dragonfly.c 111 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
#define CREDIT_SZ 8
26 27
#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 4999
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

Nikhil's avatar
Nikhil committed
95
static 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

Nikhil's avatar
Nikhil committed
103
static void delete_terminal_message_list(terminal_message_list *this) {
104 105 106
    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
    int num_real_groups;
124 125
    int radix;
    int total_routers;
126
    int total_terminals;
127
    int num_global_channels;
128 129 130 131
    double cn_delay;
    double local_delay;
    double global_delay;
    double credit_delay;
132
    double router_delay;
133 134
};

135 136 137 138 139 140
struct dfly_hash_key
{
    uint64_t message_id;
    tw_lpid sender_id;
};

141 142 143 144
struct dfly_router_sample
{
    tw_lpid router_id;
    tw_stime* busy_time;
145
    int64_t* link_traffic_sample;
146
    tw_stime end_time;
147 148
    long fwd_events;
    long rev_events;
149 150 151
};

struct dfly_cn_sample
152 153 154 155 156 157 158 159
{
   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;
160 161
   long fwd_events;
   long rev_events;
162 163
};

164 165 166 167 168 169 170 171 172
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;
};

173 174 175 176 177 178 179 180
/* 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
{
181
   uint64_t packet_counter;
182

183 184 185
   int packet_gen;
   int packet_fin;

186
   // Dragonfly specific parameters
187 188
   unsigned int router_id;
   unsigned int terminal_id;
189 190 191

   // Each terminal will have an input and output channel with the router
   int* vc_occupancy; // NUM_VC
192
   int num_vcs;
193
   tw_stime terminal_available_time;
194 195 196
   terminal_message_list **terminal_msgs;
   terminal_message_list **terminal_msgs_tail;
   int in_send_loop;
197 198 199 200
// 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];
201 202 203
  /* collective init time */
  tw_stime collective_init_time;

204
  /* node ID in the tree */ 
205 206
   tw_lpid node_id;

207
   /* messages sent & received in collectives may get interchanged several times so we have to save the 
208
     origin server information in the node's state */
209 210
   tw_lpid origin_svr; 
  
211 212 213 214 215 216 217 218 219 220 221
  /* 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;

222
   struct rc_stack * st;
223 224
   int issueIdle;
   int terminal_length;
225

226 227 228
   /* to maintain a count of child nodes that have fanned in at the parent during the collective
      fan-in phase*/
   int num_fan_nodes;
229 230 231

   const char * anno;
   const dragonfly_param *params;
232

233 234 235
   struct qhash_table *rank_tbl;
   uint64_t rank_tbl_pop;

236
   tw_stime   total_time;
237
   uint64_t total_msg_size;
238
   double total_hops;
239
   long finished_msgs;
240
   long finished_chunks;
241
   long finished_packets;
242

243 244
   tw_stime last_buf_full;
   tw_stime busy_time;
245
   char output_buf[4096];
246 247
   /* For LP suspend functionality */
   int error_ct;
248 249 250 251 252 253 254 255 256

   /* 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];
257
   struct dfly_cn_sample * sample_stat;
258 259
   int op_arr_size;
   int max_arr_size;
260
   
261 262 263
   /* for logging forward and reverse events */
   long fwd_events;
   long rev_events;
264 265 266 267 268 269 270

   /* following used for ROSS model-level stats collection */
   long fin_chunks_ross_sample;
   long data_size_ross_sample;
   long fin_hops_ross_sample;
   tw_stime fin_chunks_time_ross_sample;
   tw_stime busy_time_ross_sample;
271
};
272

273 274 275 276 277
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
278
  T_SEND,
279
  T_BUFFER,
280 281
  R_SEND,
  R_ARRIVE,
282 283 284 285
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
};
/* 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
{
309 310
    MINIMAL = 0,
    NON_MINIMAL,
311 312
    ADAPTIVE,
    PROG_ADAPTIVE
313 314 315 316 317
};

struct router_state
{
   unsigned int router_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
318
   int group_id;
319 320 321
   int op_arr_size;
   int max_arr_size;

322 323
   int* global_channel; 
   
324
   tw_stime* next_output_available_time;
325
   tw_stime* cur_hist_start_time;
326
   tw_stime* last_buf_full;
327

328
   tw_stime* busy_time;
329
   tw_stime* busy_time_sample;
330

331 332 333 334 335
   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;
336
   int *queued_count;
337
   struct rc_stack * st;
338
   
339
   int** vc_occupancy;
340
   int64_t* link_traffic;
341
   int64_t * link_traffic_sample;
342 343 344

   const char * anno;
   const dragonfly_param *params;
345 346 347

   int* prev_hist_num;
   int* cur_hist_num;
348
   
349
   char output_buf[4096];
350
   char output_buf2[4096];
351 352

   struct dfly_router_sample * rsamples;
353
   
354 355
   long fwd_events;
   long rev_events;
356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380

   /* following used for ROSS model-level stats collection */
   tw_stime* busy_time_ross_sample;
   int64_t * link_traffic_ross_sample;
};

/* had to pull some of the ROSS model stats collection stuff up here */
void dragonfly_event_collect(terminal_message *m, tw_lp *lp, char *buffer, int *collect_flag);
void dragonfly_model_stat_collect(terminal_state *s, tw_lp *lp, char *buffer);
void dfly_router_model_stat_collect(router_state *s, tw_lp *lp, char *buffer);

st_model_types dragonfly_model_types[] = {
    {(rbev_trace_f) dragonfly_event_collect,
     sizeof(int),
     (ev_trace_f) dragonfly_event_collect,
     sizeof(int),
     (model_stat_f) dragonfly_model_stat_collect,
     sizeof(tw_lpid) + sizeof(long) * 2 + sizeof(double) + sizeof(tw_stime) *2}, 
    {(rbev_trace_f) dragonfly_event_collect,
     sizeof(int),
     (ev_trace_f) dragonfly_event_collect,
     sizeof(int),
     (model_stat_f) dfly_router_model_stat_collect,
     0}, //updated in router_setup() since it's based on the radix
    {0}
381
};
382
/* End of ROSS model stats collection */
383 384 385

static short routing = MINIMAL;

386 387
static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
388
static tw_stime         max_collective = 0;
389

390

391 392
static long long       total_hops = 0;
static long long       N_finished_packets = 0;
393 394 395
static long long       total_msg_sz = 0;
static long long       N_finished_msgs = 0;
static long long       N_finished_chunks = 0;
396

397 398 399 400
static int dragonfly_rank_hash_compare(
        void *key, struct qhash_head *link)
{
    struct dfly_hash_key *message_key = (struct dfly_hash_key *)key;
401
    struct dfly_qhash_entry *tmp = NULL;
402 403

    tmp = qhash_entry(link, struct dfly_qhash_entry, hash_link);
404
    
405 406 407 408 409 410
    if (tmp->key.message_id == message_key->message_id
            && tmp->key.sender_id == message_key->sender_id)
        return 1;

    return 0;
}
411 412
static int dragonfly_hash_func(void *k, int table_size)
{
413
    struct dfly_hash_key *tmp = (struct dfly_hash_key *)k;
414
    //uint32_t pc = 0, pb = 0;	
415 416
    //bj_hashlittle2(tmp, sizeof(*tmp), &pc, &pb);
    uint64_t key = (~tmp->message_id) + (tmp->message_id << 18);
417 418
    key = key * 21;
    key = ~key ^ (tmp->sender_id >> 4);
419
    key = key * tmp->sender_id; 
420 421
    return (int)(key & (table_size - 1));
    //return (int)(pc % (table_size - 1));
422 423
}

424 425 426 427 428 429 430
/* 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);
431
    /* GiB to s */
432 433 434 435 436 437
    time = time / GB_p_s;
    /* s to ns */
    time = time * 1000.0 * 1000.0 * 1000.0;

    return(time);
}
438

439 440
/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
441
{
442 443
	   return sizeof(terminal_message);
}
444

445 446
static void free_tmp(void * ptr)
{
447
    struct dfly_qhash_entry * dfly = ptr; 
448 449 450 451 452 453
    
    if(dfly->remote_event_data)
        free(dfly->remote_event_data);
   
    if(dfly)
        free(dfly);
454
}
455
static void append_to_terminal_message_list(  
456 457
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
458
        int index, 
459 460 461 462 463 464
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thisq[index] = msg;
    } else {
        thistail[index]->next = msg;
        msg->prev = thistail[index];
465
    } 
466
    thistail[index] = msg;
467 468
}

469
static void prepend_to_terminal_message_list(  
470 471
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
472
        int index, 
473 474 475 476 477 478
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thistail[index] = msg;
    } else {
        thisq[index]->prev = msg;
        msg->next = thisq[index];
479
    } 
480 481
    thisq[index] = msg;
}
482

483 484 485 486 487 488 489 490 491 492 493 494 495 496 497
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;
498 499
}

500 501 502 503 504
static terminal_message_list* return_tail(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index) {
    terminal_message_list *tail = thistail[index];
505
    assert(tail);
506 507 508 509 510 511 512 513 514
    if(tail->prev != NULL) {
        tail->prev->next = NULL;
        thistail[index] = tail->prev;
        tail->prev = NULL;
    } else {
        thistail[index] = NULL;
        thisq[index] = NULL;
    }
    return tail;
515 516
}

517 518 519
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
520

521
    int rc = configuration_get_value_int(&config, "PARAMS", "num_routers", anno,
522
            &p->num_routers);
523
    if(rc) {
524 525 526 527 528
        p->num_routers = 4;
        fprintf(stderr, "Number of dimensions not specified, setting to %d\n",
                p->num_routers);
    }

529
    p->num_vcs = 3;
530

531 532
    rc = configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
    if(rc) {
533 534 535 536
        p->local_vc_size = 1024;
        fprintf(stderr, "Buffer size of local channels not specified, setting to %d\n", p->local_vc_size);
    }

537 538
    rc = configuration_get_value_int(&config, "PARAMS", "global_vc_size", anno, &p->global_vc_size);
    if(rc) {
539 540 541 542
        p->global_vc_size = 2048;
        fprintf(stderr, "Buffer size of global channels not specified, setting to %d\n", p->global_vc_size);
    }

543 544
    rc = configuration_get_value_int(&config, "PARAMS", "cn_vc_size", anno, &p->cn_vc_size);
    if(rc) {
545 546 547 548
        p->cn_vc_size = 1024;
        fprintf(stderr, "Buffer size of compute node channels not specified, setting to %d\n", p->cn_vc_size);
    }

549 550
    rc = configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(rc) {
551
        p->chunk_size = 512;
552
        fprintf(stderr, "Chunk size for packets is specified, setting to %d\n", p->chunk_size);
553 554
    }

555 556
    rc = configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
    if(rc) {
557 558 559 560
        p->local_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of local channels not specified, setting to %lf\n", p->local_bandwidth);
    }

561 562
    rc = configuration_get_value_double(&config, "PARAMS", "global_bandwidth", anno, &p->global_bandwidth);
    if(rc) {
563 564 565 566
        p->global_bandwidth = 4.7;
        fprintf(stderr, "Bandwidth of global channels not specified, setting to %lf\n", p->global_bandwidth);
    }

567 568
    rc = configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", anno, &p->cn_bandwidth);
    if(rc) {
569 570 571 572
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }

573 574 575 576
    p->router_delay = 50;
    configuration_get_value_double(&config, "PARAMS", "router_delay", anno,
            &p->router_delay);

577 578 579 580
    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);
581
    
582 583
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
584
            MAX_NAME_LENGTH);
585 586
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
587
    else if(strcmp(routing_str, "nonminimal")==0 || 
588
            strcmp(routing_str,"non-minimal")==0)
589 590 591 592 593
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
594 595
    else
    {
596
        fprintf(stderr, 
597
                "No routing protocol specified, setting to minimal routing\n");
598
        routing = -1;
599 600 601 602 603 604
    }

    // 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;
605
    p->radix = (p->num_routers + p->num_global_channels + p->num_cn);
606
    p->total_routers = p->num_groups * p->num_routers;
607
    p->total_terminals = p->total_routers * p->num_cn;
608 609 610 611 612 613 614
    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);
    }
615
    
616 617 618
    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);
619
    p->credit_delay = bytes_to_ns(CREDIT_SZ, p->local_bandwidth); //assume 8 bytes packet
620 621
}

622
static void dragonfly_configure(){
623
    anno_map = codes_mapping_get_lp_anno_map(LP_CONFIG_NM_TERM);
624 625
    assert(anno_map);
    num_params = anno_map->num_annos + (anno_map->has_unanno_lp > 0);
626
    all_params = malloc(num_params * sizeof(*all_params));
627

Jonathan Jenkins's avatar
Jonathan Jenkins committed
628
    for (int i = 0; i < anno_map->num_annos; i++){
629
        const char * anno = anno_map->annotations[i].ptr;
630 631 632 633 634
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
635 636 637 638 639
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
640 641
   long long avg_hops, total_finished_packets, total_finished_chunks;
   long long total_finished_msgs, final_msg_sz;
642
   tw_stime avg_time, max_time;
643
   int total_minimal_packets, total_nonmin_packets;
644
   long total_gen, total_fin;
645 646 647

   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);
648 649 650
   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);
651 652
   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);
653
   
654 655
   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);
656
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
657 658 659 660
    {
	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);
    }
661

662 663
   /* print statistics */
   if(!g_tw_mynode)
664 665
   {	
      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", 
666
              (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);
667
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
668
              printf("\n ADAPTIVE ROUTING STATS: %d chunks routed minimally %d chunks routed non-minimally completed packets %lld \n", 
669
                      total_minimal_packets, total_nonmin_packets, total_finished_chunks);
670
 
671
      printf("\n Total packets generated %ld finished %ld \n", total_gen, total_fin);
672
   }
673 674
   return;
}
675

Nikhil's avatar
Nikhil committed
676
static void dragonfly_collective_init(terminal_state * s,
677 678
           		   tw_lp * lp)
{
679 680 681
    // 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);
682
    int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
683
            NULL, 1);
684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711
    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
712
        if(next_child < ((tw_lpid)num_lps * (tw_lpid)num_reps))
713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733
        {
            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
}

734
/* initialize a dragonfly compute node terminal */
Nikhil's avatar
Nikhil committed
735
static void 
736
terminal_init( terminal_state * s, 
737 738
	       tw_lp * lp )
{
739 740 741
    s->packet_gen = 0;
    s->packet_fin = 0;

742
    uint32_t h1 = 0, h2 = 0; 
743
    bj_hashlittle2(LP_METHOD_NM_TERM, strlen(LP_METHOD_NM_TERM), &h1, &h2);
744
    terminal_magic_num = h1 + h2;
745
    
746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762
    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];
    }

763
   int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
764 765
           s->anno, 0);

766
   s->terminal_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);  
767
   
768
   s->router_id=(int)s->terminal_id / s->params->num_cn;
769 770
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;
771
   
772
   s->finished_msgs = 0;
773 774 775
   s->finished_chunks = 0;
   s->finished_packets = 0;
   s->total_time = 0.0;
776
   s->total_msg_size = 0;
777

778 779 780
   s->last_buf_full = 0.0;
   s->busy_time = 0.0;

781 782 783
   s->fwd_events = 0;
   s->rev_events = 0;

784
   rc_stack_create(&s->st);
785 786 787 788 789 790 791 792
   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;
    }

793

794
   s->rank_tbl = NULL;
795
   s->terminal_msgs = 
796
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
797
   s->terminal_msgs_tail = 
798 799 800
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
   s->terminal_msgs[0] = NULL;
   s->terminal_msgs_tail[0] = NULL;
801
   s->terminal_length = 0;
802
   s->in_send_loop = 0;
803
   s->issueIdle = 0;
804

805 806 807 808 809 810 811
   /* set up for ROSS stats sampling */
   s->fin_chunks_ross_sample = 0;
   s->data_size_ross_sample = 0;
   s->fin_hops_ross_sample = 0;
   s->fin_chunks_time_ross_sample = 0.0;
   s->busy_time_ross_sample = 0.0;

812 813 814 815
   dragonfly_collective_init(s, lp);
   return;
}

816
/* sets up the router virtual channels, global channels, 
817
 * local channels, compute node channels */
Nikhil's avatar
Nikhil committed
818
static void router_setup(router_state * r, tw_lp * lp)
819
{
820
    uint32_t h1 = 0, h2 = 0; 
821
    bj_hashlittle2(LP_METHOD_NM_ROUT, strlen(LP_METHOD_NM_ROUT), &h1, &h2);
822
    router_magic_num = h1 + h2;
823
    
824 825 826 827 828 829 830 831 832 833 834 835 836
    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];
    }

837 838 839 840 841 842 843 844 845 846 847 848
    dragonfly_param *p = r->params;
    p->num_real_groups = codes_mapping_get_lp_count(lp_group_name, 0, LP_CONFIG_NM_ROUT, NULL, 1);
    assert(p->num_real_groups > 0);
    if(p->num_real_groups % p->num_routers)
    {
        tw_error(TW_LOC, "\n Config error: num_routers specified %d "
                "does not divide num_router per group %d  ",
                p->num_real_groups , p->num_routers);
    }
    p->num_real_groups = p->num_real_groups/p->num_routers;
    
    num_routers_per_mgrp = codes_mapping_get_lp_count (lp_group_name, 1, LP_METHOD_NM_ROUT,
849
            NULL, 0);
850
    /*int num_grp_reps = codes_mapping_get_group_reps(lp_group_name);
851 852 853 854
    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);
855
    */
856 857 858
   r->router_id=mapping_rep_id + mapping_offset;
   r->group_id=r->router_id/p->num_routers;

859 860 861
   r->fwd_events = 0;
   r->rev_events = 0;

862 863 864
   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));
865
   r->link_traffic = (int64_t*)malloc(p->radix * sizeof(int64_t));
866
   r->link_traffic_sample = (int64_t*)malloc(p->radix * sizeof(int64_t));
867 868
   r->cur_hist_num = (int*)malloc(p->radix * sizeof(int));
   r->prev_hist_num = (int*)malloc(p->radix * sizeof(int));
869
   
870 871
   r->vc_occupancy = (int**)malloc(p->radix * sizeof(int*));
   r->in_send_loop = (int*)malloc(p->radix * sizeof(int));
872
   r->pending_msgs = 
873
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
874
   r->pending_msgs_tail = 
875
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
876
   r->queued_msgs = 
877
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
878
   r->queued_msgs_tail = 
879
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
880
   r->queued_count = (int*)malloc(p->radix * sizeof(int));
881 882
   r->last_buf_full = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
   r->busy_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
883
   r->busy_time_sample = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
884

885 886 887 888 889 890
   /* set up for ROSS stats sampling */
   r->link_traffic_ross_sample = (int64_t*)calloc(p->radix, sizeof(int64_t));
   r->busy_time_ross_sample = (tw_stime*)calloc(p->radix, sizeof(tw_stime));
   if (g_st_model_stats)
       lp->model_types->mstat_sz = sizeof(tw_lpid) + (sizeof(int64_t) + sizeof(tw_stime)) * p->radix;

891
   rc_stack_create(&r->st);
892
   for(int i=0; i < p->radix; i++)
893 894
    {
       // Set credit & router occupancy
895 896
    r->last_buf_full[i] = 0.0;
    r->busy_time[i] = 0.0;
897
    r->busy_time_sample[i] = 0.0;
898 899
	r->next_output_available_time[i]=0;
	r->cur_hist_start_time[i] = 0;
900
    r->link_traffic[i]=0;
901
    r->link_traffic_sample[i] = 0;
902 903
	r->cur_hist_num[i] = 0;
	r->prev_hist_num[i] = 0;
904
    r->queued_count[i] = 0;    
905 906
    r->in_send_loop[i] = 0;
    r->vc_occupancy[i] = (int*)malloc(p->num_vcs * sizeof(int));
907
    r->pending_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
908
        sizeof(terminal_message_list*));
909
    r->pending_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
910
        sizeof(terminal_message_list*));
911
    r->queued_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
912
        sizeof(terminal_message_list*));
913
    r->queued_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
914
        sizeof(terminal_message_list*));
915
        for(int j = 0; j < p->num_vcs; j++) {
916 917 918 919 920 921 922 923 924
            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
925
//   printf("\n LP ID %d VC occupancy radix %d Router %d is connected to ", lp->gid, p->radix, r->router_id);
926
#endif 
927 928
   //round the number of global channels to the nearest even number
#if USE_DIRECT_SCHEME
929
       int first = r->router_id % p->num_routers;
930
       for(int i=0; i < p->num_global_channels; i++)
931
        {
932
            int target_grp = first;
933 934 935
            if(target_grp == r->group_id) {
                target_grp = p->num_groups - 1;
            }
936
            int my_pos = r->group_id % p->num_routers;
937 938 939 940 941 942 943
            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
944
   int router_offset = (r->router_id % p->num_routers) * 
945 946
    (p->num_global_channels / 2) + 1;
   for(int i=0; i < p->num_global_channels; i++)
947 948 949 950 951 952 953 954 955 956 957 958
    {
      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)
         {
959
           r->global_channel[i]=p->total_routers+r->global_channel[i]; 
960 961 962
	 }
#if DEBUG == 1
    printf("\n channel %d ", r->global_channel[i]);
963
#endif 
964 965 966 967 968 969 970
    }
#endif

#if DEBUG == 1
   printf("\n");
#endif
   return;
971
}	
972 973 974


/* dragonfly packet event , generates a dragonfly packet on the compute node */
975 976 977 978 979 980 981 982 983 984
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)
985
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
986 987
    (void)message_offset;
    (void)sched_params;
988 989 990 991 992
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

993
    xfer_to_nic_time = codes_local_latency(sender); 
994 995
    //e_new = tw_event_new(sender->gid, xfer_to_nic_time+offset, sender);
    //msg = tw_event_data(e_new);
996 997
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
998 999
    strcpy(msg->category, req->category);
    msg->final_dest_gid = req->final_dest_lp;
1000
    msg->total_size = req->msg_size;
1001
    msg->sender_lp=req->src_lp;
1002
    msg->sender_mn_lp = sender->gid;
1003
    msg->packet_size = packet_size;
1004
    msg->travel_start_time = tw_now(sender);
1005 1006 1007
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
1008
    msg->dest_terminal_id = req->dest_mn_lp;
1009
    msg->message_id = req->msg_id;
1010 1011
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;
1012
    msg->magic = terminal_magic_num; 
1013 1014
    msg->msg_start_time = req->msg_start_time;

1015 1016
    if(is_last_pckt) /* Its the last packet so pass in remote and local event information*/
      {
1017
	if(req->remote_event_size > 0)
1018
	 {
1019 1020 1021
		msg->remote_event_size_bytes = req->remote_event_size;
		memcpy(tmp_ptr, remote_event, req->remote_event_size);
		tmp_ptr += req->remote_event_size;
1022
	}
1023
	if(req->self_event_size > 0)
1024
	{
1025 1026 1027
		msg->local_event_size_bytes = req->self_event_size;
		memcpy(tmp_ptr, self_event, req->self_event_size);
		tmp_ptr += req->self_event_size;
1028 1029
	}
     }
1030
	   //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);
1031
    tw_event_send(e_new);
1032
    return xfer_to_nic_time;
1033 1034 1035 1036 1037 1038 1039 1040 1041
}

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

1042
/* given two group IDs, find the router of the src_gid that connects to the dest_gid*/
1043
tw_lpid getRouterFromGroupID(int dest_gid, 
1044
		    int src_gid,
1045
		    int num_routers,
1046
            int total_groups)
1047
{
1048 1049 1050 1051 1052 1053 1054
#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
1055 1056 1057
  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;
1058
  
1059 1060
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
1061
  
1062 1063
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
1064
  
1065
  offset=(offset - 1) % (half_channel * num_routers);
1066 1067

  // If the destination router is in the same group
1068
  tw_lpid router_id;
1069 1070 1071 1072 1073 1074 1075

  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;
1076
#endif
1077
}	
1078 1079

/*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*/
Nikhil's avatar
Nikhil committed
1080
static void router_credit_send(router_state * s, terminal_message * msg, 
1081
  tw_lp * lp, int sq) {
1082 1083 1084 1085
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

1086
  int dest = 0,  type = R_BUFFER;
1087
  int is_terminal = 0;
1088

1089
  const dragonfly_param *p = s->params;
1090
 
1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
  // 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");

1103
  ts = g_tw_lookahead + p->credit_delay +  tw_rand_unif(lp->rng);
1104
	
1105
  if (is_terminal) {
1106
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY, 
1107 1108 1109
      (void**)&buf_msg, NULL);
    buf_msg->magic = terminal_magic_num;
  } else {
1110 1111
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY_ROUTER,
            (void**)&buf_msg, NULL);
1112 1113
    buf_msg->magic = router_magic_num;
  }
1114
 
1115 1116 1117 1118
  if(sq == -1) {
    buf_msg->vc_index = msg->vc_index;
    buf_msg->output_chan = msg->output_chan;
  } else {
1119
    buf_msg->vc_index = msg->saved_vc;
1120 1121
    buf_msg->output_chan = msg->saved_channel;
  }
1122
  
1123
  buf_msg->type = type;
1124

1125 1126
  tw_event_send(buf_e);
  return;
1127 1128
}

Nikhil's avatar
Nikhil committed
1129
static void packet_generate_rc(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
1130
{
1131 1132
   s->packet_gen--;
   packet_gen--;
1133
   
1134
   tw_rand_reverse_unif(lp->rng);
1135

1136 1137 1138
   int num_chunks = msg->packet_size/s->params->chunk_size;
   if(msg->packet_size % s->params->chunk_size)
       num_chunks++;
1139

1140
   if(!num_chunks)
1141
       num_chunks = 1;
1142

1143 1144
   int i;
   for(i = 0; i < num_chunks; i++) {
1145
        delete_terminal_message_list(return_tail(s->terminal_msgs, 
1146
          s->terminal_msgs_tail, 0));
1147
        s->terminal_length -= s->params->chunk_size;
1148 1149
   }
    if(bf->c5) {
1150
        codes_local_latency_reverse(lp);
1151 1152
        s->in_send_loop = 0;
    }
1153 1154
      if(bf->c11) {
        s->issueIdle = 0;
1155
        s->last_buf_full = msg->saved_busy_time;
1156
      }
1157 1158
     struct mn_stats* stat;
     stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
1159 1160 1161 1162
     stat->send_count--;
     stat->send_bytes -= msg->packet_size;
     stat->send_time -= (1/s->params->cn_bandwidth) * msg->packet_size;
}
1163

1164
/* generates packet at the current dragonfly compute node */
Nikhil's avatar
Nikhil committed
1165
static void packet_generate(terminal_state * s, tw_bf * bf, terminal_message * msg, 
1166
  tw_lp * lp) {
1167 1168
  packet_gen++;
  s->packet_gen++;
1169

1170
  tw_stime ts, nic_ts;
1171

1172
  assert(lp->gid != msg->dest_terminal_id);
1173
  const dragonfly_param *p = s->params;
1174

Jonathan Jenkins's avatar
Jonathan Jenkins committed
1175 1176
  int total_event_size;
  uint64_t num_chunks = msg->packet_size / p->chunk_size;
1177
  if (msg->packet_size % s->params->chunk_size) 
1178
      num_chunks++;
1179 1180