model-net-mpi-replay.c 90.4 KB
Newer Older
1 2 3 4 5 6 7
/*
 * Copyright (C) 2014 University of Chicago.
 * 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 512
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
/* NOTE: Message tracking works in sequential mode only! */
36
static int debug_cols = 0;
37 38 39
/* 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;
40
static int enable_msg_tracking = 1;
41
static int is_synthetic = 0;
42
tw_lpid TRACK_LP = -1;
43
int nprocs = 0;
44
static double total_syn_data = 0;
45
static int unmatched = 0;
46
char workload_type[128];
47
char workload_name[128];
48 49 50 51
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
52
static int num_dumpi_traces = 0;
53
static int64_t EAGER_THRESHOLD = 8192;
54

55 56
static long num_ops = 0;
static upper_threshold = 1048576;
57
static int alloc_spec = 0;
58 59
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
60
static int payload_sz = 1024;
61 62

/* Doing LP IO*/
63
static char * params = NULL;
64
static char lp_io_dir[256] = {'\0'};
65
static char sampling_dir[32] = {'\0'};
66
static char mpi_msg_dir[32] = {'\0'};
67 68 69 70
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

71 72 73 74
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
75 76 77
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
78 79 80 81 82
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

83 84
/* Variables for Cortex Support */
/* Matthieu's additions start */
85
#ifdef ENABLE_CORTEX_PYTHON
86 87 88
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
89
#endif
90 91
/* Matthieu's additions end */

92 93
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
Misbah Mubarak's avatar
Misbah Mubarak committed
94
typedef unsigned int dumpi_req_id;
95 96

static int net_id = 0;
97
static float noise = 0.1;
98 99 100
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
101
static int syn_type = 0;
102

103
FILE * workload_log = NULL;
104
FILE * msg_size_log = NULL;
105 106 107 108
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
109

110 111 112
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

113 114 115
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

116 117 118 119 120 121
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;
122 123 124
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
125
static int enable_debug = 0;
126

127 128 129
/* set group context */
struct codes_mctx group_ratio;

130
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
131
* 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.
132 133 134 135 136 137 138
* 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,
139 140 141 142 143 144
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
145 146
};

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

/* 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
191
    unsigned int req_ids[MAX_WAIT_REQS];
192
	int num_completed;
193 194
	int count;
    tw_stime start_time;
195 196 197
    struct qlist_head ql;
};

198 199 200 201 202 203
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
204
    struct qhash_head hash_link;
205 206
    struct qlist_head ql; 
};
207 208 209 210 211 212 213 214 215 216
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;
217 218
    int app_id;
    int local_rank;
219

220
    int synthetic_pattern;
221 222 223
    int is_finished;
    int neighbor_completed;

224
    struct rc_stack * processed_ops;
225
    struct rc_stack * processed_wait_op;
226
    struct rc_stack * matched_reqs;
227
//    struct rc_stack * indices;
228 229 230 231 232 233 234 235 236 237

    /* 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;

238

239 240
	/* time spent by the LP in executing the app trace*/
	double start_time;
241 242 243 244 245 246 247 248 249

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
	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;
265

266
    tw_stime cur_interval_end;
267
    
268 269
    /* Pending wait operation */
    struct pending_waits * wait_op;
270

271 272 273 274 275 276
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

277 278 279
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

280 281 282
    unsigned long syn_data;
    unsigned long gen_data;
    
283 284 285 286
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
287
    char output_buf[512];
288
    //char output_buf2[512];
289
    char col_stats[64];
290 291 292 293
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
294
* 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.
295 296 297
* 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
{
298
   // forward message handler
299
   int msg_type;
300
   int op_type;
301
   model_net_event_return event_rc;
302

303 304 305
   struct
   {
       tw_lpid src_rank;
306
       int dest_rank;
307
       int64_t num_bytes;
308 309 310 311 312
       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
313
       unsigned int req_id;
314
       int matched_req;
315
       int tag;
316
       int app_id;
317 318
       int found_match;
       short wait_completed;
319
       short rend_send;
320 321 322 323 324 325 326
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
327
       int64_t saved_num_bytes;
328
       int saved_syn_length;
329
       struct codes_workload_op * mpi_op;
330
   } rc;
331 332
};

333
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);
334 335

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
336 337
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
338
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
339 340
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
341
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
342 343
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
344
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
345 346
/* execute the computational delay */
static void codes_exec_comp_delay(
347
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
348 349 350 351 352 353 354 355 356 357 358 359 360 361 362
/* 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 */
363 364 365 366 367 368 369 370
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);
371 372 373 374 375 376 377 378 379 380 381 382 383 384 385
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);

386
/*static void update_message_size_rc(
387 388 389 390
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
391
{*/
392
/*TODO: Complete reverse handler */
393
/*    (void)ns;
394 395 396
    (void)lp;
    (void)bf;
    (void)m;
397
}*/
398 399 400 401 402 403 404 405 406 407
/* 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)
{
408 409 410
            (void)bf;
            (void)is_eager;

411 412 413
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
414
            if(ns->msg_sz_table == NULL)
415 416 417 418 419 420 421 422 423 424
                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)
            {
425
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
426 427
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
428
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
429
                msg_info->avg_latency = msg_info->agg_latency;
430
                assert(ns->msg_sz_table);
431
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", msg_info->msg_size, msg_info->agg_latency, msg_info->num_msgs);
432
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
433
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
434
                
435 436 437 438 439
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
440
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
441
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
442
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
443 444 445 446 447 448 449 450
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
451 452 453
    (void)ns;
    (void)bf;
    (void)m;
454 455 456 457 458
        
    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); 
459 460 461 462 463 464 465 466
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
467 468 469
        (void)bf;
        (void)m;

470 471 472 473 474 475 476 477 478 479 480 481 482 483 484
        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);

485
            lprintf("\n Other ranks %d ", num_other_ranks);
486 487 488 489 490 491 492 493 494 495 496 497
            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);
498
                m_new = (struct nw_message*)tw_event_data(e);
499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531
                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)
    {
532
//        printf("\n All workloads completed, notifying background traffic ");
533 534 535 536 537 538 539 540 541 542 543 544 545
        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;

546
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
547 548 549 550 551 552 553 554 555 556
        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);
557
        m_new = (struct nw_message*)tw_event_data(e); 
558 559 560 561 562 563 564 565 566 567
        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)
{
568 569 570 571
        (void)b;
        (void)msg;
        (void)lp;

572 573 574 575 576 577 578 579 580
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
581 582
        (void)b;
        (void)msg;
583
        ns->is_finished = 1;
584
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613
        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;

614 615 616 617 618 619 620
    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);
621 622 623 624 625 626 627 628 629 630 631 632 633 634

}

/* 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;
635 636
    int intm_dest_id;
    nw_message remote_m;
637 638 639 640 641 642

    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);

643 644 645 646
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
647
    {
648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732
        case UNIFORM:
        {
            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;
        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; 
        }
733 734
    }
    /* New event after MEAN_INTERVAL */  
735
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
736 737 738
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
739
    m_new = (struct nw_message*)tw_event_data(e);
740 741
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
742 743

    free(dest_svr);
744 745 746 747
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
748 749 750
    (void)bf;
    (void)m;
    (void)lp;
751 752 753 754 755 756 757
//    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)
{
758 759 760
    (void)bf;
    (void)lp;

761 762 763 764 765 766 767 768 769 770 771 772
    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)
	{ 
        	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;
	}
     }
773 774 775 776
    int data = m->fwd.num_bytes;
    s->syn_data += data;
    num_syn_bytes_recvd += data;
}
777
/* Debugging functions, may generate unused function warning */
778
/*static void print_waiting_reqs(uint32_t * reqs, int count)
779
{
780
    lprintf("\n Waiting reqs: %d count", count);
781 782
    int i;
    for(i = 0; i < count; i++ )
783
        lprintf(" %d ", reqs[i]);
784
}*/
785 786 787 788 789 790 791 792 793 794 795 796
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);
797
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
798 799
       }
}
Misbah Mubarak's avatar
Misbah Mubarak committed
800
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
801
{
802
//    printf("\n Completed queue: ");
803 804
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
805
      tw_output(lp, "\n");
806 807 808
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
809
            tw_output(lp, " %llu ", current->req_id);
810
       }
Misbah Mubarak's avatar
Misbah Mubarak committed
811
}
812
static int clear_completed_reqs(nw_state * s,
813
        tw_lp * lp,
Misbah Mubarak's avatar
Misbah Mubarak committed
814
        unsigned int * reqs, int count)
815
{
816 817 818
    (void)s;
    (void)lp;

819
    int i, matched = 0;
820

821 822 823
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
824 825 826
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

827
      int index = 0;
828 829
      qlist_for_each(ent, &s->completed_reqs)
       {
830 831
           if(prev)
           {
832
              rc_stack_push(lp, prev, free, s->matched_reqs);
833 834
              prev = NULL;
           }
835
            
836 837
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
838 839
            if(current->req_id == reqs[i])
            {
840
                ++matched;
841
                qlist_del(&current->ql);
842
                prev = current;
843
            }
844
            ++index;
845
       }
846 847

      if(prev)
848 849 850 851
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
852
    }
853
    return matched;
854
}
855
static void add_completed_reqs(nw_state * s,
856 857
        tw_lp * lp,
        int count)
858
{
859
    (void)lp;
860
    for(int i = 0; i < count; i++)
861
    {
862 863
       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.
864
       qlist_add(&req->ql, &s->completed_reqs);
865
    }//end for
866
}
867

868 869 870 871 872 873
/* 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);
}

874
static int notify_posted_wait(nw_state* s,
875
        tw_bf * bf, nw_message * m, tw_lp * lp,
Misbah Mubarak's avatar
Misbah Mubarak committed
876
        unsigned int completed_req)
877
{
878 879
    (void)bf;

880 881
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
882

883
    m->fwd.wait_completed = 0;
884

885 886
    if(!wait_elem)
        return 0;
887

888
    int op_type = wait_elem->op_type;
889

890 891 892
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
893
            m->fwd.wait_completed = 1;
894 895
            wait_completed = 1;
    }
896 897
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
898 899 900 901 902 903
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
904
            {
905
                wait_elem->num_completed++;
906
                if(wait_elem->num_completed > wait_elem->count)
907
                    printf("\n Num completed %d count %d LP %llu ",
908 909
                            wait_elem->num_completed,
                            wait_elem->count,
910
                            LLU(lp->gid));
911 912
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
913

914
                if(wait_elem->num_completed >= wait_elem->count)
915
                {
916
                    if(enable_debug)
917
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
918
                    wait_completed = 1;
919
                }
920

921
                m->fwd.wait_completed = 1;
922
            }
923
        }
924
    }
925
    return wait_completed;
926
}
927

928
/* reverse handler of MPI wait operation */
929
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
930
{
931
   if(bf->c1)
932
    {
933
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
934 935
        if(m->fwd.found_match == 0)
        {
936
            qlist_add(&qi->ql, &s->completed_reqs);
937
        }
938 939 940 941 942 943 944 945 946 947 948 949 950
        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++;
           }
951
        }
952
        //get_next_mpi_operation_rc(s, bf, lp, m);
953
        codes_issue_next_event_rc(lp);
954
        return;
955
    }
956 957 958
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
959
}
960

961
/* execute MPI wait operation */
962
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
963
{
964
    /* check in the completed receives queue if the request ID has already been completed.*/
Misbah Mubarak's avatar
Misbah Mubarak committed
965 966
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
967
    assert(!s->wait_op);
Misbah Mubarak's avatar
Misbah Mubarak committed
968
    unsigned int req_id = mpi_op->u.wait.req_id;
969

970
    struct completed_requests* current = NULL;
971

972
    struct qlist_head * ent = NULL;
973
    int index = 0;
974 975 976 977 978
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
979
            bf->c1=1;
980
            qlist_del(&current->ql);
981
            rc_stack_push(lp, current, free, s->processed_ops);
982
            codes_issue_next_event(lp);
983
            m->fwd.found_match = index;
Misbah Mubarak's avatar
Misbah Mubarak committed
984
            if(s->nw_id == (tw_lpid)TRACK_LP)
985 986 987
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
Misbah Mubarak's avatar
Misbah Mubarak committed
988
            }
989
            //get_next_mpi_operation(s, bf, m, lp);
990 991
            return;
        }
992
        ++index;
993
    }
994

995 996 997 998 999
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
1000
    /* If not, add the wait operation in the pending 'waits' list. */
1001
    struct pending_waits* wait_op = (struct pending_waits*)calloc(1, sizeof(struct pending_waits));
1002 1003 1004
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
1005 1006
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
1007
    s->wait_op = wait_op;
1008

1009
    return;
1010 1011
}

1012
static void codes_exec_mpi_wait_all_rc(
1013
        nw_state* s,
1014 1015
        tw_bf * bf,
        nw_message * m,
1016
        tw_lp* lp)
1017
{
1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028
  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--;
    }
  }
1029 1030 1031 1032 1033 1034 1035 1036
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1037
      add_completed_reqs(s, lp, m->fwd.num_matched);
1038
      codes_issue_next_event_rc(lp);
1039
      //get_next_mpi_operation_rc(s, bf, lp, m);
1040 1041
  }
  return;
1042
}
1043

1044
static void codes_exec_mpi_wait_all(
1045
        nw_state* s,
1046 1047
        tw_bf * bf,
        nw_message * m,
1048
        tw_lp* lp,
1049
        struct codes_workload_op * mpi_op)
1050
{
1051
  if(enable_debug)
1052
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1053

1054 1055 1056 1057 1058 1059 1060 1061
  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;
1062
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1063 1064 1065 1066 1067 1068
        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)
    {
1069
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1070 1071 1072 1073 1074 1075 1076 1077
        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++;
  }
1078
  int count = mpi_op->u.waits.count;
1079 1080
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1081

1082
  int i = 0, num_matched = 0;
1083
  m->fwd.num_matched = 0;
1084

1085
  /*if(lp->gid == TRACK_LP)
1086
  {
1087
      printf("\n MPI Wait all posted ");
1088
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1089 1090
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1091
      /* check number of completed irecvs in the completion queue */
1092 1093
  for(i = 0; i < count; i++)
  {
Misbah Mubarak's avatar
Misbah Mubarak committed
1094
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1095 1096 1097 1098
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1099
            current = qlist_entry(ent, struct completed_requests, ql);
1100 1101 1102 1103
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1104

1105
  m->fwd.found_match = num_matched;
1106 1107 1108 1109
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1110 1111 1112
      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);
1113 1114
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1115
      //get_next_mpi_operation(s, bf, lp, m);
1116 1117
  }
  else
1118 1119
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1120
	  struct pending_waits* wait_op = (struct pending_waits*)calloc(1, sizeof(struct pending_waits));
1121 1122 1123 1124 1125 1126 1127 1128
	  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;
1129
	  wait_op->start_time = tw_now(lp);
1130
      s->wait_op = wait_op;
1131
  }
1132 1133
  return;
}