model-net.c 18.2 KB
Newer Older
1
/*
Philip Carns's avatar
Philip Carns committed
2
 * Copyright (C) 2013 University of Chicago.
3
 * See COPYRIGHT notice in top-level directory.
Philip Carns's avatar
Philip Carns committed
4
 *
5
 */
Philip Carns's avatar
Philip Carns committed
6

7 8 9
#include <string.h>
#include <assert.h>

10
#include "codes/model-net.h"
11
#include "codes/model-net-method.h"
12
#include "codes/model-net-lp.h"
13
#include "codes/model-net-sched.h"
14
#include "codes/codes.h"
15

16 17 18 19
#define STR_SIZE 16
#define PROC_TIME 10.0

extern struct model_net_method simplenet_method;
20
extern struct model_net_method simplewan_method;
21
extern struct model_net_method torus_method;
22
extern struct model_net_method dragonfly_method;
Philip Carns's avatar
Philip Carns committed
23
extern struct model_net_method loggp_method;
24

25 26 27 28 29 30 31 32 33 34 35 36
#define X(a,b,c,d) b,
char * model_net_lp_config_names[] = {
    NETWORK_DEF
};
#undef X

#define X(a,b,c,d) c,
char * model_net_method_names[] = {
    NETWORK_DEF
};
#undef X

37
/* Global array initialization, terminated with a NULL entry */
38 39 40 41 42
#define X(a,b,c,d) d,
struct model_net_method* method_array[] = { 
    NETWORK_DEF
};
#undef X
43

44 45 46
int in_sequence = 0;
tw_stime mn_msg_offset = 0.0;

47
int model_net_setup(char* name,
48
		    uint64_t packet_size,
49 50
		    const void* net_params)
{
51
     int i;
52 53 54
    /* find struct for underlying method (according to configuration file) */
     for(i=0; method_array[i] != NULL; i++)
     {
55
     	if(strcmp(model_net_method_names[i], name) == 0)
56 57 58 59 60 61
	{
	   method_array[i]->mn_setup(net_params);
	   method_array[i]->packet_size = packet_size;
	   return(i);
	}
     }
62
     fprintf(stderr, "Error: undefined network name %s (Available options simplenet, torus, dragonfly) \n", name);
63 64 65
     return -1; // indicating error
}

66 67 68
int model_net_get_id(char *name){
    int i;
    for(i=0; method_array[i] != NULL; i++) {
69
        if(strcmp(model_net_method_names[i], name) == 0) {
70 71 72 73 74 75
            return i;
        }
    }
    return -1;
}

76 77 78
void model_net_write_stats(tw_lpid lpid, struct mn_stats* stat)
{
    int ret;
79
    char id[19+CATEGORY_NAME_MAX+1];
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
    char data[1024];

    sprintf(id, "model-net-category-%s", stat->category);
    sprintf(data, "lp:%ld\tsend_count:%ld\tsend_bytes:%ld\tsend_time:%f\t" 
        "recv_count:%ld\trecv_bytes:%ld\trecv_time:%f\tmax_event_size:%ld\n",
        (long)lpid,
        stat->send_count,
        stat->send_bytes,
        stat->send_time,
        stat->recv_count,
        stat->recv_bytes,
        stat->recv_time,
        stat->max_event_size);

    ret = lp_io_write(lpid, id, strlen(data), data);
    assert(ret == 0);

    return;
}

void model_net_print_stats(tw_lpid lpid, mn_stats mn_stats_array[])
{

    int i;
    struct mn_stats all;

    memset(&all, 0, sizeof(all));
    sprintf(all.category, "all");

    for(i=0; i<CATEGORY_MAX; i++)
    {
        if(strlen(mn_stats_array[i].category) > 0)
        {
            all.send_count += mn_stats_array[i].send_count;
            all.send_bytes += mn_stats_array[i].send_bytes;
            all.send_time += mn_stats_array[i].send_time;
            all.recv_count += mn_stats_array[i].recv_count;
            all.recv_bytes += mn_stats_array[i].recv_bytes;
            all.recv_time += mn_stats_array[i].recv_time;
            if(mn_stats_array[i].max_event_size > all.max_event_size)
                all.max_event_size = mn_stats_array[i].max_event_size;

            model_net_write_stats(lpid, &mn_stats_array[i]);
        }
    }
    model_net_write_stats(lpid, &all);
}

struct mn_stats* model_net_find_stats(const char* category, mn_stats mn_stats_array[])
{
    int i;
    int new_flag = 0;
    int found_flag = 0;

    for(i=0; i<CATEGORY_MAX; i++)
    {
        if(strlen(mn_stats_array[i].category) == 0)
        {
            found_flag = 1;
            new_flag = 1;
            break;
        }
        if(strcmp(category, mn_stats_array[i].category) == 0)
        {
            found_flag = 1;
            new_flag = 0;
            break;
        }
    }
    assert(found_flag);

    if(new_flag)
    {
        strcpy(mn_stats_array[i].category, category);
    }
    return(&mn_stats_array[i]);
}

158
static void model_net_event_impl_base(
159 160 161 162 163 164 165 166 167 168 169
        int net_id,
        char* category, 
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        int is_pull,
        tw_stime offset,
        int remote_event_size,
        const void* remote_event,
        int self_event_size,
        const void* self_event,
        tw_lp *sender) {
170

171 172 173 174 175 176 177 178
    if (remote_event_size + self_event_size + sizeof(model_net_wrap_msg) 
            > g_tw_msg_sz){
        tw_error(TW_LOC, "Error: model_net trying to transmit an event of size "
                         "%d but ROSS is configured for events of size %zd\n",
                         remote_event_size+self_event_size+sizeof(model_net_wrap_msg),
                         g_tw_msg_sz);
        return;
    }
179

180
    tw_lpid mn_lp = model_net_find_local_device(net_id, sender);
181
    tw_stime poffset = codes_local_latency(sender);
182 183 184 185 186 187 188 189 190 191 192 193 194 195
    if (in_sequence){
        tw_stime tmp = mn_msg_offset;
        mn_msg_offset += poffset;
        poffset += tmp;
    }
    tw_event *e = codes_event_new(mn_lp, poffset+offset, sender);

    model_net_wrap_msg *m = tw_event_data(e);
    m->event_type = MN_BASE_NEW_MSG;
    m->magic = model_net_base_magic;

    // set the request struct 
    model_net_request *r = &m->msg.m_base.u.req;
    r->net_id = net_id;
196
    r->packet_size = model_net_get_packet_size(net_id);
197
    r->final_dest_lp = final_dest_lp;
198
    r->src_lp = sender->gid;
199 200 201
    r->msg_size = message_size;
    r->remote_event_size = remote_event_size;
    r->self_event_size = self_event_size;
202
    r->is_pull = is_pull;
203 204 205 206 207 208 209 210 211 212 213
    strncpy(r->category, category, CATEGORY_NAME_MAX-1);
    r->category[CATEGORY_NAME_MAX-1]='\0';
    
    void *e_msg = (m+1);
    if (remote_event_size > 0){
        memcpy(e_msg, remote_event, remote_event_size);
        e_msg = (char*)e_msg + remote_event_size; 
    }
    if (self_event_size > 0){
        memcpy(e_msg, self_event, self_event_size);
    }
214

215 216 217 218 219
    //print_base(m);
    tw_event_send(e);
}
static void model_net_event_impl_base_rc(tw_lp *sender){
    codes_local_latency_reverse(sender);
220
}
221 222 223 224 225 226 227 228 229 230 231 232 233

void model_net_event(
    int net_id,
    char* category, 
    tw_lpid final_dest_lp, 
    uint64_t message_size, 
    tw_stime offset,
    int remote_event_size,
    const void* remote_event,
    int self_event_size,
    const void* self_event,
    tw_lp *sender)
{
234 235 236
    model_net_event_impl_base(net_id, category, final_dest_lp, message_size,
            0, offset, remote_event_size, remote_event, self_event_size,
            self_event, sender);
237 238 239 240 241 242 243 244 245 246 247 248 249
}

void model_net_pull_event(
        int net_id,
        char *category,
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
        const void *self_event,
        tw_lp *sender){
    /* NOTE: for a pull, we are filling the *remote* event - it will be remote
     * from the destination's POV */
250 251
    model_net_event_impl_base(net_id, category, final_dest_lp, message_size,
            1, offset, self_event_size, self_event, 0, NULL, sender);
252 253 254
}


255 256 257
int model_net_set_params()
{
  char mn_name[MAX_NAME_LENGTH];
258
  char sched[MAX_NAME_LENGTH];
259 260
  long int packet_size_l = 0;
  uint64_t packet_size;
261
  int net_id=-1;
262
  int ret;
263 264 265 266

  config_lpgroups_t paramconf;
  configuration_get_lpgroups(&config, "PARAMS", &paramconf);
  configuration_get_value(&config, "PARAMS", "modelnet", mn_name, MAX_NAME_LENGTH);
267 268
  ret = configuration_get_value(&config, "PARAMS", "modelnet_scheduler", sched,
          MAX_NAME_LENGTH);
269 270

  configuration_get_value_longint(&config, "PARAMS", "packet_size", &packet_size_l);
271
  packet_size = packet_size_l;
272

273
    if (ret > 0){
274 275 276 277 278 279
        int i;
        for (i = 0; i < MAX_SCHEDS; i++){
            if (strcmp(sched_names[i], sched) == 0){
                mn_sched_type = i;
                break;
            }
280
        }
281 282 283
        if (i == MAX_SCHEDS){
            fprintf(stderr, 
                    "Unknown value for PARAMS:modelnet-scheduler : %s\n", 
284 285 286 287 288 289 290 291 292
                    sched);
            abort();
        }
    }
    else{
        // default: FCFS
        mn_sched_type = MN_SCHED_FCFS;
    }

293 294 295 296 297 298 299 300 301 302 303
    if (mn_sched_type == MN_SCHED_FCFS_FULL){
        // override packet size to something huge (leave a bit in the unlikely
        // case that an op using packet size causes overflow)
        packet_size = 1ull << 62;
    }
    else if (!packet_size && mn_sched_type != MN_SCHED_FCFS_FULL)
    {
        packet_size = 512;
        fprintf(stderr, "\n Warning, no packet size specified, setting packet size to %llu ", packet_size);
    }

304
  if(strcmp(model_net_method_names[SIMPLENET],mn_name)==0)
305 306 307 308 309 310
   {
     double net_startup_ns, net_bw_mbps;
     simplenet_param net_params;
     
     configuration_get_value_double(&config, "PARAMS", "net_startup_ns", &net_startup_ns);
     configuration_get_value_double(&config, "PARAMS", "net_bw_mbps", &net_bw_mbps);
311 312
     net_params.net_startup_ns = net_startup_ns;
     net_params.net_bw_mbps =  net_bw_mbps;
313
     net_id = model_net_setup(model_net_method_names[SIMPLENET], packet_size, (const void*)&net_params); /* Sets the network as simplenet and packet size 512 */
314
   }
315
  else if (strcmp(model_net_method_names[SIMPLEWAN],mn_name)==0){
316 317 318
    simplewan_param net_params;
    configuration_get_value_relpath(&config, "PARAMS", "net_startup_ns_file", net_params.startup_filename, MAX_NAME_LENGTH);
    configuration_get_value_relpath(&config, "PARAMS", "net_bw_mbps_file", net_params.bw_filename, MAX_NAME_LENGTH);
319
    net_id = model_net_setup(model_net_method_names[SIMPLEWAN], packet_size, (const void*)&net_params);
320
  }
321
   else if(strcmp(model_net_method_names[LOGGP],mn_name)==0)
Philip Carns's avatar
Philip Carns committed
322 323 324 325
   {
     char net_config_file[256];
     loggp_param net_params;
     
326
     configuration_get_value_relpath(&config, "PARAMS", "net_config_file", net_config_file, 256);
Philip Carns's avatar
Philip Carns committed
327
     net_params.net_config_file = net_config_file;
328
     net_id = model_net_setup(model_net_method_names[LOGGP], packet_size, (const void*)&net_params); /* Sets the network as loggp and packet size 512 */
Philip Carns's avatar
Philip Carns committed
329 330
   }

331
  else if(strcmp(model_net_method_names[DRAGONFLY], mn_name)==0)	  
332
    {
333
       dragonfly_param net_params;
334
       int num_routers=0, num_vcs=0, local_vc_size=0, global_vc_size=0, cn_vc_size=0, chunk_size=0;
335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376
       double local_bandwidth=0.0, cn_bandwidth=0.0, global_bandwidth=0.0;
       
       configuration_get_value_int(&config, "PARAMS", "num_routers", &num_routers);
       if(!num_routers)
	{
	   num_routers = 4; 
	   printf("\n Number of dimensions not specified, setting to %d ", num_routers);
        } 
       net_params.num_routers = num_routers; 

       configuration_get_value_int(&config, "PARAMS", "num_vcs", &num_vcs);
       if(!num_vcs)
       {
          num_vcs = 1;
	  printf("\n Number of virtual channels not specified, setting to %d ", num_vcs);
       }
       net_params.num_vcs = num_vcs;

       configuration_get_value_int(&config, "PARAMS", "local_vc_size", &local_vc_size);
       if(!local_vc_size)
	{
	   local_vc_size = 1024;
	   printf("\n Buffer size of local channels not specified, setting to %d ", local_vc_size);
	}
       net_params.local_vc_size = local_vc_size;

       configuration_get_value_int(&config, "PARAMS", "global_vc_size", &global_vc_size);
       if(!global_vc_size)
	{
	  global_vc_size = 2048;
	  printf("\n Buffer size of global channels not specified, setting to %d ", global_vc_size);
	}
       net_params.global_vc_size = global_vc_size;

       configuration_get_value_int(&config, "PARAMS", "cn_vc_size", &cn_vc_size);
       if(!cn_vc_size)
	 {
	    cn_vc_size = 1024;
	    printf("\n Buffer size of compute node channels not specified, setting to %d ", cn_vc_size);
	 }
       net_params.cn_vc_size = cn_vc_size;

377 378 379 380 381 382 383 384
	configuration_get_value_int(&config, "PARAMS", "chunk_size", &chunk_size);
	if(!chunk_size)
	  {
		chunk_size = 64;
		printf("\n Chunk size for packets is specified, setting to %d ", chunk_size);
	  }
	net_params.chunk_size = chunk_size;

385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408
	configuration_get_value_double(&config, "PARAMS", "local_bandwidth", &local_bandwidth);
        if(!local_bandwidth)
	  {
	    local_bandwidth = 5.25;
	    printf("\n Bandwidth of local channels not specified, setting to %lf ", local_bandwidth);
	 }
       net_params.local_bandwidth = local_bandwidth;

       configuration_get_value_double(&config, "PARAMS", "global_bandwidth", &global_bandwidth);
        if(!global_bandwidth)
	{
	     global_bandwidth = 4.7;
	     printf("\n Bandwidth of global channels not specified, setting to %lf ", global_bandwidth);
	}
	net_params.global_bandwidth = global_bandwidth;

	configuration_get_value_double(&config, "PARAMS", "cn_bandwidth", &cn_bandwidth);
	if(!cn_bandwidth)
	 {
	     cn_bandwidth = 5.25;
	     printf("\n Bandwidth of compute node channels not specified, setting to %lf ", cn_bandwidth);
	}
	net_params.cn_bandwidth = cn_bandwidth;

409
	
410 411 412 413 414 415
       char routing[MAX_NAME_LENGTH];
       configuration_get_value(&config, "PARAMS", "routing", routing, MAX_NAME_LENGTH);
       if(strcmp(routing, "minimal") == 0)
	   net_params.routing = 0;
       else if(strcmp(routing, "nonminimal")==0 || strcmp(routing,"non-minimal")==0)
	       net_params.routing = 1;
416 417
	else if (strcmp(routing, "adaptive") == 0)
		net_params.routing = 2;
418 419 420 421 422
       else
       {
       	   printf("\n No routing protocol specified, setting to minimal routing");
   	   net_params.routing = 0;	   
       }
423
    net_id = model_net_setup(model_net_method_names[DRAGONFLY], packet_size, (const void*)&net_params);   
424
    }
425
   else if(strcmp(model_net_method_names[TORUS], mn_name)==0)
426
     {
427 428
	torus_param net_params;
	char dim_length[MAX_NAME_LENGTH];
429
	int n_dims=0, buffer_size=0, num_vc=0, i=0, chunk_size = 0;
430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452
	double link_bandwidth=0;

	configuration_get_value_int(&config, "PARAMS", "n_dims", &n_dims);
	if(!n_dims)
	{
	   n_dims = 4; /* a 4-D torus */
	   printf("\n Number of dimensions not specified, setting to %d ", n_dims);
	}
	
	configuration_get_value_double(&config, "PARAMS", "link_bandwidth", &link_bandwidth);	
	if(!link_bandwidth)
	{
		link_bandwidth = 2.0; /*default bg/q configuration */
		printf("\n Link bandwidth not specified, setting to %lf ", link_bandwidth);
	}

	configuration_get_value_int(&config, "PARAMS", "buffer_size", &buffer_size);
	if(!buffer_size)
	{
		buffer_size = 2048;
		printf("\n Buffer size not specified, setting to %d ",buffer_size);
	}

453 454 455 456 457 458
	configuration_get_value_int(&config, "PARAMS", "chunk_size", &chunk_size);
	if(!chunk_size)
	{
	       chunk_size = 32;
	       printf("\n Chunk size not specified, setting to %d ", chunk_size);
	}
459 460 461 462 463 464 465 466 467 468 469 470
	configuration_get_value_int(&config, "PARAMS", "num_vc", &num_vc);
	if(!num_vc)
	{
		num_vc = 1; /*by default, we have one for taking packets, another for taking credit*/
		printf("\n num_vc not specified, setting to %d ", num_vc);
	}

        configuration_get_value(&config, "PARAMS", "dim_length", dim_length, MAX_NAME_LENGTH);
        char* token;
	net_params.n_dims=n_dims;
	net_params.num_vc=num_vc;
	net_params.buffer_size=buffer_size;
471
	net_params.chunk_size = chunk_size;
472 473 474 475 476 477 478 479 480 481 482 483 484 485 486
	net_params.link_bandwidth=link_bandwidth;
	net_params.dim_length=malloc(n_dims*sizeof(int));
        token = strtok(dim_length, ",");	
	while(token != NULL)
	{
	   sscanf(token, "%d", &net_params.dim_length[i]);
	   if(!net_params.dim_length[i])
	   {
	      printf("\n Invalid torus dimension specified %d, exitting... ", net_params.dim_length[i]);
	      MPI_Finalize();
	      exit(-1);
	   }
	   i++;
	   token = strtok(NULL,",");
	}
487
	net_id = model_net_setup(model_net_method_names[TORUS], packet_size, (const void*)&net_params);
488 489 490
     }
  else
       printf("\n Invalid network argument %s ", mn_name);
491
  model_net_base_init();
492 493
  return net_id;
}
494

495 496 497 498
void model_net_event_rc(
        int net_id,
        tw_lp *sender,
        uint64_t message_size){
499
    model_net_event_impl_base_rc(sender);
500 501 502 503 504
}

void model_net_pull_event_rc(
        int net_id,
        tw_lp *sender) {
505
    model_net_event_impl_base_rc(sender);
506 507
}

508
/* returns the message size, can be either simplenet, dragonfly or torus message size*/
509
int model_net_get_msg_sz(int net_id)
510 511 512
{
   // TODO: Add checks on network name
   // TODO: Add dragonfly and torus network models
513 514
   return sizeof(model_net_wrap_msg);
#if 0
515
   if(net_id < 0 || net_id >= MAX_NETS)
516 517 518 519 520 521
     {
      printf("%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
      exit(-1);
     }

       return method_array[net_id]->mn_get_msg_sz();
522
#endif
523 524 525
}

/* returns the packet size in the modelnet struct */
526
uint64_t model_net_get_packet_size(int net_id)
527
{
528
  if(net_id < 0 || net_id >= MAX_NETS)
529 530 531 532 533 534 535
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }
  return method_array[net_id]->packet_size; // TODO: where to set the packet size?
}

536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557
/* This event does a collective operation call for model-net */
void model_net_event_collective(int net_id, char* category, int message_size, int remote_event_size, const void* remote_event, tw_lp* sender)
{
  if(net_id < 0 || net_id > MAX_NETS)
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }
  return method_array[net_id]->mn_collective_call(category, message_size, remote_event_size, remote_event, sender);
}

/* reverse event of the collective operation call */
void model_net_event_collective_rc(int net_id, int message_size, tw_lp* sender)
{
  if(net_id < 0 || net_id > MAX_NETS)
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }
  return method_array[net_id]->mn_collective_call_rc(message_size, sender);
}

558 559 560
/* returns lp type for modelnet */
const tw_lptype* model_net_get_lp_type(int net_id)
{
561
    if(net_id < 0 || net_id >= MAX_NETS)
562 563 564 565 566 567 568 569 570 571
     {
       fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
       exit(-1);
     }

   // TODO: ADd checks by network names
   // Add dragonfly and torus network models
   return method_array[net_id]->mn_get_lp_type();
}

572 573
void model_net_report_stats(int net_id)
{
574
  if(net_id < 0 || net_id >= MAX_NETS)
575 576 577 578 579 580 581
  {
    fprintf(stderr, "%s Error: Uninitializied modelnet network, call modelnet_init first\n", __FUNCTION__);
    exit(-1);
   }

     // TODO: ADd checks by network names
     //    // Add dragonfly and torus network models
582 583
   method_array[net_id]->mn_report_stats();
   return;
584
}
585 586 587 588 589 590

tw_lpid model_net_find_local_device(int net_id, tw_lp *sender)
{
    return(method_array[net_id]->model_net_method_find_local_device(sender));
}

591 592 593 594 595 596 597 598
/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */