dragonfly.c 68.5 KB
Newer Older
Philip Carns's avatar
Philip Carns committed
1 2 3 4 5 6
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

7 8 9 10
// Local router ID: 0 --- total_router-1
// Router LP ID 
// Terminal LP ID

11 12 13
#include <ross.h>

#include "codes/codes_mapping.h"
14
#include "codes/jenkins-hash.h"
15 16 17
#include "codes/codes.h"
#include "codes/model-net.h"
#include "codes/model-net-method.h"
18 19
#include "codes/model-net-lp.h"
#include "codes/net/dragonfly.h"
20 21 22 23

#define CREDIT_SIZE 8
#define MEAN_PROCESS 1.0

24 25 26
/* collective specific parameters */
#define TREE_DEGREE 4
#define LEVEL_DELAY 1000
27
#define DRAGONFLY_COLLECTIVE_DEBUG 0
28 29 30
#define NUM_COLLECTIVES  1
#define COLLECTIVE_COMPUTATION_DELAY 5700
#define DRAGONFLY_FAN_OUT_DELAY 20.0
31
#define WINDOW_LENGTH 0
32

33
// debugging parameters
34
#define TRACK -1
35
#define PRINT_ROUTER_TABLE 1
36
#define DEBUG 0
37

38 39 40
#define LP_CONFIG_NM (model_net_lp_config_names[DRAGONFLY])
#define LP_METHOD_NM (model_net_method_names[DRAGONFLY])

41 42
static double maxd(double a, double b) { return a < b ? b : a; }

43 44
// arrival rate
static double MEAN_INTERVAL=200.0;
45 46
// threshold for adaptive routing
static int adaptive_threshold = 10;
47 48

/* minimal and non-minimal packet counts for adaptive routing*/
49
unsigned int minimal_count=0, nonmin_count=0, completed_packets = 0;
50

51 52 53 54 55 56
typedef struct dragonfly_param dragonfly_param;
/* annotation-specific parameters (unannotated entry occurs at the 
 * last index) */
static uint64_t                  num_params = 0;
static dragonfly_param         * all_params = NULL;
static const config_anno_map_t * anno_map   = NULL;
57 58

/* global variables for codes mapping */
59
static char lp_group_name[MAX_NAME_LENGTH];
60 61
static int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;

62 63 64 65 66 67 68 69
/* router magic number */
int router_magic_num = 0;

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

/* number of routers in a mapping group */
static int num_routers_per_mgrp = 0;
70

71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
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;
    int num_groups;
    int radix;
    int total_routers;
89
    int total_terminals;
90 91 92
    int num_global_channels;
};

93 94 95 96 97 98 99 100 101 102 103 104
/* 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
{
   unsigned long long packet_counter;

   // Dragonfly specific parameters
105 106
   tw_lpid router_id;
   tw_lpid terminal_id;
107 108 109 110 111 112 113 114 115 116

   // Each terminal will have an input and output channel with the router
   int* vc_occupancy; // NUM_VC
   int* output_vc_state;
   tw_stime terminal_available_time;
   tw_stime next_credit_available_time;
// 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];
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140
  /* collective init time */
  tw_stime collective_init_time;

  /* node ID in the tree */ 
   tw_lpid node_id;

   /* messages sent & received in collectives may get interchanged several times so we have to save the 
     origin server information in the node's state */
   tw_lpid origin_svr; 
  
  /* parent node ID of the current node */
   tw_lpid parent_node_id;
   /* array of children to be allocated in terminal_init*/
   tw_lpid* children;

   /* children of a node can be less than or equal to the tree degree */
   int num_children;

   short is_root;
   short is_leaf;

   /* to maintain a count of child nodes that have fanned in at the parent during the collective
      fan-in phase*/
   int num_fan_nodes;
141 142 143

   const char * anno;
   const dragonfly_param *params;
144
};
145

146 147 148 149 150 151 152 153 154
/* terminal event type (1-4) */
enum event_t
{
  T_GENERATE=1,
  T_ARRIVE,
  T_SEND,
  T_BUFFER,
  R_SEND,
  R_ARRIVE,
155 156 157 158
  R_BUFFER,
  D_COLLECTIVE_INIT,
  D_COLLECTIVE_FAN_IN,
  D_COLLECTIVE_FAN_OUT
159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
};
/* 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
{
182 183
    MINIMAL = 0,
    NON_MINIMAL,
184 185
    ADAPTIVE,
    PROG_ADAPTIVE
186 187 188 189 190 191 192 193
};

struct router_state
{
   unsigned int router_id;
   unsigned int group_id;
  
   int* global_channel; 
194
   
195 196
   tw_stime* next_output_available_time;
   tw_stime* next_credit_available_time;
197 198
   tw_stime* cur_hist_start_time;
   
199 200
   int* vc_occupancy;
   int* output_vc_state;
201 202 203

   const char * anno;
   const dragonfly_param *params;
204 205 206

   int* prev_hist_num;
   int* cur_hist_num;
207 208 209 210 211 212
};

static short routing = MINIMAL;

static tw_stime         dragonfly_total_time = 0;
static tw_stime         dragonfly_max_latency = 0;
213
static tw_stime         max_collective = 0;
214 215 216 217 218 219 220 221 222 223 224 225 226 227


static long long       total_hops = 0;
static long long       N_finished_packets = 0;

/* returns the dragonfly router lp type for lp registration */
static const tw_lptype* dragonfly_get_router_lp_type(void);

/* returns the dragonfly message size */
static int dragonfly_get_msg_sz(void)
{
	   return sizeof(terminal_message);
}

228 229 230
static void dragonfly_read_config(const char * anno, dragonfly_param *params){
    // shorthand
    dragonfly_param *p = params;
231

232 233 234 235 236 237 238 239 240 241
    configuration_get_value_int(&config, "PARAMS", "num_routers", anno,
            &p->num_routers);
    if(p->num_routers <= 0) {
        p->num_routers = 4;
        fprintf(stderr, "Number of dimensions not specified, setting to %d\n",
                p->num_routers);
    }

    configuration_get_value_int(&config, "PARAMS", "num_vcs", anno,
            &p->num_vcs);
242
    if(!p->num_vcs) {
243 244 245 246 247
        p->num_vcs = 1;
        fprintf(stderr, "Number of virtual channels not specified, setting to %d\n", p->num_vcs);
    }

    configuration_get_value_int(&config, "PARAMS", "local_vc_size", anno, &p->local_vc_size);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
248
    if(p->local_vc_size <= 0) {
249 250 251 252 253
        p->local_vc_size = 1024;
        fprintf(stderr, "Buffer size of local channels not specified, setting to %d\n", p->local_vc_size);
    }

    configuration_get_value_int(&config, "PARAMS", "global_vc_size", anno, &p->global_vc_size);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
254
    if(p->global_vc_size <= 0) {
255 256 257 258 259
        p->global_vc_size = 2048;
        fprintf(stderr, "Buffer size of global channels not specified, setting to %d\n", p->global_vc_size);
    }

    configuration_get_value_int(&config, "PARAMS", "cn_vc_size", anno, &p->cn_vc_size);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
260
    if(p->cn_vc_size <= 0) {
261 262 263 264 265
        p->cn_vc_size = 1024;
        fprintf(stderr, "Buffer size of compute node channels not specified, setting to %d\n", p->cn_vc_size);
    }

    configuration_get_value_int(&config, "PARAMS", "chunk_size", anno, &p->chunk_size);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
266
    if(p->chunk_size <= 0) {
267
        p->chunk_size = 64;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
268
        fprintf(stderr, "Chunk size for packets is not specified, setting to %d\n", p->chunk_size);
269 270 271
    }

    configuration_get_value_double(&config, "PARAMS", "local_bandwidth", anno, &p->local_bandwidth);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
272
    if(p->local_bandwidth <= 0) {
273 274 275 276 277
        p->local_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of local channels not specified, setting to %lf\n", p->local_bandwidth);
    }

    configuration_get_value_double(&config, "PARAMS", "global_bandwidth", anno, &p->global_bandwidth);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
278
    if(p->global_bandwidth <= 0) {
279 280 281 282 283
        p->global_bandwidth = 4.7;
        fprintf(stderr, "Bandwidth of global channels not specified, setting to %lf\n", p->global_bandwidth);
    }

    configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", anno, &p->cn_bandwidth);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
284
    if(p->cn_bandwidth <= 0) {
285 286 287 288 289
        p->cn_bandwidth = 5.25;
        fprintf(stderr, "Bandwidth of compute node channels not specified, setting to %lf\n", p->cn_bandwidth);
    }


290 291
    char routing_str[MAX_NAME_LENGTH];
    configuration_get_value(&config, "PARAMS", "routing", anno, routing_str,
292
            MAX_NAME_LENGTH);
293 294 295 296 297 298 299 300
    if(strcmp(routing_str, "minimal") == 0)
        routing = MINIMAL;
    else if(strcmp(routing_str, "nonminimal")==0 || strcmp(routing_str,"non-minimal")==0)
        routing = NON_MINIMAL;
    else if (strcmp(routing_str, "adaptive") == 0)
        routing = ADAPTIVE;
    else if (strcmp(routing_str, "prog-adaptive") == 0)
	routing = PROG_ADAPTIVE;
301 302 303 304
    else
    {
        fprintf(stderr, 
                "No routing protocol specified, setting to minimal routing\n");
305
        routing = -1;
306 307 308 309 310 311 312 313 314
    }

    // 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;
    p->radix = p->num_vcs *
        (p->num_cn + p->num_global_channels + p->num_routers);
    p->total_routers = p->num_groups * p->num_routers;
315
    p->total_terminals = p->total_routers * p->num_cn;
316
    printf("\n Total nodes %d routers %d groups %d radix %d num_vc %d ", p->num_cn * p->total_routers,
317 318
								p->total_routers,
								p->num_groups,
319 320
								p->radix,
                                                                p->num_vcs);
321 322 323 324 325 326
}

static void dragonfly_configure(){
    anno_map = codes_mapping_get_lp_anno_map(LP_CONFIG_NM);
    assert(anno_map);
    num_params = anno_map->num_annos + (anno_map->has_unanno_lp > 0);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
327
    all_params = calloc(num_params, sizeof(*all_params));
328 329 330 331 332 333 334 335

    for (uint64_t i = 0; i < anno_map->num_annos; i++){
        const char * anno = anno_map->annotations[i];
        dragonfly_read_config(anno, &all_params[i]);
    }
    if (anno_map->has_unanno_lp > 0){
        dragonfly_read_config(NULL, &all_params[anno_map->num_annos]);
    }
336 337 338 339 340 341 342
}

/* report dragonfly statistics like average and maximum packet latency, average number of hops traversed */
static void dragonfly_report_stats()
{
   long long avg_hops, total_finished_packets;
   tw_stime avg_time, max_time;
343
   int total_minimal_packets, total_nonmin_packets, total_completed_packets;
344 345 346 347 348

   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);
   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);
349
   if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
350 351 352
    {
	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);
353
 	MPI_Reduce(&completed_packets, &total_completed_packets, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
354
    }
355 356
   /* print statistics */
   if(!g_tw_mynode)
357
   {	
358
   printf("\n Total hops %lld total finished packets %d ", avg_hops, total_finished_packets);
359
      printf(" Average number of hops traversed %f average message latency %lf us maximum message latency %lf us \n", (float)avg_hops/total_finished_packets, avg_time/(total_finished_packets*1000), max_time/1000);
360
     if(routing == ADAPTIVE || routing == PROG_ADAPTIVE)
361
              printf("\n ADAPTIVE ROUTING STATS: %d percent packets routed minimally %d percent packets routed non-minimally completed packets %d ", total_minimal_packets, total_nonmin_packets, total_completed_packets);
362 363
 
  }
364 365
   return;
}
366

367 368 369
void dragonfly_collective_init(terminal_state * s,
           		   tw_lp * lp)
{
370 371 372 373 374
    // TODO: be annotation-aware
    codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id, NULL,
            &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
    int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
            NULL, 1);
375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
    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;
        if(next_child < (num_lps * num_reps))
        {
            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
}

425
/* dragonfly packet event , generates a dragonfly packet on the compute node */
426
static tw_stime dragonfly_packet_event(char* category, tw_lpid final_dest_lp, uint64_t packet_size, int is_pull, uint64_t pull_size, tw_stime offset, const mn_sched_params *sched_params, int remote_event_size, const void* remote_event, int self_event_size, const void* self_event, tw_lpid src_lp, tw_lp *sender, int is_last_pckt)
427 428 429 430 431 432
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    char* tmp_ptr;

433
    xfer_to_nic_time = codes_local_latency(sender); /* Throws an error of found last KP time > current event time otherwise when LPs of one type are placed together*/
434 435
    e_new = model_net_method_event_new(sender->gid, xfer_to_nic_time+offset,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);
436 437
    strcpy(msg->category, category);
    msg->final_dest_gid = final_dest_lp;
438
    msg->sender_lp=src_lp;
439 440 441 442
    msg->packet_size = packet_size;
    msg->remote_event_size_bytes = 0;
    msg->local_event_size_bytes = 0;
    msg->type = T_GENERATE;
443
    msg->magic = terminal_magic_num;
444 445
    msg->is_pull = is_pull;
    msg->pull_size = pull_size;
446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461

    if(is_last_pckt) /* Its the last packet so pass in remote and local event information*/
      {
	if(remote_event_size > 0)
	 {
		msg->remote_event_size_bytes = remote_event_size;
		memcpy(tmp_ptr, remote_event, remote_event_size);
		tmp_ptr += remote_event_size;
	}
	if(self_event_size > 0)
	{
		msg->local_event_size_bytes = self_event_size;
		memcpy(tmp_ptr, self_event, self_event_size);
		tmp_ptr += self_event_size;
	}
     }
462
	   //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);
463
    tw_event_send(e_new);
464
    return xfer_to_nic_time;
465 466 467 468 469 470 471 472 473
}

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

474 475 476 477
/* given two group IDs, find the router of the src_gid that connects to the dest_gid*/
tw_lpid getRouterFromGroupID(int dest_gid, 
		    int src_gid,
		    int num_routers)
478
{
479 480 481
  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;
482
  
483 484
  if((dest_gid * num_routers) < group_begin)
    offset = (group_begin - dest_gid * num_routers) / num_routers; // take absolute value
485
  
486 487
  int half_channel = num_routers / 4;
  int index = (offset - 1)/(half_channel * num_routers);
488
  
489
  offset=(offset - 1) % (half_channel * num_routers);
490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509

  // If the destination router is in the same group
  tw_lpid router_id;

  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;
}	

/*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*/
void router_credit_send(router_state * s, tw_bf * bf, terminal_message * msg, tw_lp * lp)
{
  tw_event * buf_e;
  tw_stime ts;
  terminal_message * buf_msg;

  int dest=0, credit_delay=0, type = R_BUFFER;
510
  int is_terminal = 0;
511
  int found_magic = router_magic_num;
512

513
  const dragonfly_param *p = s->params;
514
  int sender_radix;
515 516 517 518
 // Notify sender terminal about available buffer space
  if(msg->last_hop == TERMINAL)
  {
   dest = msg->src_terminal_id;
519
   sender_radix = msg->local_id % p->num_cn;  
520
   //determine the time in ns to transfer the credit
521
   credit_delay = (1 / p->cn_bandwidth) * CREDIT_SIZE;
522
   type = T_BUFFER;
523
   is_terminal = 1;
524
   found_magic = terminal_magic_num;
525 526 527 528
  }
   else if(msg->last_hop == GLOBAL)
   {
     dest = msg->intm_lp_id;
529
     sender_radix = p->num_cn + (msg->local_id % p->num_global_channels);
530
     credit_delay = (1 / p->global_bandwidth) * CREDIT_SIZE;
531 532 533 534
   }
    else if(msg->last_hop == LOCAL)
     {
        dest = msg->intm_lp_id;
535
        sender_radix = p->num_cn + p->num_global_channels + (msg->local_id % p->num_routers);
536
     	credit_delay = (1/p->local_bandwidth) * CREDIT_SIZE;
537 538 539 540 541
     }
    else
      printf("\n Invalid message type");

   // Assume it takes 0.1 ns of serialization latency for processing the credits in the queue
542
    assert(sender_radix < s->params->radix );
543
    int output_port = msg->saved_vc / p->num_vcs;
544 545 546

    msg->saved_available_time = s->next_credit_available_time[sender_radix];
    s->next_credit_available_time[sender_radix] = maxd(tw_now(lp), s->next_credit_available_time[output_port]);
547
    ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, (double)credit_delay/1000);
548
	
549
    s->next_credit_available_time[sender_radix]+=ts;
550 551
    if (is_terminal){
        buf_e = model_net_method_event_new(dest, 
552
                s->next_credit_available_time[sender_radix] - tw_now(lp), lp,
553 554 555
                DRAGONFLY, (void**)&buf_msg, NULL);
    }
    else{
556
        buf_e = tw_event_new(dest, s->next_credit_available_time[sender_radix] - tw_now(lp) , lp);
557 558
        buf_msg = tw_event_data(buf_e);
    }
559
    buf_msg->origin_router_id = s->router_id;
560 561
    buf_msg->vc_index = msg->saved_vc;
    buf_msg->type=type;
562
    buf_msg->magic = found_magic;
563 564 565 566 567 568 569 570
    buf_msg->last_hop = msg->last_hop;
    buf_msg->packet_ID=msg->packet_ID;

    tw_event_send(buf_e);

    return;
}

571 572 573 574
static void packet_generate_rc(terminal_state * s, 
			    tw_bf * bf, 
			    terminal_message * msg, 
			    tw_lp * lp)
575
{
576 577
    int i;
    tw_rand_reverse_unif(lp->rng);
578

579
    for(i = 0; i < msg->num_chunks; i++)
580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608
       tw_rand_reverse_unif(lp->rng);
	 
     mn_stats* stat;
     stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
     stat->send_count--;
     stat->send_bytes -= msg->packet_size;
     stat->send_time -= (1/s->params->cn_bandwidth) * msg->packet_size;
}
/* generates packet at the current dragonfly compute node */
static void packet_generate(terminal_state * s, 
			    tw_bf * bf, 
			    terminal_message * msg, 
			    tw_lp * lp)
{
	tw_lpid dest_terminal_id;
	dest_terminal_id = model_net_find_local_device(DRAGONFLY, s->anno, 0,
	msg->final_dest_gid);
	msg->dest_terminal_id = dest_terminal_id;

	const dragonfly_param *p = s->params;

	tw_stime ts;
	tw_event *e;
	terminal_message *m;
	int i, total_event_size;

	uint64_t num_chunks = msg->packet_size / p->chunk_size;
	if (msg->packet_size % s->params->chunk_size)
	  num_chunks++;
609 610 611
	
	if(!num_chunks)
	   num_chunks = 1;
612

613
	msg->num_chunks = num_chunks;
614 615
	msg->packet_ID = lp->gid + g_tw_nlp * s->packet_counter + tw_rand_integer(lp->rng, 0, lp->gid + g_tw_nlp * s->packet_counter);
	msg->travel_start_time = tw_now(lp);
616
	
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
	for(i = 0; i < num_chunks; i++)
	  {
		ts = g_tw_lookahead + 0.1 + tw_rand_exponential(lp->rng, MEAN_INTERVAL/200);
		int chan = -1, j;
		for(j = 0; j < p->num_vcs; j++)
		 {
		     if(s->vc_occupancy[j] < p->cn_vc_size * num_chunks)
		      {
		       chan=j;
		       break;
		      }
		 }
		// this is a terminal event, so use the method-event version
	       void * m_data;
	       e = model_net_method_event_new(lp->gid, i+ts, lp, DRAGONFLY,
		       (void**)&m, &m_data);
	       memcpy(m, msg, sizeof(terminal_message));
	       m->dest_terminal_id = dest_terminal_id;
	       void * m_data_src = model_net_method_get_edata(DRAGONFLY, msg);
	       if (msg->remote_event_size_bytes){
		    memcpy(m_data, m_data_src, msg->remote_event_size_bytes);
	       }
	       if (msg->local_event_size_bytes){ 
		    memcpy((char*)m_data + msg->remote_event_size_bytes,
			    (char*)m_data_src + msg->remote_event_size_bytes,
			    msg->local_event_size_bytes);
	       }
	       m->intm_group_id = -1;
	       m->saved_vc=0;
	       m->chunk_id = i;
	       m->magic = terminal_magic_num;
648
 
649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
	      /* if(msg->packet_ID == TRACK && msg->chunk_id == num_chunks-1)
		 printf("\n packet generated %lld at terminal %d chunk id %d ", msg->packet_ID, (int)lp->gid, i);
	       */
	       m->output_chan = -1;
	       if(chan != -1) // If the input queue is available
		{
		    // Send the packet out
		     m->type = T_SEND;
		     tw_event_send(e);
		}
	      else
		 {
		  printf("\n Exceeded queue size, exitting %d", s->vc_occupancy[0]);
		  MPI_Finalize();
		  exit(-1);
		} //else
665
  } // for
666 667 668 669 670 671 672 673
	total_event_size = model_net_get_msg_sz(DRAGONFLY) + 
	msg->remote_event_size_bytes + msg->local_event_size_bytes;
	mn_stats* stat;
	stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
	stat->send_count++;
	stat->send_bytes += msg->packet_size;
	stat->send_time += (1/p->cn_bandwidth) * msg->packet_size;
	if(stat->max_event_size < total_event_size)
674
	  stat->max_event_size = total_event_size;
675

676 677 678
  return;
}

679 680 681 682 683 684 685 686 687 688 689 690
static void packet_send_rc(terminal_state * s, 
			tw_bf * bf, 
			terminal_message * msg, 
			tw_lp * lp)
{
   s->terminal_available_time = msg->saved_available_time;
   tw_rand_reverse_unif(lp->rng);
   int vc = msg->saved_vc;
   s->vc_occupancy[vc]--;
   s->packet_counter--;
   s->output_vc_state[vc] = VC_IDLE;

691
   if (msg->chunk_id == (msg->num_chunks)-1){
692 693 694 695
     codes_local_latency_reverse(lp);
   }

}
696
/* sends the packet from the current dragonfly compute node to the attached router */
697 698 699 700
static void packet_send(terminal_state * s, 
			tw_bf * bf, 
			terminal_message * msg, 
			tw_lp * lp)
701 702 703 704 705 706 707 708 709 710
{
  tw_stime ts;
  tw_event *e;
  terminal_message *m;
  tw_lpid router_id;
  /* Route the packet to its source router */ 
   int vc=msg->saved_vc;

   //  Each packet is broken into chunks and then sent over the channel
   msg->saved_available_time = s->terminal_available_time;
711
   double head_delay = (1/s->params->cn_bandwidth) * s->params->chunk_size;
712
   ts = head_delay + tw_rand_exponential(lp->rng, (double)head_delay/200);
713
   s->terminal_available_time = maxd(s->terminal_available_time, tw_now(lp));
714 715
   s->terminal_available_time += ts;

716 717 718
   //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);
719
   
720
   codes_mapping_get_lp_id(lp_group_name, "dragonfly_router", NULL, 1,
721 722
           s->router_id/num_routers_per_mgrp, s->router_id % num_routers_per_mgrp, &router_id);

723
   // we are sending an event to the router, so no method_event here
724 725
   e = tw_event_new(router_id, s->terminal_available_time - tw_now(lp), lp);

726 727
   //if(msg->packet_ID == TRACK && msg->chunk_id == num_chunks-1)
   //  printf("\n terminal %d packet %lld chunk %d being sent to router %d router id %d ", (int)lp->gid, (long long)msg->packet_ID, msg->chunk_id, (int)router_id, s->router_id);
728
   m = tw_event_data(e);
729 730 731 732 733
   memcpy(m, msg, sizeof(terminal_message));
   if (msg->remote_event_size_bytes){
        memcpy(m+1, model_net_method_get_edata(DRAGONFLY, msg),
                msg->remote_event_size_bytes);
   }
734
   m->magic = router_magic_num;
735
   m->origin_router_id = s->router_id;
736 737 738 739 740
   m->type = R_ARRIVE;
   m->src_terminal_id = lp->gid;
   m->saved_vc = vc;
   m->last_hop = TERMINAL;
   m->intm_group_id = -1;
741
   m->path_type = -1;
742
   m->local_event_size_bytes = 0;
743
   m->local_id = s->terminal_id;
744 745
   tw_event_send(e);

746
   if(msg->chunk_id == msg->num_chunks - 1) 
747
    {
748 749 750
      // now that message is sent, issue an "idle" event to tell the scheduler
      // when I'm next available
      model_net_method_idle_event(codes_local_latency(lp) +
751
              s->terminal_available_time - tw_now(lp), 0, lp);
752

753 754 755 756 757
      /* local completion message */
      if(msg->local_event_size_bytes > 0)
	 {
           tw_event* e_new;
	   terminal_message* m_new;
758 759 760
	   void* local_event = 
               (char*)model_net_method_get_edata(DRAGONFLY, msg) + 
               msg->remote_event_size_bytes;
761
	   ts = g_tw_lookahead + (1/s->params->cn_bandwidth) * msg->local_event_size_bytes;
762
	   e_new = tw_event_new(msg->sender_lp, ts, lp);
763 764 765 766 767 768 769 770 771
	   m_new = tw_event_data(e_new);
	   memcpy(m_new, local_event, msg->local_event_size_bytes);
	   tw_event_send(e_new);
	}
    }
   
   s->packet_counter++;
   s->vc_occupancy[vc]++;

772
   if(s->vc_occupancy[vc] >= s->params->cn_vc_size)
773
      s->output_vc_state[vc] = VC_CREDIT;
774
   
775 776 777
   return;
}

778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793
static void packet_arrive_rc(terminal_state * s, 
		   tw_bf * bf, 
                   terminal_message * msg, 
                   tw_lp * lp)
{
    
   completed_packets--;

   if(msg->path_type == MINIMAL)
	minimal_count--;

   if(msg->path_type == NON_MINIMAL)
	nonmin_count--;

   tw_rand_reverse_unif(lp->rng);
   s->next_credit_available_time = msg->saved_credit_time;
794
   if(msg->chunk_id == (msg->num_chunks)-1)
795 796 797 798 799 800 801 802 803 804 805 806 807 808
   {
    mn_stats* stat;
    stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
    stat->recv_count--;
    stat->recv_bytes -= msg->packet_size;
    stat->recv_time -= tw_now(lp) - msg->travel_start_time;
    
    N_finished_packets--;
    
    dragonfly_total_time -= (tw_now(lp) - msg->travel_start_time);
    if(bf->c3)
	 dragonfly_max_latency = msg->saved_available_time;
   }
    
809
   if (msg->chunk_id == (msg->num_chunks)-1 && 
810 811 812 813 814 815 816 817
		msg->remote_event_size_bytes && 
		msg->is_pull)
    {
	    int net_id = model_net_get_id(LP_METHOD_NM);
	    model_net_event_rc(net_id, lp, msg->pull_size);
    }

}
818
/* packet arrives at the destination terminal */
819 820 821 822
static void packet_arrive(terminal_state * s, 
		   tw_bf * bf, 
                   terminal_message * msg, 
                   tw_lp * lp)
823
{
824 825 826
    uint64_t num_chunks = msg->packet_size / s->params->chunk_size;
    if (msg->packet_size % s->params->chunk_size)
        num_chunks++;
827 828 829 830 831 832 833 834

    completed_packets++;

    if(msg->path_type == MINIMAL)
	minimal_count++;

    if(msg->path_type == NON_MINIMAL)
	nonmin_count++;
835 836 837
    
    if(msg->path_type != MINIMAL && msg->path_type != NON_MINIMAL)
	printf("\n Wrong message path type %d ", msg->path_type);
838
#if DEBUG == 1
839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860
if( msg->packet_ID == TRACK && msg->chunk_id == num_chunks-1)
    {
	printf( "(%lf) [Terminal %d] packet %lld has arrived  \n",
              tw_now(lp), (int)lp->gid, msg->packet_ID);

	printf("travel start time is %f\n",
                msg->travel_start_time);

    }
#endif

  // Packet arrives and accumulate # queued
  // Find a queue with an empty buffer slot
   tw_event * e, * buf_e;
   terminal_message * m, * buf_msg;
   tw_stime ts;
   bf->c3 = 0;
   bf->c2 = 0;

  if(msg->chunk_id == num_chunks-1)
  {
	 bf->c2 = 1;
861 862 863 864 865
	 mn_stats* stat = model_net_find_stats(msg->category, s->dragonfly_stats_array);
	 stat->recv_count++;
	 stat->recv_bytes += msg->packet_size;
	 stat->recv_time += tw_now(lp) - msg->travel_start_time;

866 867 868 869 870 871 872 873 874 875 876 877
	 N_finished_packets++;
	 dragonfly_total_time += tw_now( lp ) - msg->travel_start_time;

	 if (dragonfly_max_latency < tw_now( lp ) - msg->travel_start_time) 
	 {
		bf->c3 = 1;
		msg->saved_available_time = dragonfly_max_latency;
		dragonfly_max_latency=tw_now( lp ) - msg->travel_start_time;
	 }
	// Trigger an event on receiving server
	if(msg->remote_event_size_bytes)
	{
878
            void * tmp_ptr = model_net_method_get_edata(DRAGONFLY, msg);
879
            ts = g_tw_lookahead + 0.1 + (1/s->params->cn_bandwidth) * msg->remote_event_size_bytes;
880
            if (msg->is_pull){
881
                int net_id = model_net_get_id(LP_METHOD_NM);
882 883 884 885 886 887 888 889 890 891
                model_net_event(net_id, msg->category, msg->sender_lp,
                        msg->pull_size, ts, msg->remote_event_size_bytes,
                        tmp_ptr, 0, NULL, lp);
            }
            else{
                e = tw_event_new(msg->final_dest_gid, ts, lp);
                m = tw_event_data(e);
                memcpy(m, tmp_ptr, msg->remote_event_size_bytes);
                tw_event_send(e); 
            }
892 893 894
	}
  }

895
  int credit_delay = (1 / s->params->cn_bandwidth) * CREDIT_SIZE;
896
  ts = credit_delay + 0.1 + tw_rand_exponential(lp->rng, credit_delay/1000);
897
  
898
  msg->saved_credit_time = s->next_credit_available_time;
899
  s->next_credit_available_time = maxd(s->next_credit_available_time, tw_now(lp));
900 901
  s->next_credit_available_time += ts;

902 903
  if(msg->intm_lp_id == TRACK)
	printf("\n terminal sending credit at chan %d ", msg->saved_vc);
904
  //TODO: be annotation-aware
905
  // no method_event here - message going to router
906
  buf_e = tw_event_new(msg->intm_lp_id, s->next_credit_available_time - tw_now(lp), lp);
907
  buf_msg = tw_event_data(buf_e);
908
  buf_msg->magic = router_magic_num;
909 910 911 912 913 914 915 916 917 918 919 920 921 922
  buf_msg->vc_index = msg->saved_vc;
  buf_msg->type=R_BUFFER;
  buf_msg->packet_ID=msg->packet_ID;
  buf_msg->last_hop = TERMINAL;
  tw_event_send(buf_e);

  return;
}

/* initialize a dragonfly compute node terminal */
void 
terminal_init( terminal_state * s, 
	       tw_lp * lp )
{
923 924 925 926
    uint32_t h1 = 0, h2 = 0; 
    bj_hashlittle2(LP_METHOD_NM, strlen(LP_METHOD_NM), &h1, &h2);
    terminal_magic_num = h1 + h2;
    
927
    int i;
928 929
    char anno[MAX_NAME_LENGTH];

930
    // Assign the global router ID
931
    // TODO: be annotation-aware
932 933 934 935 936 937 938
    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{
939
        s->anno = strdup(anno);
940 941 942 943
        int id = configuration_get_annotation_index(anno, anno_map);
        s->params = &all_params[id];
    }

944
   int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
945
           s->anno, 0);
946 947

   s->terminal_id = (mapping_rep_id * num_lps) + mapping_offset;  
Misbah Mubarak's avatar
Misbah Mubarak committed
948
   s->router_id=(int)s->terminal_id / (s->params->num_routers/2);
949 950 951
   s->terminal_available_time = 0.0;
   s->packet_counter = 0;

952 953
   s->vc_occupancy = (int*)malloc(s->params->num_vcs * sizeof(int));
   s->output_vc_state = (int*)malloc(s->params->num_vcs * sizeof(int));
954

955
   for( i = 0; i < s->params->num_vcs; i++ )
956 957 958 959
    {
      s->vc_occupancy[i]=0;
      s->output_vc_state[i]=VC_IDLE;
    }
960
//   printf("\n Terminal ID %d Router ID %d ", s->terminal_id, s->router_id);
961
   dragonfly_collective_init(s, lp);
962 963 964
   return;
}

965
/* collective operation for the torus network */
966
void dragonfly_collective(char const * category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
967 968 969 970 971 972 973
{
    tw_event * e_new;
    tw_stime xfer_to_nic_time;
    terminal_message * msg;
    tw_lpid local_nic_id;
    char* tmp_ptr;

974 975 976 977
    codes_mapping_get_lp_info(sender->gid, lp_group_name, &mapping_grp_id,
            NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
    codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, NULL, 1,
            mapping_rep_id, mapping_offset, &local_nic_id);
978

979
    xfer_to_nic_time = codes_local_latency(sender);
980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
    e_new = model_net_method_event_new(local_nic_id, xfer_to_nic_time,
            sender, DRAGONFLY, (void**)&msg, (void**)&tmp_ptr);

    msg->remote_event_size_bytes = message_size;
    strcpy(msg->category, category);
    msg->sender_svr=sender->gid;
    msg->type = D_COLLECTIVE_INIT;

    tmp_ptr = (char*)msg;
    tmp_ptr += dragonfly_get_msg_sz();
    if(remote_event_size > 0)
     {
            msg->remote_event_size_bytes = remote_event_size;
            memcpy(tmp_ptr, remote_event, remote_event_size);
            tmp_ptr += remote_event_size;
     }

    tw_event_send(e_new);
    return;
}

/* reverse for collective operation of the dragonfly network */
void dragonfly_collective_rc(int message_size, tw_lp* sender)
{
     codes_local_latency_reverse(sender);
     return;
}

static void send_remote_event(terminal_state * s,
                        tw_bf * bf,
                        terminal_message * msg,
                        tw_lp * lp)
{
    // Trigger an event on receiving server
    if(msg->remote_event_size_bytes)
     {
            tw_event* e;
            tw_stime ts;
            terminal_message * m;
1019
            ts = (1/s->params->cn_bandwidth) * msg->remote_event_size_bytes;
1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047
            e = codes_event_new(s->origin_svr, ts, lp);
            m = tw_event_data(e);
            char* tmp_ptr = (char*)msg;
            tmp_ptr += dragonfly_get_msg_sz();
            memcpy(m, tmp_ptr, msg->remote_event_size_bytes);
            tw_event_send(e);
     }
}

static void node_collective_init(terminal_state * s,
                        tw_bf * bf,
                        terminal_message * msg,
                        tw_lp * lp)
{
        tw_event * e_new;
        tw_lpid parent_nic_id;
        tw_stime xfer_to_nic_time;
        terminal_message * msg_new;
        int num_lps;

        msg->saved_collective_init_time = s->collective_init_time;
        s->collective_init_time = tw_now(lp);
	s->origin_svr = msg->sender_svr;
	
        if(s->is_leaf)
        {
            //printf("\n LP %ld sending message to parent %ld ", s->node_id, s->parent_node_id);
            /* get the global LP ID of the parent node */
1048 1049 1050 1051 1052
            // 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);
            num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
1053 1054
                    s->anno, 0);
            codes_mapping_get_lp_id(lp_group_name, LP_CONFIG_NM, s->anno, 0,
1055 1056
                    s->parent_node_id/num_lps, (s->parent_node_id % num_lps),
                    &parent_nic_id);
1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086

           /* send a message to the parent that the LP has entered the collective operation */
            xfer_to_nic_time = g_tw_lookahead + LEVEL_DELAY;
            //e_new = codes_event_new(parent_nic_id, xfer_to_nic_time, lp);
	    void* m_data;
	    e_new = model_net_method_event_new(parent_nic_id, xfer_to_nic_time,
            	lp, DRAGONFLY, (void**)&msg_new, (void**)&m_data);
	    	
            memcpy(msg_new, msg, sizeof(terminal_message));
	    if (msg->remote_event_size_bytes){
        	memcpy(m_data, model_net_method_get_edata(DRAGONFLY, msg),
                	msg->remote_event_size_bytes);
      	    }
	    
            msg_new->type = D_COLLECTIVE_FAN_IN;
            msg_new->sender_node = s->node_id;

            tw_event_send(e_new);
        }
        return;
}

static void node_collective_fan_in(terminal_state * s,
                        tw_bf * bf,
                        terminal_message * msg,
                        tw_lp * lp)
{
        int i;
        s->num_fan_nodes++;

1087 1088 1089
        codes_mapping_get_lp_info(lp->gid, lp_group_name, &mapping_grp_id,
                NULL, &mapping_type_id, NULL, &mapping_rep_id, &mapping_offset);
        int num_lps = codes_mapping_get_lp_count(lp_group_name, 1, LP_CONFIG_NM,
1090
                s->anno, 0);
1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102