model-net-mpi-replay.c 91.5 KB
Newer Older
1
/*
2
* Copyright (C) 2014 University of Chicago.
3 4 5 6 7
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
8
#include <sys/stat.h>
mubarak's avatar
mubarak committed
9
#include <sys/resource.h>
10 11 12 13 14 15 16
#include "codes/codes-workload.h"
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
#include "codes/rc-stack.h"
#include "codes/quicklist.h"
17
#include "codes/quickhash.h"
18
#include "codes/codes-jobmap.h"
19

20
/* turning on track lp will generate a lot of output messages */
21
#define MN_LP_NM "modelnet_dragonfly_custom"
22
#define CONTROL_MSG_SZ 64
23
#define TRACE -1
24 25
#define MAX_SYN_SENDS 10
#define MAX_WAIT_REQS 1024
26
#define CS_LP_DBG 1
27
#define RANK_HASH_TABLE_SZ 2000
28
#define NW_LP_NM "nw-lp"
29 30 31
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
32
#define INCAST_ID 511
33

34 35 36
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

37
static unsigned long perm_switch_thresh = 8388608;
38

39
/* NOTE: Message tracking works in sequential mode only! */
40
static int debug_cols = 0;
41
static int synthetic_pattern = 1;
42 43 44
/* Turning on this option slows down optimistic mode substantially. Only turn
 * on if you get issues with wait-all completion with traces. */
static int preserve_wait_ordering = 0;
45
static int enable_msg_tracking = 0;
46
static int is_synthetic = 0;
47
tw_lpid TRACK_LP = -1;
48
int nprocs = 0;
49
static double total_syn_data = 0;
50
static int unmatched = 0;
51
char workload_type[128];
52
char workload_name[128];
53 54 55 56
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
57
static int num_dumpi_traces = 0;
58
static int64_t EAGER_THRESHOLD = INT_MAX;
59

mubarak's avatar
mubarak committed
60
static long num_ops = 0;
61
static int upper_threshold = 1048576;
62
static int alloc_spec = 0;
63
static tw_stime self_overhead = 10.0;
64
static tw_stime mean_interval = 10;
65
static int payload_sz = 1024;
66 67

/* Doing LP IO*/
68
static char * params = NULL;
69
static char lp_io_dir[256] = {'\0'};
70
static char sampling_dir[32] = {'\0'};
71
static char mpi_msg_dir[32] = {'\0'};
72 73 74 75
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

76 77 78 79
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
80 81 82
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
83 84 85 86 87
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

88 89
/* Variables for Cortex Support */
/* Matthieu's additions start */
90
#ifdef ENABLE_CORTEX_PYTHON
91 92 93
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
94
#endif
95 96
/* Matthieu's additions end */

97 98
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
Misbah Mubarak's avatar
Misbah Mubarak committed
99
typedef unsigned int dumpi_req_id;
100 101

static int net_id = 0;
102
static float noise = 1.0;
103 104 105
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
106
static int syn_type = 0;
107

108
FILE * workload_log = NULL;
109
FILE * msg_size_log = NULL;
110 111 112 113
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
114

115 116 117
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

118 119 120
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

121 122 123 124 125 126
double max_time = 0,  max_comm_time = 0, max_wait_time = 0, max_send_time = 0, max_recv_time = 0;
double avg_time = 0, avg_comm_time = 0, avg_wait_time = 0, avg_send_time = 0, avg_recv_time = 0;


/* runtime option for disabling computation time simulation */
static int disable_delay = 0;
127 128 129
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
130
static int enable_debug = 0;
131

132 133 134
/* set group context */
struct codes_mctx group_ratio;

135
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
136
* MPI_SEND_ARRIVED is issued when a MPI message arrives at its destination (the message is transported by model-net and an event is invoked when it arrives.
137 138 139 140 141 142 143
* MPI_SEND_POSTED is issued when a MPI message has left the source LP (message is transported via model-net). */
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
	MPI_SEND_POSTED,
144 145 146 147 148 149
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
150 151
};

152 153 154 155 156 157
/* type of synthetic traffic */
enum TRAFFIC
{
    UNIFORM = 1, /* sends message to a randomly selected node */
    NEAREST_NEIGHBOR = 2, /* sends message to the next node (potentially connected to the same router) */
    ALLTOALL = 3, /* sends message to all other nodes */
158 159
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
160
};
161 162 163 164
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
165
    int app_id;
166 167 168 169 170
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
171 172 173 174 175 176 177
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
struct mpi_msgs_queue
{
    int op_type;
    int tag;
    int source_rank;
    int dest_rank;
178
    int64_t num_bytes;
179
    int64_t seq_id;
180 181 182 183 184 185 186 187
    tw_stime req_init_time;
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* stores request IDs of completed MPI operations (Isends or Irecvs) */
struct completed_requests
{
Misbah Mubarak's avatar
Misbah Mubarak committed
188
	unsigned int req_id;
189
    struct qlist_head ql;
190
    int index;
191 192 193 194 195 196
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
Misbah Mubarak's avatar
Misbah Mubarak committed
197
    unsigned int req_ids[MAX_WAIT_REQS];
198
	int num_completed;
199 200
	int count;
    tw_stime start_time;
201 202 203
    struct qlist_head ql;
};

204 205 206 207 208 209
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
210
    struct qhash_head  hash_link;
211 212
    struct qlist_head ql; 
};
213 214 215 216 217 218 219 220 221 222
typedef struct mpi_msgs_queue mpi_msgs_queue;
typedef struct completed_requests completed_requests;
typedef struct pending_waits pending_waits;

/* state of the network LP. It contains the pointers to send/receive lists */
struct nw_state
{
	long num_events_per_lp;
	tw_lpid nw_id;
	short wrkld_end;
223 224
    int app_id;
    int local_rank;
225

226
    int synthetic_pattern;
227 228 229
    int is_finished;
    int neighbor_completed;

230
    struct rc_stack * processed_ops;
231
    struct rc_stack * processed_wait_op;
232
    struct rc_stack * matched_reqs;
233
//    struct rc_stack * indices;
234 235 236 237 238 239 240 241 242 243

    /* count of sends, receives, collectives and delays */
	unsigned long num_sends;
	unsigned long num_recvs;
	unsigned long num_cols;
	unsigned long num_delays;
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;

244

245 246
	/* time spent by the LP in executing the app trace*/
	double start_time;
247 248 249 250 251 252 253 254 255

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

256 257 258 259 260
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
261 262
    /* max time for synthetic traffic message */
    double max_time;
263 264 265 266 267 268 269 270 271 272
	/* time spent in message receive */
	double recv_time;
	/* time spent in wait operation */
	double wait_time;
	/* FIFO for isend messages arrived on destination */
	struct qlist_head arrival_queue;
	/* FIFO for irecv messages posted but not yet matched with send operations */
	struct qlist_head pending_recvs_queue;
	/* List of completed send/receive requests */
	struct qlist_head completed_reqs;
273

274
    tw_stime cur_interval_end;
275
    
276 277
    /* Pending wait operation */
    struct pending_waits * wait_op;
278

279 280 281 282 283 284
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

285 286 287
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

288 289
    unsigned long syn_data;
    unsigned long gen_data;
290 291
  
    unsigned long prev_switch;
292
    int saved_perm_dest;
293

294 295 296 297
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
298
    char output_buf[512];
299
    char col_stats[64];
300 301 302 303
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
304
* ptr_match_op holds the matched MPI operation which are removed from the queues when a send is matched with the receive in forward event handler.
305 306 307
* network event being sent. op is the MPI operation issued by the network workloads API. rv_data holds the data for reverse computation (TODO: Fill this data structure only when the simulation runs in optimistic mode). */
struct nw_message
{
308
   // forward message handler
309
   int msg_type;
310
   int op_type;
311
   model_net_event_return event_rc;
312
   struct codes_workload_op * mpi_op;
313

314 315 316
   struct
   {
       tw_lpid src_rank;
317
       int dest_rank;
318
       int64_t num_bytes;
319 320 321 322 323
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
Misbah Mubarak's avatar
Misbah Mubarak committed
324
       unsigned int req_id;
325
       int matched_req;
326
       int tag;
327
       int app_id;
328 329
       int found_match;
       short wait_completed;
330
       short rend_send;
331 332 333 334 335 336 337
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
338
       int64_t saved_num_bytes;
339
       int saved_syn_length;
340
       unsigned long saved_prev_switch;
341
   } rc;
342 343
};

344
static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op, int matched_req);
345 346

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
347 348
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
349
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
350 351
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
352
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
353 354
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
355
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
356 357
/* execute the computational delay */
static void codes_exec_comp_delay(
358
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
359 360 361 362 363 364 365 366 367 368 369 370 371 372 373
/* gets the next MPI operation from the network-workloads API. */
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* reverse handler of get next mpi operation. */
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);
/* reverse handler of next operation */
static void codes_issue_next_event_rc(tw_lp* lp);


///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
/* upon arrival of local completion message, inserts operation in completed send queue */
/* upon arrival of an isend operation, updates the arrival queue of the network */
374 375 376 377 378 379 380 381
static void update_completed_queue(
        nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp, dumpi_req_id req_id);
/* reverse of the above function */
static void update_completed_queue_rc(
        nw_state*s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp);
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse of the above function */
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* callback to a message sender for computing message time */
static void update_message_time(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse for computing message time */
static void update_message_time_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);

397
/*static void update_message_size_rc(
398 399 400 401
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
402
{*/
403
/*TODO: Complete reverse handler */
404
/*    (void)ns;
405 406 407
    (void)lp;
    (void)bf;
    (void)m;
408
}*/
409 410 411 412 413 414 415 416 417 418
/* update the message size */
static void update_message_size(
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m,
        mpi_msgs_queue * qitem,
        int is_eager,
        int is_send)
{
419 420 421
            (void)bf;
            (void)is_eager;

422 423 424
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
425
            if(ns->msg_sz_table == NULL)
426 427 428 429 430 431 432 433 434 435
                ns->msg_sz_table = qhash_init(msg_size_hash_compare, quickhash_64bit_hash, RANK_HASH_TABLE_SZ); 
            
            hash_link = qhash_search(ns->msg_sz_table, &(qitem->num_bytes));

            if(is_send)
                msg_init_time = m->fwd.sim_start_time;
            
            /* update hash table */
            if(!hash_link)
            {
436
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
437 438
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
439
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
440
                msg_info->avg_latency = msg_info->agg_latency;
441 442
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
443
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
444
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
445 446 447 448 449
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
450
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
451
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
452
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
453 454 455 456 457 458 459 460
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
461 462 463
    (void)ns;
    (void)bf;
    (void)m;
464 465 466 467 468
        
    int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
    
    for(int i = 0; i < num_jobs - 1; i++)
        tw_rand_reverse_unif(lp->rng); 
469 470 471 472 473 474 475 476
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
477 478 479
        (void)bf;
        (void)m;

480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
        struct codes_jobmap_id jid; 
        jid = codes_jobmap_to_local_id(ns->nw_id, jobmap_ctx);
        
        int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
        
        for(int other_id = 0; other_id < num_jobs; other_id++)
        {
            if(other_id == jid.job)
                continue;

            struct codes_jobmap_id other_jid;
            other_jid.job = other_id;

            int num_other_ranks = codes_jobmap_get_num_ranks(other_id, jobmap_ctx);

495
            lprintf("\n Other ranks %d ", num_other_ranks);
496
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
497 498 499 500 501 502 503 504 505 506 507
            tw_lpid global_dest_id;
     
            for(int k = 0; k < num_other_ranks; k++)    
            {
                other_jid.rank = k;
                int intm_dest_id = codes_jobmap_to_global_id(other_jid, jobmap_ctx); 
                global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

                tw_event * e;
                struct nw_message * m_new;  
                e = tw_event_new(global_dest_id, ts, lp);
508
                m_new = (struct nw_message*)tw_event_data(e);
509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
                m_new->msg_type = CLI_BCKGND_FIN;
                tw_event_send(e);   
            }
        }
        return;
}
static void notify_neighbor_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
       if(bf->c0)
       {
            notify_background_traffic_rc(ns, lp, bf, m);
            return;
       }
   
       if(bf->c1)
       {
          tw_rand_reverse_unif(lp->rng); 
       }
} 
static void notify_neighbor(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
    if(ns->local_rank == num_dumpi_traces - 1 
            && ns->is_finished == 1
            && ns->neighbor_completed == 1)
    {
542
//        printf("\n All workloads completed, notifying background traffic ");
543 544 545 546 547 548 549 550 551 552 553 554 555
        bf->c0 = 1;
        notify_background_traffic(ns, lp, bf, m);
        return;
    }
    
    struct codes_jobmap_id nbr_jid;
    nbr_jid.job = ns->app_id;
    tw_lpid global_dest_id;

    if(ns->is_finished == 1 && (ns->neighbor_completed == 1 || ns->local_rank == 0))
    {
        bf->c1 = 1;

556
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
557
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
558 559 560 561 562 563 564 565 566
        nbr_jid.rank = ns->local_rank + 1;
        
        /* Send a notification to the neighbor about completion */
        int intm_dest_id = codes_jobmap_to_global_id(nbr_jid, jobmap_ctx); 
        global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);
       
        tw_event * e;
        struct nw_message * m_new;  
        e = tw_event_new(global_dest_id, ts, lp);
567
        m_new = (struct nw_message*)tw_event_data(e); 
568 569 570 571 572 573 574 575 576 577
        m_new->msg_type = CLI_NBR_FINISH;
        tw_event_send(e);   
    }
}
void finish_bckgnd_traffic_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
578 579 580 581
        (void)b;
        (void)msg;
        (void)lp;

582 583 584 585 586 587 588 589 590
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
591 592
        (void)b;
        (void)msg;
593
        ns->is_finished = 1;
594
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
595
        
596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624
        return;
}

void finish_nbr_wkld_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 0;
    
    notify_neighbor_rc(ns, lp, b, msg);
}

void finish_nbr_wkld(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 1;

    notify_neighbor(ns, lp, b, msg);
}
static void gen_synthetic_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(bf->c0)
        return;

625 626 627 628 629 630
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
631
        s->prev_switch = m->rc.saved_prev_switch;
632
//        s->saved_perm_dest = s->rc_perm;
633 634
        tw_rand_reverse_unif(lp->rng);
    }
635 636 637 638 639
    int i;
    for (i=0; i < m->rc.saved_syn_length; i++){
        model_net_event_rc2(lp, &m->event_rc);
        s->gen_data -= payload_sz;
        num_syn_bytes_sent -= payload_sz;
640
        s->num_bytes_sent -= payload_sz;
641 642
    }
        tw_rand_reverse_unif(lp->rng);
643
        s->num_sends--;
644 645 646 647 648 649 650 651 652 653 654 655 656 657

}

/* generate synthetic traffic */
static void gen_synthetic_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(s->is_finished == 1)
    {
        bf->c0 = 1;
        return;
    }

    /* Get job information */
    tw_lpid global_dest_id;
658 659
    int intm_dest_id;
    nw_message remote_m;
660 661 662 663 664 665

    struct codes_jobmap_id jid;
    jid = codes_jobmap_to_local_id(s->nw_id, jobmap_ctx); 

    int num_clients = codes_jobmap_get_num_ranks(jid.job, jobmap_ctx);

666 667 668 669
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
670
    {
671 672
        case UNIFORM:
        {
673
            bf->c1 = 1;
674 675 676 677 678 679 680
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
            dest_svr[0] = tw_rand_integer(lp->rng, 0, num_clients - 1);
            if(dest_svr[0] == s->local_rank)
                dest_svr[0] = (s->local_rank + 1) % num_clients;
        }
        break;
681 682 683

        case PERMUTATION:
        {
684 685
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

686 687
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
688 689
            //if(s->gen_data - s->prev_switch >= perm_switch_thresh)
            /*if(s->saved_perm_dest == -1)
690
            {
691
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
692
                bf->c2 = 1;
693
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
694 695
                dest_svr[0] = tw_rand_integer(lp->rng, 0, num_clients - 1);
                if(dest_svr[0] == s->local_rank)
696 697 698 699 700
                    dest_svr[0] = (s->local_rank + num_clients/2) % num_clients;// */
                /* TODO: Fix random number generation code */
//                s->rc_perm = s->saved_perm_dest;
                /*s->saved_perm_dest = dest_svr[0];
                assert(s->saved_perm_dest != s->local_rank);
701 702
            }
            else
703 704 705 706 707 708 709
                dest_svr[0] = s->saved_perm_dest;*/
            if(s->local_rank != INCAST_ID)
                dest_svr[0] = INCAST_ID;
            else
                dest_svr[0] = 0;

            assert(dest_svr[0] != s->local_rank);
710 711
        }
        break;
712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780
        case NEAREST_NEIGHBOR:
        {
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
            dest_svr[0] = (s->local_rank + 1) % num_clients;
        }
        break;
        case ALLTOALL:
        {
            dest_svr = (int*) calloc(num_clients-1, sizeof(int));
            int index = 0;
            for (i=0;i<num_clients;i++)
            {
                if(i!=s->local_rank) 
                {
                    dest_svr[index] = i;
                    index++;
                    length++;
                }
            }
        }
        break;
        case STENCIL:  //2D 4-point stencil
        {
            /* I think this code snippet is coming from the LLNL stencil patterns. */
            int digits, x=1, y=1, row, col, temp=num_clients;
            length = 4;
            dest_svr = (int*) calloc(4, sizeof(int));
            for (digits = 0; temp > 0; temp >>= 1)
                digits++;
            digits = digits/2;
            for (i = 0; i < digits; i++)
                x = x * 2;
            y = num_clients / x;
            //printf("\nStencil Syn: x=%d, y=%d", x, y);
            row = s->local_rank / y;
            col = s->local_rank % y;

            dest_svr[0] = row * y + ((col-1+y)%y);   /* left neighbor */
            dest_svr[1] = row * y + ((col+1+y)%y);   /* right neighbor */
            dest_svr[2] = ((row-1+x)%x) * y + col;   /* bottom neighbor */
            dest_svr[3] = ((row+1+x)%x) * y + col;   /* up neighbor */
        }
        break;
        default:
            tw_error(TW_LOC, "Undefined traffic pattern");
    }   
    /* Record length for reverse handler*/
    m->rc.saved_syn_length = length;

    if(length > 0)
    {
        // m->event_array_rc = (model_net_event_return) malloc(length * sizeof(model_net_event_return));
        //printf("\nRANK %d Dests %d", s->local_rank, length);
        for (i = 0; i < length; i++)
        {
            /* Generate synthetic traffic */
            jid.rank = dest_svr[i];
            intm_dest_id = codes_jobmap_to_global_id(jid, jobmap_ctx); 
            global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

            remote_m.fwd.sim_start_time = tw_now(lp);
            remote_m.fwd.dest_rank = dest_svr[i];
            remote_m.msg_type = CLI_BCKGND_ARRIVE;
            remote_m.fwd.num_bytes = payload_sz;
            remote_m.fwd.app_id = s->app_id;
            remote_m.fwd.src_rank = s->local_rank;

            // printf("\nAPP %d SRC %d Dest %d (twid %llu)", jid.job, s->local_rank, dest_svr[i], global_dest_id);
781
            m->event_rc = model_net_event(net_id, "medium", global_dest_id, payload_sz, 0.0, 
782 783 784 785
                    sizeof(nw_message), (const void*)&remote_m, 
                    0, NULL, lp);
            
            s->gen_data += payload_sz;
786
            s->num_bytes_sent += payload_sz;
787 788
            num_syn_bytes_sent += payload_sz; 
        }
789
    }
790 791
    s->num_sends++;

792
    /* New event after MEAN_INTERVAL */  
793
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
794 795 796
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
797
    m_new = (struct nw_message*)tw_event_data(e);
798 799
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
800 801 802
    
    if(s->num_sends == MAX_SYN_SENDS)
        s->is_finished = 1;
803 804

    free(dest_svr);
805 806 807 808
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
809 810 811
    (void)bf;
    (void)m;
    (void)lp;
812
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
813
    s->num_recvs--;
814 815 816
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
817
    s->num_bytes_recvd -= data;
818
    s->send_time = m->rc.saved_send_time;
819 820 821
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
822 823 824
    (void)bf;
    (void)lp;

825 826 827
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    if(s->local_rank == 0)
     {
828
//    	printf("\n Data arrived %lld rank %llu total data %ld ", m->fwd.num_bytes, s->nw_id, s->syn_data);
829
/*	if(s->syn_data > upper_threshold)
mubarak's avatar
mubarak committed
830 831 832 833
    if(s->local_rank == 0)
     {
    	printf("\n Data arrived %lld rank %llu total data %ld ", m->fwd.num_bytes, s->nw_id, s->syn_data);
	if(s->syn_data > upper_threshold)
834 835 836 837 838 839 840
	{ 
        	struct rusage mem_usage;
		int who = RUSAGE_SELF;
		int err = getrusage(who, &mem_usage);
		printf("\n Memory usage %lf gigabytes", ((double)mem_usage.ru_maxrss / (1024.0 * 1024.0)));
		upper_threshold += 1048576;
	}*/
mubarak's avatar
mubarak committed
841
	}
842
    m->rc.saved_send_time = s->send_time;
843
    if((tw_now(lp) - m->fwd.sim_start_time) > s->max_time)
844 845
        s->max_time = tw_now(lp) - m->fwd.sim_start_time;

846 847
    s->send_time += (tw_now(lp) - m->fwd.sim_start_time);
    s->num_recvs++;
848 849
    int data = m->fwd.num_bytes;
    s->syn_data += data;
850
    s->num_bytes_recvd += data;
851 852
    num_syn_bytes_recvd += data;
}
853
/* Debugging functions, may generate unused function warning */
854
/*static void print_waiting_reqs(uint32_t * reqs, int count)
855
{
856
    lprintf("\n Waiting reqs: %d count", count);
857 858
    int i;
    for(i = 0; i < count; i++ )
859
        lprintf(" %d ", reqs[i]);
860
}*/
861 862 863 864 865 866 867 868 869 870 871 872
static void print_msgs_queue(struct qlist_head * head, int is_send)
{
    if(is_send)
        printf("\n Send msgs queue: ");
    else
        printf("\n Recv msgs queue: ");

    struct qlist_head * ent = NULL;
    mpi_msgs_queue * current = NULL;
    qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, mpi_msgs_queue, ql);
mubarak's avatar
mubarak committed
873
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
874 875
       }
}
Misbah Mubarak's avatar
Misbah Mubarak committed
876
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
877
{
878
//    printf("\n Completed queue: ");
879 880
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
881
      tw_output(lp, "\n");
882 883 884
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
885
            tw_output(lp, " %llu ", current->req_id);
886
       }
Misbah Mubarak's avatar
Misbah Mubarak committed
887
}
888
static int clear_completed_reqs(nw_state * s,
889
        tw_lp * lp,
Misbah Mubarak's avatar
Misbah Mubarak committed
890
        unsigned int * reqs, int count)
891
{
892 893 894
    (void)s;
    (void)lp;

895
    int i, matched = 0;
896

897 898 899
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
900 901 902
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

903
      int index = 0;
904 905
      qlist_for_each(ent, &s->completed_reqs)
       {
906 907
           if(prev)
           {
908
              rc_stack_push(lp, prev, free, s->matched_reqs);
909 910
              prev = NULL;
           }
911
            
912 913
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
914 915
            if(current->req_id == reqs[i])
            {
916
                ++matched;
917
                qlist_del(&current->ql);
918
                prev = current;
919
            }
920
            ++index;
921
       }
922 923

      if(prev)
924 925 926 927
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
928
    }
929
    return matched;
930
}
931
static void add_completed_reqs(nw_state * s,
932 933
        tw_lp * lp,
        int count)
934
{
935
    (void)lp;
936
    for(int i = 0; i < count; i++)
937
    {
938 939
       struct completed_requests * req = (struct completed_requests*)rc_stack_pop(s->matched_reqs);
       // turn on only if wait-all unmatched error arises in optimistic mode.
940
       qlist_add(&req->ql, &s->completed_reqs);
941
    }//end for
942
}
943

944 945 946 947 948 949
/* helper function - maps an MPI rank to an LP id */
static tw_lpid rank_to_lpid(int rank)
{
    return codes_mapping_get_lpid_from_relative(rank, NULL, "nw-lp", NULL, 0);
}

950
static int notify_posted_wait(nw_state* s,
951
        tw_bf * bf, nw_message * m, tw_lp * lp,
Misbah Mubarak's avatar
Misbah Mubarak committed
952
        unsigned int completed_req)
953
{
954 955
    (void)bf;

956 957
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
958

959
    m->fwd.wait_completed = 0;
960