model-net-mpi-replay.c 89.7 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>
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
#define MAX_WAIT_REQS 512
25
#define CS_LP_DBG 1
26
#define RANK_HASH_TABLE_SZ 2000
27
#define NW_LP_NM "nw-lp"
28 29 30
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
31

32 33 34
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

35
static unsigned long perm_switch_thresh = 8388608;
36

37
/* NOTE: Message tracking works in sequential mode only! */
38
static int debug_cols = 0;
39 40 41
/* 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;
42
static int enable_msg_tracking = 0;
43
static int is_synthetic = 0;
44
tw_lpid TRACK_LP = -1;
45
int nprocs = 0;
46
static double total_syn_data = 0;
47
static int unmatched = 0;
48
char workload_type[128];
49
char workload_name[128];
50 51 52 53
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
54
static int num_dumpi_traces = 0;
55
static int64_t EAGER_THRESHOLD = 8192;
56

57
static long num_ops = 0;
58
static int upper_threshold = 1048576;
59
static int alloc_spec = 0;
60 61
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
62
static int payload_sz = 1024;
63 64

/* Doing LP IO*/
65
static char * params = NULL;
66
static char lp_io_dir[256] = {'\0'};
67
static char sampling_dir[32] = {'\0'};
68
static char mpi_msg_dir[32] = {'\0'};
69 70 71 72
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

73 74 75 76
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
77 78 79
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
80 81 82 83 84
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

85 86
/* Variables for Cortex Support */
/* Matthieu's additions start */
87
#ifdef ENABLE_CORTEX_PYTHON
88 89 90
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
91
#endif
92 93
/* Matthieu's additions end */

94 95
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
96
typedef unsigned int dumpi_req_id;
97 98

static int net_id = 0;
99
static float noise = 1.0;
100 101 102
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
103
static int syn_type = 0;
104

105
FILE * workload_log = NULL;
106
FILE * msg_size_log = NULL;
107 108 109 110
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
111

112 113 114
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

115 116 117
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

118 119 120 121 122 123
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;
124 125 126
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
127
static int enable_debug = 0;
128

129 130 131
/* set group context */
struct codes_mctx group_ratio;

132
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
133
* 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.
134 135 136 137 138 139 140
* 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,
141 142 143 144 145 146
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
147 148
};

149 150 151 152 153 154
/* 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 */
155 156
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
157
};
158 159 160 161
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
162
    int app_id;
163 164 165 166 167
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
168 169 170 171 172 173 174
/* 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;
175
    int64_t num_bytes;
176
    int64_t seq_id;
177 178 179 180 181 182 183 184
    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
{
185
	unsigned int req_id;
186
    struct qlist_head ql;
187
    int index;
188 189 190 191 192 193
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
194
    unsigned int req_ids[MAX_WAIT_REQS];
195
	int num_completed;
196 197
	int count;
    tw_stime start_time;
198 199 200
    struct qlist_head ql;
};

201 202 203 204 205 206
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
207
    struct qhash_head  hash_link;
208 209
    struct qlist_head ql; 
};
210 211 212 213 214 215 216 217 218 219
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;
220 221
    int app_id;
    int local_rank;
222

223
    int synthetic_pattern;
224 225 226
    int is_finished;
    int neighbor_completed;

227
    struct rc_stack * processed_ops;
228
    struct rc_stack * processed_wait_op;
229
    struct rc_stack * matched_reqs;
230
//    struct rc_stack * indices;
231 232 233 234 235 236 237 238 239 240

    /* 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;

241

242 243
	/* time spent by the LP in executing the app trace*/
	double start_time;
244 245 246 247 248 249 250 251 252

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
	/* 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;
268

269
    tw_stime cur_interval_end;
270
    
271 272
    /* Pending wait operation */
    struct pending_waits * wait_op;
273

274 275 276 277 278 279
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

280 281 282
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

283 284
    unsigned long syn_data;
    unsigned long gen_data;
285 286 287 288 289
  
    unsigned long prev_switch;
    unsigned long saved_perm_dest;
    unsigned long rc_perm;

290 291 292 293
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
294
    char output_buf[512];
295
    char col_stats[64];
296 297 298 299
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
300
* 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.
301 302 303
* 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
{
304
   // forward message handler
305
   int msg_type;
306
   int op_type;
307
   model_net_event_return event_rc;
308
   struct codes_workload_op * mpi_op;
309

310 311 312
   struct
   {
       tw_lpid src_rank;
313
       int dest_rank;
314
       int64_t num_bytes;
315 316 317 318 319
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
320
       unsigned int req_id;
321
       int matched_req;
322
       int tag;
323
       int app_id;
324 325
       int found_match;
       short wait_completed;
326
       short rend_send;
327 328 329 330 331 332 333
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
334
       int64_t saved_num_bytes;
335
       int saved_syn_length;
336
       unsigned long saved_prev_switch;
337
   } rc;
338 339
};

340
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);
341 342

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
343 344
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
345
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
346 347
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
348
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
349 350
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
351
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
352 353
/* execute the computational delay */
static void codes_exec_comp_delay(
354
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
355 356 357 358 359 360 361 362 363 364 365 366 367 368 369
/* 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 */
370 371 372 373 374 375 376 377
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);
378 379 380 381 382 383 384 385 386 387 388 389 390 391 392
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);

393
/*static void update_message_size_rc(
394 395 396 397
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
398
{*/
399
/*TODO: Complete reverse handler */
400
/*    (void)ns;
401 402 403
    (void)lp;
    (void)bf;
    (void)m;
404
}*/
405 406 407 408 409 410 411 412 413 414
/* 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)
{
415 416 417
            (void)bf;
            (void)is_eager;

418 419 420
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
421
            if(ns->msg_sz_table == NULL)
422 423 424 425 426 427 428 429 430 431
                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)
            {
432
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
433 434
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
435
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
436
                msg_info->avg_latency = msg_info->agg_latency;
437 438
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
439
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
440
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
441 442 443 444 445
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
446
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
447
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
448
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
449 450 451 452 453 454 455 456
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
457 458 459
    (void)ns;
    (void)bf;
    (void)m;
460 461 462 463 464
        
    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); 
465 466 467 468 469 470 471 472
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
473 474 475
        (void)bf;
        (void)m;

476 477 478 479 480 481 482 483 484 485 486 487 488 489 490
        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);

491
            lprintf("\n Other ranks %d ", num_other_ranks);
492 493 494 495 496 497 498 499 500 501 502 503
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
            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);
504
                m_new = (struct nw_message*)tw_event_data(e);
505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537
                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)
    {
538
//        printf("\n All workloads completed, notifying background traffic ");
539 540 541 542 543 544 545 546 547 548 549 550 551
        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;

552
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
553 554 555 556 557 558 559 560 561 562
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
        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);
563
        m_new = (struct nw_message*)tw_event_data(e); 
564 565 566 567 568 569 570 571 572 573
        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)
{
574 575 576 577
        (void)b;
        (void)msg;
        (void)lp;

578 579 580 581 582 583 584 585 586
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
587 588
        (void)b;
        (void)msg;
589
        ns->is_finished = 1;
590
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619
        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;

620 621 622 623 624 625
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
626
        s->prev_switch = m->rc.saved_prev_switch;
627 628 629
        s->saved_perm_dest = s->rc_perm;
        tw_rand_reverse_unif(lp->rng);
    }
630 631 632 633 634 635 636
    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;
    }
        tw_rand_reverse_unif(lp->rng);
637 638 639 640 641 642 643 644 645 646 647 648 649 650

}

/* 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;
651 652
    int intm_dest_id;
    nw_message remote_m;
653 654 655 656 657 658

    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);

659 660 661 662
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
663
    {
664 665
        case UNIFORM:
        {
666
            bf->c1 = 1;
667 668 669 670 671 672 673
            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;
674 675 676

        case PERMUTATION:
        {
677 678
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

679 680
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
681
            if(s->gen_data - s->prev_switch >= perm_switch_thresh)
682
            {
683
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
684
                bf->c2 = 1;
685
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
686 687 688 689 690 691 692 693 694 695
                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 + num_clients/2) % num_clients;
                s->rc_perm = s->saved_perm_dest;
                s->saved_perm_dest = dest_svr[0];
            }
            else
                dest_svr[0] = s->saved_perm_dest;
        }
        break;
696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771
        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);
            m->event_rc = model_net_event(net_id, "synthetic-tr", global_dest_id, payload_sz, 0.0, 
                    sizeof(nw_message), (const void*)&remote_m, 
                    0, NULL, lp);
            
            s->gen_data += payload_sz;
            num_syn_bytes_sent += payload_sz; 
        }
772 773
    }
    /* New event after MEAN_INTERVAL */  
774
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
775 776 777
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
778
    m_new = (struct nw_message*)tw_event_data(e);
779 780
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
781 782

    free(dest_svr);
783 784 785 786
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
787 788 789
    (void)bf;
    (void)m;
    (void)lp;
790 791 792 793 794 795 796
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
797 798 799
    (void)bf;
    (void)lp;

800 801 802 803 804
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    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)
805 806 807 808
    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)
809 810 811 812 813 814 815
	{ 
        	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;
	}*/
816
	}
817 818 819 820
    int data = m->fwd.num_bytes;
    s->syn_data += data;
    num_syn_bytes_recvd += data;
}
821
/* Debugging functions, may generate unused function warning */
822
/*static void print_waiting_reqs(uint32_t * reqs, int count)
823
{
824
    lprintf("\n Waiting reqs: %d count", count);
825 826
    int i;
    for(i = 0; i < count; i++ )
827
        lprintf(" %d ", reqs[i]);
828
}*/
829 830 831 832 833 834 835 836 837 838 839 840
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);
841
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
842 843
       }
}
844
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
845
{
846
//    printf("\n Completed queue: ");
847 848
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
849
      tw_output(lp, "\n");
850 851 852
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
853
            tw_output(lp, " %llu ", current->req_id);
854
       }
855
}
856
static int clear_completed_reqs(nw_state * s,
857
        tw_lp * lp,
858
        unsigned int * reqs, int count)
859
{
860 861 862
    (void)s;
    (void)lp;

863
    int i, matched = 0;
864

865 866 867
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
868 869 870
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

871
      int index = 0;
872 873
      qlist_for_each(ent, &s->completed_reqs)
       {
874 875
           if(prev)
           {
876
              rc_stack_push(lp, prev, free, s->matched_reqs);
877 878
              prev = NULL;
           }
879
            
880 881
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
882 883
            if(current->req_id == reqs[i])
            {
884
                ++matched;
885
                qlist_del(&current->ql);
886
                prev = current;
887
            }
888
            ++index;
889
       }
890 891

      if(prev)
892 893 894 895
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
896
    }
897
    return matched;
898
}
899
static void add_completed_reqs(nw_state * s,
900 901
        tw_lp * lp,
        int count)
902
{
903
    (void)lp;
904
    for(int i = 0; i < count; i++)
905
    {
906 907
       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.
908
       qlist_add(&req->ql, &s->completed_reqs);
909
    }//end for
910
}
911

912 913 914 915 916 917
/* 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);
}

918
static int notify_posted_wait(nw_state* s,
919
        tw_bf * bf, nw_message * m, tw_lp * lp,
920
        unsigned int completed_req)
921
{
922 923
    (void)bf;

924 925
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
926

927
    m->fwd.wait_completed = 0;
928

929 930
    if(!wait_elem)
        return 0;
931

932
    int op_type = wait_elem->op_type;
933

934 935 936
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
937
            m->fwd.wait_completed = 1;
938 939
            wait_completed = 1;
    }
940 941
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
942 943 944 945 946 947
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
948
            {
949
                wait_elem->num_completed++;
950
                if(wait_elem->num_completed > wait_elem->count)
951
                    printf("\n Num completed %d count %d LP %llu ",
952 953
                            wait_elem->num_completed,
                            wait_elem->count,
954
                            LLU(lp->gid));
955 956
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
957

958
                if(wait_elem->num_completed >= wait_elem->count)
959
                {
960
                    if(enable_debug)
961
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
962
                    wait_completed = 1;
963
                }
964

965
                m->fwd.wait_completed = 1;
966
            }
967
        }
968
    }
969
    return wait_completed;
970
}
971

972
/* reverse handler of MPI wait operation */
973
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
974
{
975
   if(bf->c1)
976
    {
977
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
978 979
        if(m->fwd.found_match == 0)
        {
980
            qlist_add(&qi->ql, &s->completed_reqs);
981
        }
982 983 984 985 986 987 988 989 990 991 992 993 994
        else
        {
           int index = 1;
           struct qlist_head * ent = NULL;
           qlist_for_each(ent, &s->completed_reqs)
           {
                if(index == m->fwd.found_match)
                {
                    qlist_add(&qi->ql, ent);
                    break;
                }
                index++;
           }
995
        }
996
        codes_issue_next_event_rc(lp);
997
        return;
998
    }
999 1000 1001
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
1002
}
1003

1004
/* execute MPI wait operation */
1005
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1006
{
1007
    /* check in the completed receives queue if the request ID has already been completed.*/
1008 1009
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
1010
    assert(!s->wait_op);
1011
    unsigned int req_id = mpi_op->u.wait.req_id;
1012

1013
    struct completed_requests* current = NULL;
1014

1015
    struct qlist_head * ent = NULL;
1016
    int index = 0;
1017 1018 1019 1020 1021
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
1022
            bf->c1=1;
1023
            qlist_del(&current->ql);
1024
            rc_stack_push(lp, current, free, s->processed_ops);
1025
            codes_issue_next_event(lp);
1026
            m->fwd.found_match = index;
1027
            if(s->nw_id == (tw_lpid)TRACK_LP)
1028 1029 1030
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
1031
            }
1032 1033
            return;
        }
1034
        ++index;
1035
    }
1036

1037 1038 1039 1040 1041
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
1042
    /* If not, add the wait operation in the pending 'waits' list. */
1043
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1044 1045 1046
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
1047 1048
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
1049
    s->wait_op = wait_op;
1050

1051
    return;
1052 1053
}

1054
static void codes_exec_mpi_wait_all_rc(
1055
        nw_state* s,
1056 1057
        tw_bf * bf,
        nw_message * m,
1058
        tw_lp* lp)
1059
{
1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070
  if(bf->c1)
  {
    int sampling_indx = s->sampling_indx;
    s->mpi_wkld_samples[sampling_indx].num_waits_sample--;

    if(bf->c2)
    {
        s->cur_interval_end -= sampling_interval;
        s->sampling_indx--;
    }
  }
1071 1072 1073 1074 1075 1076 1077 1078
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1079
      add_completed_reqs(s, lp, m->fwd.num_matched);
1080 1081 1082
      codes_issue_next_event_rc(lp);
  }
  return;
1083
}
1084

1085
static void codes_exec_mpi_wait_all(
1086
        nw_state* s,
1087 1088
        tw_bf * bf,
        nw_message * m,
1089
        tw_lp* lp,
1090
        struct codes_workload_op * mpi_op)
1091
{
1092
  if(enable_debug)
1093
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1094

1095 1096 1097 1098 1099 1100 1101 1102
  if(enable_sampling)
  {
    bf->c1 = 1;
    if(tw_now(lp) >= s->cur_interval_end)
    {
        bf->c2 = 1;
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].nw_id = s->nw_id;
1103
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1104 1105 1106 1107 1108 1109
        s->mpi_wkld_samples[indx].sample_end_time = s->cur_interval_end;
        s->cur_interval_end += sampling_interval;
        s->sampling_indx++;
    }
    if(s->sampling_indx >= MAX_STATS)
    {
1110
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1111 1112 1113 1114 1115 1116 1117 1118
        memcpy(tmp, s->mpi_wkld_samples, s->sampling_indx);
        free(s->mpi_wkld_samples);
        s->mpi_wkld_samples = tmp;
        s->max_arr_size += MAX_STATS;
    }
    int indx = s->sampling_indx;
    s->mpi_wkld_samples[indx].num_waits_sample++;
  }
1119
  int count = mpi_op->u.waits.count;
1120 1121
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1122

1123
  int i = 0, num_matched = 0;
1124
  m->fwd.num_matched = 0;
1125

1126
  /*if(lp->gid == TRACK_LP)
1127
  {
1128
      printf("\n MPI Wait all posted ");
1129
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1130 1131
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1132
      /* check number of completed irecvs in the completion queue */
1133 1134
  for(i = 0; i < count; i++)
  {
1135
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1136 1137 1138 1139
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1140
            current = qlist_entry(ent, struct completed_requests, ql);
1141 1142 1143 1144
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1145

1146
  m->fwd.found_match = num_matched;
1147 1148 1149 1150
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1151 1152 1153
      m->fwd.num_matched = clear_completed_reqs(s, lp, mpi_op->u.waits.req_ids, count);
      struct pending_waits* wait_op = s->wait_op;
      free(wait_op);
1154 1155
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1156 1157
  }
  else
1158 1159
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1160
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1161 1162 1163 1164 1165 1166 1167 1168
	  wait_op->count = count;
      wait_op->op_type = mpi_op->op_type;
      assert(count < MAX_WAIT_REQS);

      for(i = 0; i < count; i++)
          wait_op->req_ids[i] =  mpi_op->u.waits.req_ids[i];

	  wait_op->num_completed = num_matched;
1169
	  wait_op->start_time = tw_now(lp);
1170
      s->wait_op = wait_op;
1171
  }
1172 1173
  return;
}
1174

1175 1176
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1177
 * Index is used for inserting the element once again in the queue for reverse computation. */
1178
static int rm_matching_rcv(nw_state * ns,
1179
        tw_bf * bf,
1180 1181
        nw_message * m,
        tw_lp * lp,
1182
        mpi_msgs_queue * qitem)
1183 1184
{
    int matched = 0;
1185
    int index = 0;
1186
    int is_rend = 0;
1187 1188
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1189

1190 1191
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1192 1193
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1194
               ((qi->tag == qitem->tag) || qi->tag == -1)
1195
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1196 1197
        {
            matched = 1;
1198
            qi->num_bytes = qitem->num_bytes;
1199 1200
            break;
        }
1201
        ++index;
1202
    }
1203

1204 1205
    if(matched)
    {
1206 1207 1208 1209 1210 1211 1212
        if(enable_msg_tracking && qitem->num_bytes < EAGER_THRESHOLD)
        {
            update_message_size(ns, lp, bf, m, qitem, 1, 1);
        }
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
1213 1214
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
1215 1216
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qitem, qi->req_id);
1217
        }
1218 1219
        else
        {
1220
            bf->c12 = 1;
1221 1222 1223 1224
            m->rc.saved_recv_time = ns->recv_time;
            ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
        }
        if(qi->op_type == CODES_WK_IRECV && !is_rend)
1225
        {
1226
            bf->c9 = 1;
1227 1228 1229 1230
            /*if(ns->nw_id == (tw_lpid)TRACK_LP)
            {
                printf("\n Completed irecv req id %d ", qi->req_id);
            }*/
1231
            update_completed_queue(ns, bf, m, lp, qi->req_id);
1232
        }
1233 1234 1235
        else if(qi->op_type == CODES_WK_RECV && !is_rend)
        {
            bf->c8 = 1;
1236
            codes_issue_next_event(lp);
1237
        }
1238

1239
        qlist_del(&qi->ql);
1240

1241
        rc_stack_push(lp, qi, free, ns->processed_ops);
1242
        return index;
1243 1244 1245 1246
    }
    return -1;
}

1247
static int rm_matching_send(nw_state * ns,
1248 1249 1250
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
1251 1252 1253 1254 1255
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

1256
    int index = 0;
1257 1258
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1259 1260 1261
        if(//(qi->num_bytes == qitem->num_bytes) // it is not a requirement in MPI that the send and receive sizes match
                // && 
		(qi->tag == qitem->tag || qitem->tag == -1)
1262 1263
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
1264
            qitem->num_bytes = qi->num_bytes;
1265 1266 1267
            matched = 1;
            break;
        }
1268
        ++index;
1269 1270 1271 1272
    }

    if(matched)
    {
1273 1274 1275
        if(enable_msg_tracking && (qi->num_bytes < EAGER_THRESHOLD))
            update_message_size(ns, lp, bf, m, qi, 1, 0);
        
1276 1277
        m->fwd.matched_req = qitem->req_id;
        int is_rend = 0;
1278 1279 1280 1281
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
             * the data */
1282
            bf->c10 = 1;
1283 1284
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qi, qitem->req_id);
1285
        }
Misbah Mubarak's avatar