dragonfly.c 120 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
   struct dfly_cn_sample ross_sample;
282
};
283

284 285 286 287 288
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
289
  T_SEND,
290
  T_BUFFER,
291 292
  R_SEND,
  R_ARRIVE,
293 294 295 296
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319
};
/* 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
{
320 321
    MINIMAL = 0,
    NON_MINIMAL,
322 323
    ADAPTIVE,
    PROG_ADAPTIVE
324 325 326 327 328
};

struct router_state
{
   unsigned int router_id;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
329
   int group_id;
330 331 332
   int op_arr_size;
   int max_arr_size;

333 334
   int* global_channel; 
   
335
   tw_stime* next_output_available_time;
336
   tw_stime* cur_hist_start_time;
337
   tw_stime** last_buf_full;
338

339
   tw_stime* busy_time;
340
   tw_stime* busy_time_sample;
341

342 343 344 345 346
   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;
347
   int *queued_count;
348
   struct rc_stack * st;
349
   
350
   int** vc_occupancy;
351
   int64_t* link_traffic;
352
   int64_t * link_traffic_sample;
353 354

   const char * anno;
355
   dragonfly_param *params;
356 357 358

   int* prev_hist_num;
   int* cur_hist_num;
359
   
360
   char output_buf[4096];
361
   char output_buf2[4096];
362 363

   struct dfly_router_sample * rsamples;
364
   
365 366
   long fwd_events;
   long rev_events;
367 368 369 370

   /* following used for ROSS model-level stats collection */
   tw_stime* busy_time_ross_sample;
   int64_t * link_traffic_ross_sample;
371
   struct dfly_router_sample ross_rsample;
372 373
};

374
/* ROSS Instrumentation Support */
375 376 377
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);
378 379 380 381
static void ross_dragonfly_rsample_fn(router_state * s, tw_bf * bf, tw_lp * lp, struct dfly_router_sample *sample);
static void ross_dragonfly_rsample_rc_fn(router_state * s, tw_bf * bf, tw_lp * lp, struct dfly_router_sample *sample);
static void ross_dragonfly_sample_fn(terminal_state * s, tw_bf * bf, tw_lp * lp, struct dfly_cn_sample *sample);
static void ross_dragonfly_sample_rc_fn(terminal_state * s, tw_bf * bf, tw_lp * lp, struct dfly_cn_sample *sample);
382 383

st_model_types dragonfly_model_types[] = {
384
    {(ev_trace_f) dragonfly_event_collect,
385 386
     sizeof(int),
     (model_stat_f) dragonfly_model_stat_collect,
387 388 389 390
     sizeof(tw_lpid) + sizeof(long) * 2 + sizeof(double) + sizeof(tw_stime) *2,
     (sample_event_f) ross_dragonfly_sample_fn,
     (sample_revent_f) ross_dragonfly_sample_rc_fn,
     sizeof(struct dfly_cn_sample) } , 
391
    {(ev_trace_f) dragonfly_event_collect,
392 393
     sizeof(int),
     (model_stat_f) dfly_router_model_stat_collect,
394 395 396 397
     0, //updated in router_setup() since it's based on the radix
     (sample_event_f) ross_dragonfly_rsample_fn,
     (sample_revent_f) ross_dragonfly_rsample_rc_fn,
     0 } , //updated in router_setup() since it's based on the radix    
398
    {NULL, 0, NULL, 0, NULL, NULL, 0}
399
};
400
/* End of ROSS model stats collection */
401 402 403

static short routing = MINIMAL;

404 405
static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
406
static tw_stime         max_collective = 0;
407

408

409 410
static long long       total_hops = 0;
static long long       N_finished_packets = 0;
411 412 413
static long long       total_msg_sz = 0;
static long long       N_finished_msgs = 0;
static long long       N_finished_chunks = 0;
414

415 416 417 418
static int dragonfly_rank_hash_compare(
        void *key, struct qhash_head *link)
{
    struct dfly_hash_key *message_key = (struct dfly_hash_key *)key;
419
    struct dfly_qhash_entry *tmp = NULL;
420 421

    tmp = qhash_entry(link, struct dfly_qhash_entry, hash_link);
422
    
423 424 425 426 427 428
    if (tmp->key.message_id == message_key->message_id
            && tmp->key.sender_id == message_key->sender_id)
        return 1;

    return 0;
}
429 430
static int dragonfly_hash_func(void *k, int table_size)
{
431
    struct dfly_hash_key *tmp = (struct dfly_hash_key *)k;
432
    //uint32_t pc = 0, pb = 0;	
433 434
    //bj_hashlittle2(tmp, sizeof(*tmp), &pc, &pb);
    uint64_t key = (~tmp->message_id) + (tmp->message_id << 18);
435 436
    key = key * 21;
    key = ~key ^ (tmp->sender_id >> 4);
437
    key = key * tmp->sender_id; 
438 439
    return (int)(key & (table_size - 1));
    //return (int)(pc % (table_size - 1));
440 441
}

442 443 444 445 446 447 448
/* 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);
449
    /* GiB to s */
450 451 452 453 454 455
    time = time / GB_p_s;
    /* s to ns */
    time = time * 1000.0 * 1000.0 * 1000.0;

    return(time);
}
456

457 458
/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
459
{
460 461
	   return sizeof(terminal_message);
}
462

463 464
static void free_tmp(void * ptr)
{
465
    struct dfly_qhash_entry * dfly = ptr; 
466 467 468 469 470 471
    
    if(dfly->remote_event_data)
        free(dfly->remote_event_data);
   
    if(dfly)
        free(dfly);
472
}
473
static void append_to_terminal_message_list(  
474 475
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
476
        int index, 
477 478 479 480 481 482
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thisq[index] = msg;
    } else {
        thistail[index]->next = msg;
        msg->prev = thistail[index];
483
    } 
484
    thistail[index] = msg;
485 486
}

487
static void prepend_to_terminal_message_list(  
488 489
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
490
        int index, 
491 492 493 494 495 496
        terminal_message_list *msg) {
    if(thisq[index] == NULL) {
        thistail[index] = msg;
    } else {
        thisq[index]->prev = msg;
        msg->next = thisq[index];
497
    } 
498 499
    thisq[index] = msg;
}
500

501 502 503 504 505 506 507 508 509 510 511 512 513 514 515
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;
516 517
}

518 519 520 521 522
static terminal_message_list* return_tail(
        terminal_message_list ** thisq,
        terminal_message_list ** thistail,
        int index) {
    terminal_message_list *tail = thistail[index];
523
    assert(tail);
524 525 526 527 528 529 530 531 532
    if(tail->prev != NULL) {
        tail->prev->next = NULL;
        thistail[index] = tail->prev;
        tail->prev = NULL;
    } else {
        thistail[index] = NULL;
        thisq[index] = NULL;
    }
    return tail;
533 534
}

535
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
536 537 538 539 540 541
    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;
542 543
    // shorthand
    dragonfly_param *p = params;
544

545
    int rc = configuration_get_value_int(&config, "PARAMS", "num_routers", anno,
546
            &p->num_routers);
547
    if(rc) {
548 549 550 551 552
        p->num_routers = 4;
        fprintf(stderr, "Number of dimensions not specified, setting to %d\n",
                p->num_routers);
    }

553
    p->num_vcs = 3;
554

555 556
    rc = configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
    if(rc) {
557 558 559 560
        p->local_vc_size = 1024;
        fprintf(stderr, "Buffer size of local channels not specified, setting to %d\n", p->local_vc_size);
    }

561 562
    rc = configuration_get_value_int(&config, "PARAMS", "global_vc_size", anno, &p->global_vc_size);
    if(rc) {
563 564 565 566
        p->global_vc_size = 2048;
        fprintf(stderr, "Buffer size of global channels not specified, setting to %d\n", p->global_vc_size);
    }

567 568
    rc = configuration_get_value_int(&config, "PARAMS", "cn_vc_size", anno, &p->cn_vc_size);
    if(rc) {
569 570 571 572
        p->cn_vc_size = 1024;
        fprintf(stderr, "Buffer size of compute node channels not specified, setting to %d\n", p->cn_vc_size);
    }

573 574
    rc = configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
    if(rc) {
575
        p->chunk_size = 512;
576
        fprintf(stderr, "Chunk size for packets is specified, setting to %d\n", p->chunk_size);
577 578
    }

579 580
    rc = configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
    if(rc) {
581 582 583 584
        p->local_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of local channels not specified, setting to %lf\n", p->local_bandwidth);
    }

585 586
    rc = configuration_get_value_double(&config, "PARAMS", "global_bandwidth", anno, &p->global_bandwidth);
    if(rc) {
587 588 589 590
        p->global_bandwidth = 4.7;
        fprintf(stderr, "Bandwidth of global channels not specified, setting to %lf\n", p->global_bandwidth);
    }

591 592
    rc = configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", anno, &p->cn_bandwidth);
    if(rc) {
593 594 595 596
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }

597 598 599 600
    p->router_delay = 50;
    configuration_get_value_double(&config, "PARAMS", "router_delay", anno,
            &p->router_delay);

601
    configuration_get_value(&config, "PARAMS", "cn_sample_file", anno, dfly_cn_sample_file,
602
            MAX_NAME_LENGTH);
603
    configuration_get_value(&config, "PARAMS", "rt_sample_file", anno, dfly_rtr_sample_file,
604
            MAX_NAME_LENGTH);
605
    
606 607
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
608
            MAX_NAME_LENGTH);
609 610
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
611
    else if(strcmp(routing_str, "nonminimal")==0 || 
612
            strcmp(routing_str,"non-minimal")==0)
613 614 615 616 617
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
618 619
    else
    {
620
        fprintf(stderr, 
621
                "No routing protocol specified, setting to minimal routing\n");
622
        routing = -1;
623 624 625 626 627 628
    }

    // 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;
629
    p->radix = (p->num_routers + p->num_global_channels + p->num_cn);
630
    p->total_routers = p->num_groups * p->num_routers;
631
    p->total_terminals = p->total_routers * p->num_cn;
632
    int rank;
633
    MPI_Comm_rank(MPI_COMM_CODES, &rank);
634 635 636 637 638
    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);
    }
639
    
640 641 642
    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);
643
    p->credit_delay = bytes_to_ns(CREDIT_SZ, p->local_bandwidth); //assume 8 bytes packet
644 645
}

646
static void dragonfly_configure(){
647
    anno_map = codes_mapping_get_lp_anno_map(LP_CONFIG_NM_TERM);
648 649
    assert(anno_map);
    num_params = anno_map->num_annos + (anno_map->has_unanno_lp > 0);
650
    all_params = malloc(num_params * sizeof(*all_params));
651

Jonathan Jenkins's avatar
Jonathan Jenkins committed
652
    for (int i = 0; i < anno_map->num_annos; i++){
653
        const char * anno = anno_map->annotations[i].ptr;
654 655 656 657 658
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
659 660 661
#ifdef ENABLE_CORTEX
	model_net_topology = dragonfly_cortex_topology;
#endif
662 663 664 665 666
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
667 668
   long long avg_hops, total_finished_packets, total_finished_chunks;
   long long total_finished_msgs, final_msg_sz;
669
   tw_stime avg_time, max_time;
670
   int total_minimal_packets, total_nonmin_packets;
671
   long total_gen, total_fin;
672

673 674 675 676 677 678 679
   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);
680
   
681 682
   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);
683
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
684
    {
685 686
	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);
687
    }
688

689 690
   /* print statistics */
   if(!g_tw_mynode)
691 692
   {	
      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", 
693
              (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);
694
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
695
              printf("\n ADAPTIVE ROUTING STATS: %d chunks routed minimally %d chunks routed non-minimally completed packets %lld \n", 
696
                      total_minimal_packets, total_nonmin_packets, total_finished_chunks);
697
 
698
      printf("\n Total packets generated %ld finished %ld \n", total_gen, total_fin);
699
   }
700 701
   return;
}
702

Nikhil's avatar
Nikhil committed
703
static void dragonfly_collective_init(terminal_state * s,
704 705
           		   tw_lp * lp)
{
706 707 708
    // 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);
709
    int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
710
            NULL, 1);
711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738
    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
739
        if(next_child < ((tw_lpid)num_lps * (tw_lpid)num_reps))
740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760
        {
            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
}

761
/* initialize a dragonfly compute node terminal */
Nikhil's avatar
Nikhil committed
762
static void 
763
terminal_init( terminal_state * s, 
764 765
	       tw_lp * lp )
{
766 767 768
    s->packet_gen = 0;
    s->packet_fin = 0;

769
    
770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786
    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];
    }

787 788
   //int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM_TERM,
   //        s->anno, 0);
789

790
   s->terminal_id = codes_mapping_get_lp_relative_id(lp->gid, 0, 0);  
791
   
792
   s->router_id=(int)s->terminal_id / s->params->num_cn;
793 794
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;
795
   
796
   s->finished_msgs = 0;
797 798 799
   s->finished_chunks = 0;
   s->finished_packets = 0;
   s->total_time = 0.0;
800
   s->total_msg_size = 0;
801
   s->num_vcs = 1;
802

803
   s->last_buf_full = (tw_stime*)malloc(s->num_vcs * sizeof(tw_stime));
804 805
   s->busy_time = 0.0;

806 807 808
   s->fwd_events = 0;
   s->rev_events = 0;

809
   rc_stack_create(&s->st);
810 811 812 813
   s->vc_occupancy = (int*)malloc(s->num_vcs * sizeof(int));

   for( i = 0; i < s->num_vcs; i++ )
    {
814
      s->last_buf_full[i] = 0.0;
815 816 817
      s->vc_occupancy[i]=0;
    }

818

819
   s->rank_tbl = NULL;
820
   s->terminal_msgs = 
821
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
822
   s->terminal_msgs_tail = 
823 824 825
       (terminal_message_list**)malloc(1*sizeof(terminal_message_list*));
   s->terminal_msgs[0] = NULL;
   s->terminal_msgs_tail[0] = NULL;
826
   s->terminal_length = 0;
827
   s->in_send_loop = 0;
828
   s->issueIdle = 0;
829

830 831 832 833 834 835 836
   /* 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;

837 838 839 840
   dragonfly_collective_init(s, lp);
   return;
}

841
/* sets up the router virtual channels, global channels, 
842
 * local channels, compute node channels */
Nikhil's avatar
Nikhil committed
843
static void router_setup(router_state * r, tw_lp * lp)
844
{
845
    
846 847 848 849 850 851 852 853 854 855 856 857 858
    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];
    }

859 860 861 862 863 864 865 866 867 868 869 870
    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,
871
            NULL, 0);
872
    /*int num_grp_reps = codes_mapping_get_group_reps(lp_group_name);
873 874 875 876
    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);
877
    */
878 879 880
   r->router_id=mapping_rep_id + mapping_offset;
   r->group_id=r->router_id/p->num_routers;

881 882
   r->fwd_events = 0;
   r->rev_events = 0;
883 884
   r->ross_rsample.fwd_events = 0;
   r->ross_rsample.rev_events = 0;
885

886 887 888
   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));
889
   r->link_traffic = (int64_t*)malloc(p->radix * sizeof(int64_t));
890
   r->link_traffic_sample = (int64_t*)malloc(p->radix * sizeof(int64_t));
891 892
   r->cur_hist_num = (int*)malloc(p->radix * sizeof(int));
   r->prev_hist_num = (int*)malloc(p->radix * sizeof(int));
893
   
894 895
   r->vc_occupancy = (int**)malloc(p->radix * sizeof(int*));
   r->in_send_loop = (int*)malloc(p->radix * sizeof(int));
896
   r->pending_msgs = 
897
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
898
   r->pending_msgs_tail = 
899
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
900
   r->queued_msgs = 
901
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
902
   r->queued_msgs_tail = 
903
    (terminal_message_list***)malloc(p->radix * sizeof(terminal_message_list**));
904
   r->queued_count = (int*)malloc(p->radix * sizeof(int));
905
   r->last_buf_full = (tw_stime**)malloc(p->radix * sizeof(tw_stime*));
906
   r->busy_time = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
907
   r->busy_time_sample = (tw_stime*)malloc(p->radix * sizeof(tw_stime));
908

909 910 911 912 913
   /* 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;
914
   if (g_st_use_analysis_lps && g_st_model_stats)
915 916 917
       lp->model_types->sample_struct_sz = sizeof(struct dfly_router_sample) + (sizeof(tw_stime) + sizeof(int64_t)) * p->radix;
   r->ross_rsample.busy_time = (tw_stime*)calloc(p->radix, sizeof(tw_stime));
   r->ross_rsample.link_traffic_sample = (int64_t*)calloc(p->radix, sizeof(int64_t));
918

919
   rc_stack_create(&r->st);
920
   for(int i=0; i < p->radix; i++)
921 922
    {
       // Set credit & router occupancy
923
    r->busy_time[i] = 0.0;
924
    r->busy_time_sample[i] = 0.0;
925 926
	r->next_output_available_time[i]=0;
	r->cur_hist_start_time[i] = 0;
927
    r->link_traffic[i]=0;
928
    r->link_traffic_sample[i] = 0;
929 930
	r->cur_hist_num[i] = 0;
	r->prev_hist_num[i] = 0;
931
    r->queued_count[i] = 0;    
932 933
    r->in_send_loop[i] = 0;
    r->vc_occupancy[i] = (int*)malloc(p->num_vcs * sizeof(int));
934
    r->last_buf_full[i] = (tw_stime*)malloc(p->num_vcs * sizeof(tw_stime));
935
    r->pending_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
936
        sizeof(terminal_message_list*));
937
    r->pending_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
938
        sizeof(terminal_message_list*));
939
    r->queued_msgs[i] = (terminal_message_list**)malloc(p->num_vcs * 
940
        sizeof(terminal_message_list*));
941
    r->queued_msgs_tail[i] = (terminal_message_list**)malloc(p->num_vcs * 
942
        sizeof(terminal_message_list*));
943
        for(int j = 0; j < p->num_vcs; j++) {
944
            r->last_buf_full[i][j] = 0.0;
945 946 947 948 949 950 951 952 953
            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
954
//   printf("\n LP ID %d VC occupancy radix %d Router %d is connected to ", lp->gid, p->radix, r->router_id);
955
#endif 
956 957
   //round the number of global channels to the nearest even number
#if USE_DIRECT_SCHEME
958
       int first = r->router_id % p->num_routers;
959
       for(int i=0; i < p->num_global_channels; i++)
960
        {
961
            int target_grp = first;
962 963 964
            if(target_grp == r->group_id) {
                target_grp = p->num_groups - 1;
            }
965
            int my_pos = r->group_id % p->num_routers;
966 967 968 969 970 971 972
            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
973
   int router_offset = (r->router_id % p->num_routers) * 
974 975
    (p->num_global_channels / 2) + 1;
   for(int i=0; i < p->num_global_channels; i++)
976 977 978 979 980 981 982 983 984 985 986 987
    {
      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)
         {
988
           r->global_channel[i]=p->total_routers+r->global_channel[i]; 
989 990 991
	 }
#if DEBUG == 1
    printf("\n channel %d ", r->global_channel[i]);
992
#endif 
993 994 995 996 997 998 999
    }
#endif

#if DEBUG == 1
   printf("\n");
#endif
   return;
1000
}	
1001 1002 1003


/* dragonfly packet event , generates a dragonfly packet on the compute node */
1004 1005 1006 1007 1008 1009 1010 1011 1012 1013
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)
1014
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
1015 1016
    (void)message_offset;
    (void)sched_params;
1017 1018 1019 1020 1021
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

1022
    xfer_to_nic_time = codes_local_latency(sender); 
1023 1024
    //e_new = tw_event_new(sender->gid, xfer_to_nic_time+offset, sender);
    //msg = tw_event_data(e_new);
1025 1026
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
1027 1028
    strcpy(msg->category, req->category);
    msg->final_dest_gid = req->final_dest_lp;
1029
    msg->total_size = req->msg_size;
1030
    msg->sender_lp=req->src_lp;
1031
    msg->sender_mn_lp = sender->gid;
1032
    msg->packet_size = packet_size;
1033
    msg->travel_start_time = tw_now(sender);
1034 1035 1036
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
1037
    msg->dest_terminal_id = req->dest_mn_lp;
1038
    msg->message_id = req->msg_id;
1039 1040
    msg->is_pull = req->is_pull;
    msg->pull_size = req->pull_size;
1041
    msg->magic = terminal_magic_num; 
1042 1043
    msg->msg_start_time = req->msg_start_time;

1044 1045
    if(is_last_pckt) /* Its the last packet so pass in remote and local event information*/
      {
1046
	if(req->remote_event_size > 0)
1047
	 {
1048 1049 1050
		msg->remote_event_size_bytes = req->remote_event_size;
		memcpy(tmp_ptr, remote_event, req->remote_event_size);
		tmp_ptr += req->remote_event_size;
1051
	}
1052
	if(req->self_event_size > 0)
1053
	{
1054 1055 1056
		msg->local_event_size_bytes = req->self_event_size;
		memcpy(tmp_ptr, self_event, req->self_event_size);
		tmp_ptr += req->self_event_size;
1057 1058
	}
     }
1059
	   //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);
1060
    tw_event_send(e_new);
1061
    return xfer_to_nic_time;
1062 1063 1064 1065 1066 1067 1068 1069 1070
}

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

1071
/* given two group IDs, find the router of the src_gid that connects to the dest_gid*/
1072
tw_lpid getRouterFromGroupID(int dest_gid, 
1073
		    int src_gid,
1074
		    int num_routers,
1075
            int total_groups)
1076
{
1077 1078 1079 1080 1081 1082 1083
#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
1084 1085 1086
  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;
1087
  
1088 1089
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
1090
  
1091 1092
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
1093
  
1094
  offset=(offset - 1) % (half_channel * num_routers);
1095 1096

  // If the destination router is in the same group
1097
  tw_lpid router_id;
1098 1099 1100 1101 1102 1103 1104

  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;
1105
#endif
1106
}	
1107 1108

/*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
1109
static void router_credit_send(router_state * s, terminal_message * msg, 
1110
  tw_lp * lp, int sq) {
1111 1112 1113 1114
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

1115
  int dest = 0,  type = R_BUFFER;
1116
  int is_terminal = 0;
1117

1118
  const dragonfly_param *p = s->params;
1119
 
1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131
  // 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");

1132
  ts = g_tw_lookahead + p->credit_delay +  tw_rand_unif(lp->rng);
1133
	
1134
  if (is_terminal) {
1135
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY, 
1136 1137 1138
      (void**)&buf_msg, NULL);
    buf_msg->magic = terminal_magic_num;
  } else {
1139 1140
    buf_e = model_net_method_event_new(dest, ts, lp, DRAGONFLY_ROUTER,
            (void**)&buf_msg, NULL);
1141 1142
    buf_msg->magic = router_magic_num;
  }
1143
 
1144 1145 1146 1147
  if(sq == -1) {
    buf_msg->vc_index = msg->vc_index;
    buf_msg->output_chan = msg->output_chan;
  } else {
1148
    buf_msg->vc_index = msg->saved_vc;
1149 1150
    buf_msg->output_chan = msg->saved_channel;
  }
1151
  
1152
  buf_msg->type = type;
1153

1154 1155
  tw_event_send(buf_e);
  return;
1156 1157
}

Nikhil's avatar
Nikhil committed
1158
static void packet_generate_rc(terminal_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)