dragonfly.c 115 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 28 29
#ifdef ENABLE_CORTEX
#include <cortex/cortex.h>
#include <cortex/topology.h>
#endif

30
#define CREDIT_SZ 8
31 32
#define MEAN_PROCESS 1.0

33 34 35
/* collective specific parameters */
#define TREE_DEGREE 4
#define LEVEL_DELAY 1000
36
#define DRAGONFLY_COLLECTIVE_DEBUG 0
37 38 39
#define NUM_COLLECTIVES  1
#define COLLECTIVE_COMPUTATION_DELAY 5700
#define DRAGONFLY_FAN_OUT_DELAY 20.0
40
#define WINDOW_LENGTH 0
41
#define DFLY_HASH_TABLE_SIZE 4999
42

43
// debugging parameters
44 45
#define TRACK -1
#define TRACK_PKT -1
46
#define TRACK_MSG -1
47
#define PRINT_ROUTER_TABLE 1
48
#define DEBUG 0
49
#define USE_DIRECT_SCHEME 1
50
#define MAX_STATS 65536
51

52 53 54 55
#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])
56

57 58 59 60 61
#ifdef ENABLE_CORTEX
/* This structure is defined at the end of the file */
extern cortex_topology dragonfly_cortex_topology;
#endif

62
int debug_slot_count = 0;
63
long term_ecount, router_ecount, term_rev_ecount, router_rev_ecount;
64
long packet_gen = 0, packet_fin = 0;
65

66 67
static double maxd(double a, double b) { return a < b ? b : a; }

68
/* minimal and non-minimal packet counts for adaptive routing*/
69
static int minimal_count=0, nonmin_count=0;
70
static int num_routers_per_mgrp = 0;
71

72
typedef struct dragonfly_param dragonfly_param;
73
/* annotation-specific parameters (unannotated entry occurs at the 
74 75 76 77
 * last index) */
static uint64_t                  num_params = 0;
static dragonfly_param         * all_params = NULL;
static const config_anno_map_t * anno_map   = NULL;
78 79

/* global variables for codes mapping */
80
static char lp_group_name[MAX_NAME_LENGTH];
81 82
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

83 84 85 86 87 88
/* router magic number */
int router_magic_num = 0;

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

89 90
FILE * dragonfly_log = NULL;

91
int sample_bytes_written = 0;
92
int sample_rtr_bytes_written = 0;
93

94 95
char dfly_cn_sample_file[MAX_NAME_LENGTH];
char dfly_rtr_sample_file[MAX_NAME_LENGTH];
96

97 98 99 100 101 102 103
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;
};
104

Nikhil's avatar
Nikhil committed
105
static void init_terminal_message_list(terminal_message_list *this, 
106 107 108 109 110 111
    terminal_message *inmsg) {
    this->msg = *inmsg;
    this->event_data = NULL;
    this->next = NULL;
    this->prev = NULL;
}
112

Nikhil's avatar
Nikhil committed
113
static void delete_terminal_message_list(terminal_message_list *this) {
114 115 116
    if(this->event_data != NULL) free(this->event_data);
    free(this);
}
117

118 119 120 121 122 123 124 125 126 127 128 129 130 131
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;
132
    int num_groups;
133
    int num_real_groups;
134 135
    int radix;
    int total_routers;
136
    int total_terminals;
137
    int num_global_channels;
138 139 140 141
    double cn_delay;
    double local_delay;
    double global_delay;
    double credit_delay;
142
    double router_delay;
143 144
};

145 146 147 148 149 150
struct dfly_hash_key
{
    uint64_t message_id;
    tw_lpid sender_id;
};

151 152 153 154
struct dfly_router_sample
{
    tw_lpid router_id;
    tw_stime* busy_time;
155
    int64_t* link_traffic_sample;
156
    tw_stime end_time;
157 158
    long fwd_events;
    long rev_events;
159 160 161
};

struct dfly_cn_sample
162 163 164 165 166 167 168 169
{
   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;
170 171
   long fwd_events;
   long rev_events;
172 173
};

174 175 176 177
struct dfly_qhash_entry
{
   struct dfly_hash_key key;
   char * remote_event_data;
178
   uint64_t num_chunks;
179 180 181 182
   int remote_event_size;
   struct qhash_head hash_link;
};

183 184 185 186 187 188 189 190
/* 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
{
191
   uint64_t packet_counter;
192

193 194 195
   int packet_gen;
   int packet_fin;

196
   // Dragonfly specific parameters
197 198
   unsigned int router_id;
   unsigned int terminal_id;
199 200 201

   // Each terminal will have an input and output channel with the router
   int* vc_occupancy; // NUM_VC
202
   int num_vcs;
203
   tw_stime terminal_available_time;
204 205 206
   terminal_message_list **terminal_msgs;
   terminal_message_list **terminal_msgs_tail;
   int in_send_loop;
207 208 209 210
// 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];
211 212 213
  /* collective init time */
  tw_stime collective_init_time;

214
  /* node ID in the tree */ 
215 216
   tw_lpid node_id;

217
   /* messages sent & received in collectives may get interchanged several times so we have to save the 
218
     origin server information in the node's state */
219 220
   tw_lpid origin_svr; 
  
221 222 223 224 225 226 227 228 229 230 231
  /* 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;

232
   struct rc_stack * st;
233 234
   int issueIdle;
   int terminal_length;
235

236 237 238
   /* to maintain a count of child nodes that have fanned in at the parent during the collective
      fan-in phase*/
   int num_fan_nodes;
239 240

   const char * anno;
241
   dragonfly_param *params;
242

243 244 245
   struct qhash_table *rank_tbl;
   uint64_t rank_tbl_pop;

246
   tw_stime   total_time;
247
   uint64_t total_msg_size;
248
   double total_hops;
249
   long finished_msgs;
250
   long finished_chunks;
251
   long finished_packets;
252

253
   tw_stime * last_buf_full;
254
   tw_stime busy_time;
255
   char output_buf[4096];
256 257
   /* For LP suspend functionality */
   int error_ct;
258 259 260 261 262 263 264 265 266

   /* 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];
267
   struct dfly_cn_sample * sample_stat;
268 269
   int op_arr_size;
   int max_arr_size;
270
   
271 272 273
   /* for logging forward and reverse events */
   long fwd_events;
   long rev_events;
274 275 276 277 278 279 280

   /* 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;
281
};
282

283 284 285 286 287
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
288
  T_SEND,
289
  T_BUFFER,
290 291
  R_SEND,
  R_ARRIVE,
292 293 294 295
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
};
/* 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
{
319 320
    MINIMAL = 0,
    NON_MINIMAL,
321 322
    ADAPTIVE,
    PROG_ADAPTIVE
323 324 325 326 327
};

struct router_state
{
   unsigned int router_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
328
   int group_id;
329 330 331
   int op_arr_size;
   int max_arr_size;

332 333
   int* global_channel; 
   
334
   tw_stime* next_output_available_time;
335
   tw_stime* cur_hist_start_time;
336
   tw_stime** last_buf_full;
337

338
   tw_stime* busy_time;
339
   tw_stime* busy_time_sample;
340

341 342 343 344 345
   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;
346
   int *queued_count;
347
   struct rc_stack * st;
348
   
349
   int** vc_occupancy;
350
   int64_t* link_traffic;
351
   int64_t * link_traffic_sample;
352 353

   const char * anno;
354
   dragonfly_param *params;
355 356 357

   int* prev_hist_num;
   int* cur_hist_num;
358
   
359
   char output_buf[4096];
360
   char output_buf2[4096];
361 362

   struct dfly_router_sample * rsamples;
363
   
364 365
   long fwd_events;
   long rev_events;
366 367 368 369

   /* following used for ROSS model-level stats collection */
   tw_stime* busy_time_ross_sample;
   int64_t * link_traffic_ross_sample;
370 371
};

372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
/* 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
390
    {NULL, 0, NULL, 0, NULL, 0}
391
};
392
/* End of ROSS model stats collection */
393 394 395

static short routing = MINIMAL;

396 397
static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
398
static tw_stime         max_collective = 0;
399

400

401 402
static long long       total_hops = 0;
static long long       N_finished_packets = 0;
403 404 405
static long long       total_msg_sz = 0;
static long long       N_finished_msgs = 0;
static long long       N_finished_chunks = 0;
406

407 408 409 410
static int dragonfly_rank_hash_compare(
        void *key, struct qhash_head *link)
{
    struct dfly_hash_key *message_key = (struct dfly_hash_key *)key;
411
    struct dfly_qhash_entry *tmp = NULL;
412 413

    tmp = qhash_entry(link, struct dfly_qhash_entry, hash_link);
414
    
415 416 417 418 419 420
    if (tmp->key.message_id == message_key->message_id
            && tmp->key.sender_id == message_key->sender_id)
        return 1;

    return 0;
}
421 422
static int dragonfly_hash_func(void *k, int table_size)
{
423
    struct dfly_hash_key *tmp = (struct dfly_hash_key *)k;
424
    //uint32_t pc = 0, pb = 0;	
425 426
    //bj_hashlittle2(tmp, sizeof(*tmp), &pc, &pb);
    uint64_t key = (~tmp->message_id) + (tmp->message_id << 18);
427 428
    key = key * 21;
    key = ~key ^ (tmp->sender_id >> 4);
429
    key = key * tmp->sender_id; 
430 431
    return (int)(key & (table_size - 1));
    //return (int)(pc % (table_size - 1));
432 433
}

434 435 436 437 438 439 440
/* 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);
441
    /* GiB to s */
442 443 444 445 446 447
    time = time / GB_p_s;
    /* s to ns */
    time = time * 1000.0 * 1000.0 * 1000.0;

    return(time);
}
448

449 450
/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
451
{
452 453
	   return sizeof(terminal_message);
}
454

455 456
static void free_tmp(void * ptr)
{
457
    struct dfly_qhash_entry * dfly = ptr; 
458 459 460 461 462 463
    
    if(dfly->remote_event_data)
        free(dfly->remote_event_data);
   
    if(dfly)
        free(dfly);
464
}
465
static void append_to_terminal_message_list(  
466 467
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
468
        int index, 
469 470 471 472 473 474
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thisq[index] = msg;
    } else {
        thistail[index]->next = msg;
        msg->prev = thistail[index];
475
    } 
476
    thistail[index] = msg;
477 478
}

479
static void prepend_to_terminal_message_list(  
480 481
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
482
        int index, 
483 484 485 486 487 488
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thistail[index] = msg;
    } else {
        thisq[index]->prev = msg;
        msg->next = thisq[index];
489
    } 
490 491
    thisq[index] = msg;
}
492

493 494 495 496 497 498 499 500 501 502 503 504 505 506 507
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;
508 509
}

510 511 512 513 514
static terminal_message_list* return_tail(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index) {
    terminal_message_list *tail = thistail[index];
515
    assert(tail);
516 517 518 519 520 521 522 523 524
    if(tail->prev != NULL) {
        tail->prev->next = NULL;
        thistail[index] = tail->prev;
        tail->prev = NULL;
    } else {
        thistail[index] = NULL;
        thisq[index] = NULL;
    }
    return tail;
525 526
}

527
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
528 529 530 531 532 533
    uint32_t h1 = 0, h2 = 0; 
    bj_hashlittle2(LP_METHOD_NM_TERM, strlen(LP_METHOD_NM_TERM), &h1, &h2);
    terminal_magic_num = h1 + h2;
    
    bj_hashlittle2(LP_METHOD_NM_ROUT, strlen(LP_METHOD_NM_ROUT), &h1, &h2);
    router_magic_num = h1 + h2;
534 535
    // shorthand
    dragonfly_param *p = params;
536

537
    int rc = configuration_get_value_int(&config, "PARAMS", "num_routers", anno,
538
            &p->num_routers);
539
    if(rc) {
540 541 542 543 544
        p->num_routers = 4;
        fprintf(stderr, "Number of dimensions not specified, setting to %d\n",
                p->num_routers);
    }

545
    p->num_vcs = 3;
546

547 548
    rc = configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
    if(rc) {
549 550 551 552
        p->local_vc_size = 1024;
        fprintf(stderr, "Buffer size of local channels not specified, setting to %d\n", p->local_vc_size);
    }

553 554
    rc = configuration_get_value_int(&config, "PARAMS", "global_vc_size", anno, &p->global_vc_size);
    if(rc) {
555 556 557 558
        p->global_vc_size = 2048;
        fprintf(stderr, "Buffer size of global channels not specified, setting to %d\n", p->global_vc_size);
    }

559 560
    rc = configuration_get_value_int(&config, "PARAMS", "cn_vc_size", anno, &p->cn_vc_size);
    if(rc) {
561 562 563 564
        p->cn_vc_size = 1024;
        fprintf(stderr, "Buffer size of compute node channels not specified, setting to %d\n", p->cn_vc_size);
    }

565 566
    rc = configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(rc) {
567
        p->chunk_size = 512;
568
        fprintf(stderr, "Chunk size for packets is specified, setting to %d\n", p->chunk_size);
569 570
    }

571 572
    rc = configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
    if(rc) {
573 574 575 576
        p->local_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of local channels not specified, setting to %lf\n", p->local_bandwidth);
    }

577 578
    rc = configuration_get_value_double(&config, "PARAMS", "global_bandwidth", anno, &p->global_bandwidth);
    if(rc) {
579 580 581 582
        p->global_bandwidth = 4.7;
        fprintf(stderr, "Bandwidth of global channels not specified, setting to %lf\n", p->global_bandwidth);
    }

583 584
    rc = configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", anno, &p->cn_bandwidth);
    if(rc) {
585 586 587 588
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }

589 590 591 592
    p->router_delay = 50;
    configuration_get_value_double(&config, "PARAMS", "router_delay", anno,
            &p->router_delay);

593
    configuration_get_value(&config, "PARAMS", "cn_sample_file", anno, dfly_cn_sample_file,
594
            MAX_NAME_LENGTH);
595
    configuration_get_value(&config, "PARAMS", "rt_sample_file", anno, dfly_rtr_sample_file,
596
            MAX_NAME_LENGTH);
597
    
598 599
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
600
            MAX_NAME_LENGTH);
601 602
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
603
    else if(strcmp(routing_str, "nonminimal")==0 || 
604
            strcmp(routing_str,"non-minimal")==0)
605 606 607 608 609
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
610 611
    else
    {
612
        fprintf(stderr, 
613
                "No routing protocol specified, setting to minimal routing\n");
614
        routing = -1;
615 616 617 618 619 620
    }

    // 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;
621
    p->radix = (p->num_routers + p->num_global_channels + p->num_cn);
622
    p->total_routers = p->num_groups * p->num_routers;
623
    p->total_terminals = p->total_routers * p->num_cn;
624
    int rank;
625
    MPI_Comm_rank(MPI_COMM_CODES, &rank);
626 627 628 629 630
    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);
    }
631
    
632 633 634
    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);
635
    p->credit_delay = bytes_to_ns(CREDIT_SZ, p->local_bandwidth); //assume 8 bytes packet
636 637
}

638
static void dragonfly_configure(){
639
    anno_map = codes_mapping_get_lp_anno_map(LP_CONFIG_NM_TERM);
640 641
    assert(anno_map);
    num_params = anno_map->num_annos + (anno_map->has_unanno_lp > 0);
642
    all_params = malloc(num_params * sizeof(*all_params));
643

Jonathan Jenkins's avatar
Jonathan Jenkins committed
644
    for (int i = 0; i < anno_map->num_annos; i++){
645
        const char * anno = anno_map->annotations[i].ptr;
646 647 648 649 650
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
651 652 653
#ifdef ENABLE_CORTEX
	model_net_topology = dragonfly_cortex_topology;
#endif
654 655 656 657 658
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
659 660
   long long avg_hops, total_finished_packets, total_finished_chunks;
   long long total_finished_msgs, final_msg_sz;
661
   tw_stime avg_time, max_time;
662
   int total_minimal_packets, total_nonmin_packets;
663
   long total_gen, total_fin;
664

665 666 667 668 669 670 671
   MPI_Reduce( &total_hops, &avg_hops, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_CODES);
   MPI_Reduce( &N_finished_packets, &total_finished_packets, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_CODES);
   MPI_Reduce( &N_finished_msgs, &total_finished_msgs, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_CODES);
   MPI_Reduce( &N_finished_chunks, &total_finished_chunks, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_CODES);
   MPI_Reduce( &total_msg_sz, &final_msg_sz, 1, MPI_LONG_LONG, MPI_SUM, 0, MPI_COMM_CODES);
   MPI_Reduce( &dragonfly_total_time, &avg_time, 1,MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_CODES);
   MPI_Reduce( &dragonfly_max_latency, &max_time, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_CODES);
672
   
673 674
   MPI_Reduce( &packet_gen, &total_gen, 1, MPI_LONG, MPI_SUM, 0, MPI_COMM_CODES);
   MPI_Reduce( &packet_fin, &total_fin, 1, MPI_LONG, MPI_SUM, 0, MPI_COMM_CODES);
675
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
676
    {
677 678
	MPI_Reduce(&minimal_count, &total_minimal_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_CODES);
 	MPI_Reduce(&nonmin_count, &total_nonmin_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_CODES);
679
    }
680

681 682
   /* print statistics */
   if(!g_tw_mynode)
683 684
   {	
      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", 
685
              (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);
686
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
687
              printf("\n ADAPTIVE ROUTING STATS: %d chunks routed minimally %d chunks routed non-minimally completed packets %lld \n", 
688
                      total_minimal_packets, total_nonmin_packets, total_finished_chunks);
689
 
690
      printf("\n Total packets generated %ld finished %ld \n", total_gen, total_fin);
691
   }
692 693
   return;
}
694

Nikhil's avatar
Nikhil committed
695
static void dragonfly_collective_init(terminal_state * s,
696 697
           		   tw_lp * lp)
{
698 699 700
    // 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);
701
    int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
702
            NULL, 1);
703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730
    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
731
        if(next_child < ((tw_lpid)num_lps * (tw_lpid)num_reps))
732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752
        {
            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
}

753
/* initialize a dragonfly compute node terminal */
Nikhil's avatar
Nikhil committed
754
static void 
755
terminal_init( terminal_state * s, 
756 757
	       tw_lp * lp )
{
758 759 760
    s->packet_gen = 0;
    s->packet_fin = 0;

761
    
762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778
    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];
    }

779 780
   //int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
   //        s->anno, 0);
781

782
   s->terminal_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);  
783
   
784
   s->router_id=(int)s->terminal_id / s->params->num_cn;
785 786
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;
787
   
788
   s->finished_msgs = 0;
789 790 791
   s->finished_chunks = 0;
   s->finished_packets = 0;
   s->total_time = 0.0;
792
   s->total_msg_size = 0;
793
   s->num_vcs = 1;
794

795
   s->last_buf_full = (tw_stime*)malloc(s->num_vcs * sizeof(tw_stime));
796 797
   s->busy_time = 0.0;

798 799 800
   s->fwd_events = 0;
   s->rev_events = 0;

801
   rc_stack_create(&s->st);
802 803 804 805
   s->vc_occupancy = (int*)malloc(s->num_vcs * sizeof(int));

   for( i = 0; i < s->num_vcs; i++ )
    {
806
      s->last_buf_full[i] = 0.0;
807 808 809
      s->vc_occupancy[i]=0;
    }

810

811
   s->rank_tbl = NULL;
812
   s->terminal_msgs = 
813
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
814
   s->terminal_msgs_tail = 
815 816 817
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
   s->terminal_msgs[0] = NULL;
   s->terminal_msgs_tail[0] = NULL;
818
   s->terminal_length = 0;
819
   s->in_send_loop = 0;
820
   s->issueIdle = 0;
821

822 823 824 825 826 827 828
   /* 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;

829 830 831 832
   dragonfly_collective_init(s, lp);
   return;
}

833
/* sets up the router virtual channels, global channels, 
834
 * local channels, compute node channels */
Nikhil's avatar
Nikhil committed
835
static void router_setup(router_state * r, tw_lp * lp)
836
{
837
    
838 839 840 841 842 843 844 845 846 847 848 849 850
    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];
    }

851 852 853 854 855 856 857 858 859 860 861 862
    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,
863
            NULL, 0);
864
    /*int num_grp_reps = codes_mapping_get_group_reps(lp_group_name);
865 866 867 868
    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);
869
    */
870 871 872
   r->router_id=mapping_rep_id + mapping_offset;
   r->group_id=r->router_id/p->num_routers;

873 874 875
   r->fwd_events = 0;
   r->rev_events = 0;

876 877 878
   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));
879
   r->link_traffic = (int64_t*)malloc(p->radix * sizeof(int64_t));
880
   r->link_traffic_sample = (int64_t*)malloc(p->radix * sizeof(int64_t));
881 882
   r->cur_hist_num = (int*)malloc(p->radix * sizeof(int));
   r->prev_hist_num = (int*)malloc(p->radix * sizeof(int));
883
   
884 885
   r->vc_occupancy = (int**)malloc(p->radix * sizeof(int*));
   r->in_send_loop = (int*)malloc(p->radix * sizeof(int));
886
   r->pending_msgs = 
887
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
888
   r->pending_msgs_tail = 
889
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
890
   r->queued_msgs = 
891
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
892
   r->queued_msgs_tail = 
893
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
894
   r->queued_count = (int*)malloc(p->radix * sizeof(int));
895
   r->last_buf_full = (tw_stime**)malloc(p->radix * sizeof(tw_stime*));
896
   r->busy_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
897
   r->busy_time_sample = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
898

899 900 901 902 903 904
   /* 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;

905
   rc_stack_create(&r->st);
906
   for(int i=0; i < p->radix; i++)
907 908
    {
       // Set credit & router occupancy
909
    r->busy_time[i] = 0.0;
910
    r->busy_time_sample[i] = 0.0;
911 912
	r->next_output_available_time[i]=0;
	r->cur_hist_start_time[i] = 0;
913
    r->link_traffic[i]=0;
914
    r->link_traffic_sample[i] = 0;
915 916
	r->cur_hist_num[i] = 0;
	r->prev_hist_num[i] = 0;
917
    r->queued_count[i] = 0;    
918 919
    r->in_send_loop[i] = 0;
    r->vc_occupancy[i] = (int*)malloc(p->num_vcs * sizeof(int));
920
    r->last_buf_full[i] = (tw_stime*)malloc(p->num_vcs * sizeof(tw_stime));
921
    r->pending_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
922
        sizeof(terminal_message_list*));
923
    r->pending_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
924
        sizeof(terminal_message_list*));
925
    r->queued_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
926
        sizeof(terminal_message_list*));
927
    r->queued_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
928
        sizeof(terminal_message_list*));
929
        for(int j = 0; j < p->num_vcs; j++) {
930
            r->last_buf_full[i][j] = 0.0;
931 932 933 934 935 936 937 938 939
            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
940
//   printf("\n LP ID %d VC occupancy radix %d Router %d is connected to ", lp->gid, p->radix, r->router_id);
941
#endif 
942 943
   //round the number of global channels to the nearest even number
#if USE_DIRECT_SCHEME
944
       int first = r->router_id % p->num_routers;
945
       for(int i=0; i < p->num_global_channels; i++)
946
        {
947
            int target_grp = first;
948 949 950
            if(target_grp == r->group_id) {
                target_grp = p->num_groups - 1;
            }
951
            int my_pos = r->group_id % p->num_routers;
952 953 954 955 956 957 958
            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
959
   int router_offset = (r->router_id % p->num_routers) * 
960 961
    (p->num_global_channels / 2) + 1;
   for(int i=0; i < p->num_global_channels; i++)
962 963 964 965 966 967 968 969 970 971 972 973
    {
      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)
         {
974
           r->global_channel[i]=p->total_routers+r->global_channel[i]; 
975 976 977
	 }
#if DEBUG == 1
    printf("\n channel %d ", r->global_channel[i]);
978
#endif 
979 980 981 982 983 984 985
    }
#endif

#if DEBUG == 1
   printf("\n");
#endif
   return;
986
}	
987 988 989


/* dragonfly packet event , generates a dragonfly packet on the compute node */
990 991 992 993 994 995 996 997 998 999
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)
1000
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
1001 1002
    (void)message_offset;
    (void)sched_params;
1003 1004 1005 1006 1007
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

1008
    xfer_to_nic_time = codes_local_latency(sender); 
1009 1010
    //e_new = tw_event_new(sender->gid, xfer_to_nic_time+offset, sender);
    //msg = tw_event_data(e_new);
1011 1012
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
1013 1014
    strcpy(msg->category, req->category);
    msg->final_dest_gid = req->final_dest_lp;
1015
    msg->total_size = req->msg_size;
1016
    msg->sender_lp=req->src_lp;
1017
    msg->sender_mn_lp = sender->gid;
1018
    msg->packet_size = packet_size;
1019
    msg->travel_start_time = tw_now(sender);
1020 1021 1022
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
1023
    msg->dest_terminal_id = req->dest_mn_lp;
1024
    msg->message_id = req->msg_id;
1025 1026
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;
1027
    msg->magic = terminal_magic_num; 
1028 1029
    msg->msg_start_time = req->msg_start_time;

1030 1031
    if(is_last_pckt) /* Its the last packet so pass in remote and local event information*/
      {
1032
	if(req->remote_event_size > 0)
1033
	 {
1034 1035 1036
		msg->remote_event_size_bytes = req->remote_event_size;
		memcpy(tmp_ptr, remote_event, req->remote_event_size);
		tmp_ptr += req->remote_event_size;
1037
	}
1038
	if(req->self_event_size > 0)
1039
	{
1040 1041 1042
		msg->local_event_size_bytes = req->self_event_size;
		memcpy(tmp_ptr, self_event, req->self_event_size);
		tmp_ptr += req->self_event_size;
1043 1044
	}
     }
1045
	   //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);
1046
    tw_event_send(e_new);
1047
    return xfer_to_nic_time;
1048 1049 1050 1051 1052 1053 1054 1055 1056
}

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

1057
/* given two group IDs, find the router of the src_gid that connects to the dest_gid*/
1058
tw_lpid getRouterFromGroupID(int dest_gid, 
1059
		    int src_gid,
1060
		    int num_routers,
1061
            int total_groups)
1062
{
1063 1064 1065 1066 1067 1068 1069
#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
1070 1071 1072
  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;
1073
  
1074 1075
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
1076
  
1077 1078
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
1079
  
1080
  offset=(offset - 1) % (half_channel * num_routers);
1081 1082

  // If the destination router is in the same group
1083
  tw_lpid router_id;
1084 1085 1086 1087 1088 1089 1090

  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;
1091
#endif
1092
}	
1093 1094

/*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
1095
static void router_credit_send(router_state * s, terminal_message * msg, 
1096
  tw_lp * lp, int sq) {
1097 1098 1099 1100
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

1101
  int dest = 0,  type = R_BUFFER;
1102
  int is_terminal = 0;
1103

1104
  const dragonfly_param *p = s->params;
1105
 
1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117
  // 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");

1118
  ts = g_tw_lookahead + p->credit_delay +  tw_rand_unif(lp->rng);
1119
	
1120
  if (is_terminal) {
1121
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY, 
1122 1123 1124
      (void**)&buf_msg, NULL);
    buf_msg->magic = terminal_magic_num;
  } else {
1125 1126
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY_ROUTER,
            (void**)&buf_msg, NULL);
1127 1128
    buf_msg->magic = router_magic_num;
  }
1129
 
1130 1131 1132 1133
  if(sq == -1) {
    buf_msg->vc_index = msg->vc_index;
    buf_msg->output_chan = msg->output_chan;
  } else {
1134
    buf_msg->vc_index = msg->saved_vc;
1135 1136
    buf_msg->output_chan = msg->saved_channel;
  }
1137
  
1138
  buf_msg->type = type;
1139

1140 1141
  tw_event_send(buf_e);
  return;
1142 1143
}

Nikhil's avatar
Nikhil committed
1144
static void packet_generate_rc(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
1145
{
1146 1147
   s->packet_gen--;
   packet_gen--;
1148
   
1149
   tw_rand_reverse_unif(lp->rng);
1150

1151 1152 1153
   int num_chunks = msg->packet_size/s->params->chunk_size;
   if(msg->packet_size % s->params->chunk_size)
       num_chunks++;
1154

1155
   if(!num_chunks)
1156
       num_chunks = 1;
1157

1158 1159
   int i;
   for(i = 0; i < num_chunks; i++) {
1160
        delete_terminal_message_list(return_tail(s->terminal_msgs, 
1161
          s->terminal_msgs_tail, 0));
1162
        s->terminal_length -= s->params->chunk_size;
1163 1164
   }
    if(bf->c5) {
1165
        codes_local_latency_reverse(lp);
1166 1167
        s->in_send_loop = 0;
    }