhash_entry(link, struct dfly_qhash_entry, hash_link);
370

371 372 373 374 375 376
    if (tmp->key.message_id == message_key->message_id
            && tmp->key.sender_id == message_key->sender_id)
        return 1;

    return 0;
}
377 378
static int dragonfly_hash_func(void *k, int table_size)
{
379
    struct dfly_hash_key *tmp = (struct dfly_hash_key *)k;
380
    //uint32_t pc = 0, pb = 0;
381 382
    //bj_hashlittle2(tmp, sizeof(*tmp), &pc, &pb);
    uint64_t key = (~tmp->message_id) + (tmp->message_id << 18);
383 384
    key = key * 21;
    key = ~key ^ (tmp->sender_id >> 4);
385
    key = key * tmp->sender_id;
386 387
    return (int)(key & (table_size - 1));
    //return (int)(pc % (table_size - 1));
388 389
}

390 391 392 393 394 395 396
/* 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);
397
    /* GiB to s */
398 399 400 401 402 403
    time = time / GB_p_s;
    /* s to ns */
    time = time * 1000.0 * 1000.0 * 1000.0;

    return(time);
}
404

405 406
/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
407
{
408 409
	   return sizeof(terminal_message);
}
410

411 412
static void free_tmp(void * ptr)
{
413
    struct dfly_qhash_entry * dfly = ptr;
414 415 416
    free(dfly->remote_event_data);
    free(dfly);
}
417
static void append_to_terminal_message_list(
418 419
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
420
        int index,
421 422 423 424 425 426
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thisq[index] = msg;
    } else {
        thistail[index]->next = msg;
        msg->prev = thistail[index];
427
    }
428
    thistail[index] = msg;
429 430
}

431
static void prepend_to_terminal_message_list(
432 433
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
434
        int index,
435 436 437 438 439 440
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thistail[index] = msg;
    } else {
        thisq[index]->prev = msg;
        msg->next = thisq[index];
441
    }
442 443
    thisq[index] = msg;
}
444

445 446 447 448 449 450 451 452 453 454 455 456 457 458 459
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;
460 461
}

462 463 464 465 466
static terminal_message_list* return_tail(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index) {
    terminal_message_list *tail = thistail[index];
467
    assert(tail);
468 469 470 471 472 473 474 475 476
    if(tail->prev != NULL) {
        tail->prev->next = NULL;
        thistail[index] = tail->prev;
        tail->prev = NULL;
    } else {
        thistail[index] = NULL;
        thisq[index] = NULL;
    }
    return tail;
477 478
}

479 480 481
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
482

483
    int rc = configuration_get_value_int(&config, "PARAMS", "num_routers", anno,
484
            &p->num_routers);
485
    if(rc) {
486 487 488 489 490
        p->num_routers = 4;
        fprintf(stderr, "Number of dimensions not specified, setting to %d\n",
                p->num_routers);
    }

491
    p->num_vcs = 3;
492

493 494
    rc = configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
    if(rc) {
495 496 497 498
        p->local_vc_size = 1024;
        fprintf(stderr, "Buffer size of local channels not specified, setting to %d\n", p->local_vc_size);
    }

499 500
    rc = configuration_get_value_int(&config, "PARAMS", "global_vc_size", anno, &p->global_vc_size);
    if(rc) {
501 502 503 504
        p->global_vc_size = 2048;
        fprintf(stderr, "Buffer size of global channels not specified, setting to %d\n", p->global_vc_size);
    }

505 506
    rc = configuration_get_value_int(&config, "PARAMS", "cn_vc_size", anno, &p->cn_vc_size);
    if(rc) {
507 508 509 510
        p->cn_vc_size = 1024;
        fprintf(stderr, "Buffer size of compute node channels not specified, setting to %d\n", p->cn_vc_size);
    }

511 512
    rc = configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(rc) {
513
        p->chunk_size = 512;
514
        fprintf(stderr, "Chunk size for packets is specified, setting to %d\n", p->chunk_size);
515 516
    }

517 518
    rc = configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
    if(rc) {
519 520 521 522
        p->local_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of local channels not specified, setting to %lf\n", p->local_bandwidth);
    }

523 524
    rc = configuration_get_value_double(&config, "PARAMS", "global_bandwidth", anno, &p->global_bandwidth);
    if(rc) {
525 526 527 528
        p->global_bandwidth = 4.7;
        fprintf(stderr, "Bandwidth of global channels not specified, setting to %lf\n", p->global_bandwidth);
    }

529 530
    rc = configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", anno, &p->cn_bandwidth);
    if(rc) {
531 532 533 534
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }

535 536 537 538
    p->router_delay = 50;
    configuration_get_value_double(&config, "PARAMS", "router_delay", anno,
            &p->router_delay);

539 540 541 542
    configuration_get_value(&config, "PARAMS", "cn_sample_file", anno, cn_sample_file,
            MAX_NAME_LENGTH);
    configuration_get_value(&config, "PARAMS", "rt_sample_file", anno, router_sample_file,
            MAX_NAME_LENGTH);
543

544 545
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
546
            MAX_NAME_LENGTH);
547 548
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
549
    else if(strcmp(routing_str, "nonminimal")==0 ||
550
            strcmp(routing_str,"non-minimal")==0)
551 552 553 554 555
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
556 557
    else
    {
558
        fprintf(stderr,
559
                "No routing protocol specified, setting to minimal routing\n");
560
        routing = -1;
561 562 563 564 565 566
    }

    // 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;
567
    p->radix = (p->num_routers + p->num_global_channels + p->num_cn);
568
    p->total_routers = p->num_groups * p->num_routers;
569
    p->total_terminals = p->total_routers * p->num_cn;
570 571 572 573 574 575 576
    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);
    }
577

578 579 580 581
    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
582 583
}

584
static void dragonfly_configure(){
585
    anno_map = codes_mapping_get_lp_anno_map(LP_CONFIG_NM_TERM);
586 587
    assert(anno_map);
    num_params = anno_map->num_annos + (anno_map->has_unanno_lp > 0);
588
    all_params = malloc(num_params * sizeof(*all_params));
589

Jonathan Jenkins's avatar
Jonathan Jenkins committed
590
    for (int i = 0; i < anno_map->num_annos; i++){
591
        const char * anno = anno_map->annotations[i].ptr;
592 593 594 595 596
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
597 598 599 600 601
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
602 603
   long long avg_hops, total_finished_packets, total_finished_chunks;
   long long total_finished_msgs, final_msg_sz;
604
   tw_stime avg_time, max_time;
605
   int total_minimal_packets, total_nonmin_packets;
606
   long total_gen, total_fin;
607 608 609

   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);
610 611 612
   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);
613 614
   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);
615

616 617
   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);
618
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
619 620 621 622
    {
	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);
    }
623

624 625
   /* print statistics */
   if(!g_tw_mynode)
626 627
   {
      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",
628
              (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);
629
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
630
              printf("\n ADAPTIVE ROUTING STATS: %d chunks routed minimally %d chunks routed non-minimally completed packets %lld \n",
631
                      total_minimal_packets, total_nonmin_packets, total_finished_chunks);
632

633
      printf("\n Total packets generated %ld finished %ld \n", total_gen, total_fin);
634
   }
635 636
   return;
}
637

638 639 640
void dragonfly_collective_init(terminal_state * s,
           		   tw_lp * lp)
{
641 642 643
    // 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);
644
    int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
645
            NULL, 1);
646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673
    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
674
        if(next_child < ((tw_lpid)num_lps * (tw_lpid)num_reps))
675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695
        {
            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
}

696
/* initialize a dragonfly compute node terminal */
697 698
void
terminal_init( terminal_state * s,
699 700
	       tw_lp * lp )
{
701 702 703
    s->packet_gen = 0;
    s->packet_fin = 0;

704
    uint32_t h1 = 0, h2 = 0;
705
    bj_hashlittle2(LP_METHOD_NM_TERM, strlen(LP_METHOD_NM_TERM), &h1, &h2);
706
    terminal_magic_num = h1 + h2;
707

708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724
    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];
    }

725
   int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
726 727
           s->anno, 0);

728 729
   s->terminal_id = (mapping_rep_id * num_lps) + mapping_offset;

730 731 732
   s->router_id=(int)s->terminal_id / (s->params->num_routers/2);
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;
733

734
   s->finished_msgs = 0;
735 736 737
   s->finished_chunks = 0;
   s->finished_packets = 0;
   s->total_time = 0.0;
738
   s->total_msg_size = 0;
739

740 741 742
   s->last_buf_full = 0.0;
   s->busy_time = 0.0;

743 744 745
   s->fwd_events = 0;
   s->rev_events = 0;

746
   rc_stack_create(&s->st);
747 748 749 750 751 752 753 754
   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;
    }

755
   s->rank_tbl = qhash_init(dragonfly_rank_hash_compare, dragonfly_hash_func, DFLY_HASH_TABLE_SIZE);
756 757 758 759

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

760
   s->terminal_msgs =
761
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
762
   s->terminal_msgs_tail =
763 764 765
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
   s->terminal_msgs[0] = NULL;
   s->terminal_msgs_tail[0] = NULL;
766
   s->terminal_length = 0;
767
   s->in_send_loop = 0;
768
   s->issueIdle = 0;
769 770 771 772 773

   dragonfly_collective_init(s, lp);
   return;
}

774
/* sets up the router virtual channels, global channels,
775 776
 * local channels, compute node channels */
void router_setup(router_state * r, tw_lp * lp)
777
{
778
    uint32_t h1 = 0, h2 = 0;
779
    bj_hashlittle2(LP_METHOD_NM_ROUT, strlen(LP_METHOD_NM_ROUT), &h1, &h2);
780
    router_magic_num = h1 + h2;
781

782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797
    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;

798 799 800 801 802 803 804 805
    num_routers_per_mgrp = codes_mapping_get_lp_count (lp_group_name, 1, "modelnet_dragonfly_router",
            NULL, 0);
    int num_grp_reps = codes_mapping_get_group_reps(lp_group_name);
    if(p->total_routers != num_grp_reps * num_routers_per_mgrp)
        tw_error(TW_LOC, "\n Config error: num_routers specified %d total routers computed in the network %d "
                "does not match with repetitions * dragonfly_router %d  ",
                p->num_routers, p->total_routers, num_grp_reps * num_routers_per_mgrp);

806 807 808
   r->router_id=mapping_rep_id + mapping_offset;
   r->group_id=r->router_id/p->num_routers;

809 810 811
   r->fwd_events = 0;
   r->rev_events = 0;

812 813 814
   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));
815
   r->link_traffic = (int64_t*)malloc(p->radix * sizeof(int64_t));
816
   r->link_traffic_sample = (int64_t*)malloc(p->radix * sizeof(int64_t));
817 818
   r->cur_hist_num = (int*)malloc(p->radix * sizeof(int));
   r->prev_hist_num = (int*)malloc(p->radix * sizeof(int));
819

820 821
   r->vc_occupancy = (int**)malloc(p->radix * sizeof(int*));
   r->in_send_loop = (int*)malloc(p->radix * sizeof(int));
822
   r->pending_msgs =
823
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
824
   r->pending_msgs_tail =
825
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
826
   r->queued_msgs =
827
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
828
   r->queued_msgs_tail =
829
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
830
   r->queued_count = (int*)malloc(p->radix * sizeof(int));
831 832
   r->last_buf_full = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
   r->busy_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
833
   r->busy_time_sample = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
834

835
   rc_stack_create(&r->st);
836
   for(int i=0; i < p->radix; i++)
837 838
    {
       // Set credit & router occupancy
839 840
    r->last_buf_full[i] = 0.0;
    r->busy_time[i] = 0.0;
841
    r->busy_time_sample[i] = 0.0;
842 843
	r->next_output_available_time[i]=0;
	r->cur_hist_start_time[i] = 0;
844
    r->link_traffic[i]=0;
845
    r->link_traffic_sample[i] = 0;
846 847
	r->cur_hist_num[i] = 0;
	r->prev_hist_num[i] = 0;
848
    r->queued_count[i] = 0;
849 850
    r->in_send_loop[i] = 0;
    r->vc_occupancy[i] = (int*)malloc(p->num_vcs * sizeof(int));
851
    r->pending_msgs[i] = (terminal_message_list**)malloc(p->num_vcs *
852
        sizeof(terminal_message_list*));
853
    r->pending_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs *
854
        sizeof(terminal_message_list*));
855
    r->queued_msgs[i] = (terminal_message_list**)malloc(p->num_vcs *
856
        sizeof(terminal_message_list*));
857
    r->queued_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs *
858
        sizeof(terminal_message_list*));
859
        for(int j = 0; j < p->num_vcs; j++) {
860 861 862 863 864 865 866 867 868
            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
869
//   printf("\n LP ID %d VC occupancy radix %d Router %d is connected to ", lp->gid, p->radix, r->router_id);
870
#endif
871 872
   //round the number of global channels to the nearest even number
#if USE_DIRECT_SCHEME
873
       int first = r->router_id % p->num_routers;
874
       for(int i=0; i < p->num_global_channels; i++)
875
        {
876
            int target_grp = first;
877 878 879
            if(target_grp == r->group_id) {
                target_grp = p->num_groups - 1;
            }
880
            int my_pos = r->group_id % p->num_routers;
881 882 883 884 885 886 887
            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
888
   int router_offset = (r->router_id % p->num_routers) *
889 890
    (p->num_global_channels / 2) + 1;
   for(int i=0; i < p->num_global_channels; i++)
891 892 893 894 895 896 897 898 899 900 901 902
    {
      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)
         {
903
           r->global_channel[i]=p->total_routers+r->global_channel[i];
904 905 906
	 }
#if DEBUG == 1
    printf("\n channel %d ", r->global_channel[i]);
907
#endif
908 909 910 911 912 913 914
    }
#endif

#if DEBUG == 1
   printf("\n");
#endif
   return;
915
}
916 917 918


/* dragonfly packet event , generates a dragonfly packet on the compute node */
919 920 921 922 923 924 925 926 927 928
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)
929
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
930 931
    (void)message_offset;
    (void)sched_params;
932 933 934 935 936
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

937
    xfer_to_nic_time = codes_local_latency(sender);
938 939
    //e_new = tw_event_new(sender->gid, xfer_to_nic_time+offset, sender);
    //msg = tw_event_data(e_new);
940 941
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
942 943
    strcpy(msg->category, req->category);
    msg->final_dest_gid = req->final_dest_lp;
944
    msg->total_size = req->msg_size;
945
    msg->sender_lp=req->src_lp;
946
    msg->sender_mn_lp = sender->gid;
947
    msg->packet_size = packet_size;
948
    msg->travel_start_time = tw_now(sender);
949 950 951
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
952
    msg->dest_terminal_id = req->dest_mn_lp;
953
    msg->message_id = req->msg_id;
954 955
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;
956
    msg->magic = terminal_magic_num;
957 958
    msg->msg_start_time = req->msg_start_time;

959 960
    if(is_last_pckt) /* Its the last packet so pass in remote and local event information*/
      {
961
	if(req->remote_event_size > 0)
962
	 {
963 964 965
		msg->remote_event_size_bytes = req->remote_event_size;
		memcpy(tmp_ptr, remote_event, req->remote_event_size);
		tmp_ptr += req->remote_event_size;
966
	}
967
	if(req->self_event_size > 0)
968
	{
969 970 971
		msg->local_event_size_bytes = req->self_event_size;
		memcpy(tmp_ptr, self_event, req->self_event_size);
		tmp_ptr += req->self_event_size;
972 973
	}
     }
974
	   //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);
975
    tw_event_send(e_new);
976
    return xfer_to_nic_time;
977 978 979 980 981 982 983 984 985
}

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

986
/* given two group IDs, find the router of the src_gid that connects to the dest_gid*/
987
tw_lpid getRouterFromGroupID(int dest_gid,
988
		    int src_gid,
989
		    int num_routers,
990
            int total_groups)
991
{
992 993 994 995 996 997 998
#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
999 1000 1001
  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;
1002

1003 1004
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
1005

1006 1007
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
1008

1009
  offset=(offset - 1) % (half_channel * num_routers);
1010 1011

  // If the destination router is in the same group
1012
  tw_lpid router_id;
1013 1014 1015 1016 1017 1018 1019

  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;
1020
#endif
1021
}
1022 1023

/*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*/
1024
void router_credit_send(router_state * s, terminal_message * msg,
1025
  tw_lp * lp, int sq) {
1026 1027 1028 1029
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

1030
  int dest = 0,  type = R_BUFFER;
1031
  int is_terminal = 0;
1032

1033
  const dragonfly_param *p = s->params;
1034

1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046
  // 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");

1047
  ts = g_tw_lookahead + p->credit_delay +  tw_rand_unif(lp->rng);
1048

1049
  if (is_terminal) {
1050
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY,
1051 1052 1053
      (void**)&buf_msg, NULL);
    buf_msg->magic = terminal_magic_num;
  } else {
1054 1055
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY_ROUTER,
            (void**)&buf_msg, NULL);
1056 1057
    buf_msg->magic = router_magic_num;
  }
1058

1059 1060 1061 1062
  if(sq == -1) {
    buf_msg->vc_index = msg->vc_index;
    buf_msg->output_chan = msg->output_chan;
  } else {
1063
    buf_msg->vc_index = msg->saved_vc;
1064 1065
    buf_msg->output_chan = msg->saved_channel;
  }
1066

1067
  buf_msg->type = type;
1068

1069 1070
  tw_event_send(buf_e);
  return;
1071 1072
}

1073
void packet_generate_rc(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
1074
{
1075 1076
   s->packet_gen--;
   packet_gen--;
1077

1078
   tw_rand_reverse_unif(lp->rng);
1079

1080 1081 1082
   int num_chunks = msg->packet_size/s->params->chunk_size;
   if(msg->packet_size % s->params->chunk_size)
       num_chunks++;
1083

1084
   if(!num_chunks)
1085
       num_chunks = 1;
1086

1087 1088
   int i;
   for(i = 0; i < num_chunks; i++) {
1089
        delete_terminal_message_list(return_tail(s->terminal_msgs,
1090
          s->terminal_msgs_tail, 0));
1091
        s->terminal_length -= s->params->chunk_size;
1092 1093
   }
    if(bf->c5) {
1094
        codes_local_latency_reverse(lp);
1095 1096
        s->in_send_loop = 0;
    }
1097 1098
      if(bf->c11) {
        s->issueIdle = 0;
1099
        s->last_buf_full = msg->saved_busy_time;
1100
      }
1101 1102
     struct mn_stats* stat;
     stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
1103 1104 1105 1106
     stat->send_count--;
     stat->send_bytes -= msg->packet_size;
     stat->send_time -= (1/s->params->cn_bandwidth) * msg->packet_size;
}
1107

1108
/* generates packet at the current dragonfly compute node */
1109
void packet_generate(terminal_state * s, tw_bf * bf, terminal_message * msg,
1110
  tw_lp * lp) {
1111 1112
  packet_gen++;
  s->packet_gen++;
1113

1114
  tw_stime ts, nic_ts;
1115

1116
  assert(lp->gid != msg->dest_terminal_id);
1117
  const dragonfly_param *p = s->params;
1118

Jonathan Jenkins's avatar
Jonathan Jenkins committed
1119 1120
  int total_event_size;
  uint64_t num_chunks = msg->packet_size / p->chunk_size;
1121
  if (msg->packet_size % s->params->chunk_size)
1122
      num_chunks++;
1123 1124 1125 1126

  if(!num_chunks)
    num_chunks = 1;

1127
  nic_ts = g_tw_lookahead + (num_chunks * s->params->cn_delay) + tw_rand_unif(lp->rng);
1128

1129
  msg->packet_ID = lp->gid + g_tw_nlp * s->packet_counter;
1130 1131 1132 1133
  msg->my_N_hop = 0;
  msg->my_l_hop = 0;
  msg->my_g_hop = 0;
  msg->intm_group_id = -1;
1134

1135
  //if(msg->dest_terminal_id == TRACK)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
1136
  if(msg->packet_ID == LLU(TRACK_PKT))
1137
    printf("\n Packet %llu generated at terminal %d dest %llu size %llu num chunks %llu ",
Jonathan Jenkins's avatar
Jonathan Jenkins committed
1138 1139
            msg->packet_ID, s->terminal_id, LLU(msg->dest_terminal_id),
            LLU(msg->packet_size), LLU(num_chunks));
1140

Jonathan Jenkins's avatar
Jonathan Jenkins committed
1141
  for(uint64_t i = 0; i < num_chunks; i++)
1142 1143 1144
  {
    terminal_message_list *cur_chunk = (terminal_message_list*)malloc(
      sizeof(terminal_message_list));
1145
    msg->origin_router_id = s->router_id;
1146
    init_terminal_message_list(cur_chunk, msg);
1147

1148

1149 1150 1151 1152
    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);
    }
1153

1154 1155 1156 1157
    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);
    }
1158
    if (msg->local_event_size_bytes){
1159
      m_data_src = (char*)m_data_src + msg->remote_event_size_bytes;
1160
      memcpy((char*)cur_chunk->event_data + msg->remote_event_size_bytes,
1161 1162
          m_data_src, msg->local_event_size_bytes);
    }
1163

1164
    cur_chunk->msg.chunk_id = i;
1165
    cur_chunk->msg.origin_router_id = s->router_id;
1166 1