model-net-mpi-replay.c 92.7 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 1024
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
#define COL_TAG 1235
32

33 34 35
static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

36
static unsigned long perm_switch_thresh = 8388608;
37

38
/* NOTE: Message tracking works in sequential mode only! */
39
static int debug_cols = 0;
40
static int synthetic_pattern = 1;
41 42 43
/* 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;
44
static int enable_msg_tracking = 0;
45
static int is_synthetic = 0;
46
static int max_gen_data = 1310720;
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 priority_type = 0;
58
static int num_dumpi_traces = 0;
59
static int64_t EAGER_THRESHOLD = 8192;
60

61
static long num_ops = 0;
62
static int upper_threshold = 1048576;
63
static int alloc_spec = 0;
64
static tw_stime self_overhead = 10.0;
65
static tw_stime mean_interval = 100000;
66
static int payload_sz = 1024;
67 68

/* Doing LP IO*/
69
static char * params = NULL;
70
static char lp_io_dir[256] = {'\0'};
71
static char sampling_dir[32] = {'\0'};
72
static char mpi_msg_dir[32] = {'\0'};
73 74 75 76
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

77 78 79 80
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
81 82 83
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
84 85 86 87 88
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

89 90
/* Variables for Cortex Support */
/* Matthieu's additions start */
91
#ifdef ENABLE_CORTEX_PYTHON
92 93 94
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
95
#endif
96 97
/* Matthieu's additions end */

98 99
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
100
typedef unsigned int dumpi_req_id;
101 102

static int net_id = 0;
103
static float noise = 1.0;
104 105 106
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
107
static int syn_type = 0;
108

109
FILE * workload_log = NULL;
110
FILE * msg_size_log = NULL;
111 112 113 114
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
115

116 117 118
long long num_bytes_sent=0;
long long num_bytes_recvd=0;

119 120 121
long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

122 123 124 125 126 127
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;
128 129 130
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
131
static int enable_debug = 0;
132

133 134 135
/* set group context */
struct codes_mctx group_ratio;

136
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
137
* 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.
138 139 140 141 142 143 144
* 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,
145 146 147 148 149 150
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
151 152
};

153 154 155 156 157 158
/* 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 */
159 160
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
161
};
162 163 164 165
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
166
    int app_id;
167 168 169 170 171
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
172 173 174 175 176 177 178
/* 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;
179
    int64_t num_bytes;
180
    int64_t seq_id;
181 182 183 184 185 186 187 188
    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
{
189
	unsigned int req_id;
190
    struct qlist_head ql;
191
    int index;
192 193 194 195 196 197
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
198
    unsigned int req_ids[MAX_WAIT_REQS];
199
	int num_completed;
200 201
	int count;
    tw_stime start_time;
202 203 204
    struct qlist_head ql;
};

205 206 207 208 209 210
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
211
    struct qhash_head  hash_link;
212 213
    struct qlist_head ql; 
};
214 215 216 217 218 219 220 221 222 223
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;
224 225
    int app_id;
    int local_rank;
226

227
    int synthetic_pattern;
228 229 230
    int is_finished;
    int neighbor_completed;

231
    struct rc_stack * processed_ops;
232
    struct rc_stack * processed_wait_op;
233
    struct rc_stack * matched_reqs;
234
//    struct rc_stack * indices;
235 236 237 238 239 240 241 242 243 244

    /* 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;

245

246 247
	/* time spent by the LP in executing the app trace*/
	double start_time;
248 249 250 251 252 253 254 255 256

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

257 258 259 260 261
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
262 263
    /* max time for synthetic traffic message */
    double max_time;
264 265 266 267 268 269 270 271 272 273
	/* 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;
274

275
    tw_stime cur_interval_end;
276
    
277 278
    /* Pending wait operation */
    struct pending_waits * wait_op;
279

280 281 282 283 284 285
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

286 287 288
    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

289 290
    unsigned long syn_data;
    unsigned long gen_data;
291 292
  
    unsigned long prev_switch;
293
    int saved_perm_dest;
294
    unsigned long rc_perm;
295

296 297 298 299
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
300
    char output_buf[512];
301
    char col_stats[64];
302 303 304 305
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
306
* 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.
307 308 309
* 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
{
310
   // forward message handler
311
   int msg_type;
312
   int op_type;
313
   model_net_event_return event_rc;
314
   struct codes_workload_op * mpi_op;
315

316 317 318
   struct
   {
       tw_lpid src_rank;
319
       int dest_rank;
320
       int64_t num_bytes;
321 322 323 324 325
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
326
       unsigned int req_id;
327
       int matched_req;
328
       int tag;
329
       int app_id;
330 331
       int found_match;
       short wait_completed;
332
       short rend_send;
333 334 335
   } fwd;
   struct
   {
336
       int saved_perm;
337 338 339 340
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
341
       int64_t saved_num_bytes;
342
       int saved_syn_length;
343
       unsigned long saved_prev_switch;
344
   } rc;
345 346
};

347
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);
348 349

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
350 351
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
352
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
353 354
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
355
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
356 357
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
358
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
359 360
/* execute the computational delay */
static void codes_exec_comp_delay(
361
        nw_state* s, tw_bf *bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
362 363 364 365 366 367 368 369 370 371 372 373 374 375 376
/* 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 */
377 378 379 380 381 382 383 384
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);
385 386 387 388 389 390 391 392 393 394 395 396 397 398 399
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);

400
/*static void update_message_size_rc(
401 402 403 404
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
405
{*/
406
/*TODO: Complete reverse handler */
407
/*    (void)ns;
408 409 410
    (void)lp;
    (void)bf;
    (void)m;
411
}*/
412 413 414 415 416 417 418 419 420 421
/* 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)
{
422 423 424
            (void)bf;
            (void)is_eager;

425 426 427
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
428
            if(ns->msg_sz_table == NULL)
429 430 431 432 433 434 435 436 437 438
                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)
            {
439
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
440 441
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
442
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
443
                msg_info->avg_latency = msg_info->agg_latency;
444 445
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
446
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
447
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
448 449 450 451 452
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
453
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
454
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
455
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
456 457 458 459 460 461 462 463
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
464 465 466
    (void)ns;
    (void)bf;
    (void)m;
467 468 469 470 471
        
    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); 
472 473 474 475 476 477 478 479
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
480 481 482
        (void)bf;
        (void)m;

483 484 485 486 487 488 489 490 491 492 493 494 495 496 497
        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);

498
            lprintf("\n Other ranks %d ", num_other_ranks);
499
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
500 501 502 503 504 505 506 507 508 509 510
            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);
511
                m_new = (struct nw_message*)tw_event_data(e);
512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544
                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)
    {
545
//        printf("\n All workloads completed, notifying background traffic ");
546 547 548 549 550 551 552 553 554 555 556 557 558
        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;

559
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
560
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
561 562 563 564 565 566 567 568 569
        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);
570
        m_new = (struct nw_message*)tw_event_data(e); 
571 572 573 574 575 576 577 578 579 580
        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)
{
581 582 583 584
        (void)b;
        (void)msg;
        (void)lp;

585 586 587 588 589 590 591 592 593
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
594 595
        (void)b;
        (void)msg;
596
        ns->is_finished = 1;
597
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
598
        
599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627
        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;

628 629 630 631 632 633
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
634
        s->prev_switch = m->rc.saved_prev_switch;
635
        s->saved_perm_dest = m->rc.saved_perm;
636 637
        tw_rand_reverse_unif(lp->rng);
    }
638 639 640 641 642
    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;
643
        s->num_bytes_sent -= payload_sz;
644 645
    }
        tw_rand_reverse_unif(lp->rng);
646
        s->num_sends--;
647

648 649
     if(bf->c5)
         s->is_finished = 0;
650 651 652 653 654 655 656 657 658 659 660 661 662
}

/* 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;
663 664
    int intm_dest_id;
    nw_message remote_m;
665 666 667 668 669 670

    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);

671 672 673 674
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
675
    {
676 677
        case UNIFORM:
        {
678
            bf->c1 = 1;
679 680 681 682 683 684 685
            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;
686 687 688

        case PERMUTATION:
        {
689 690
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

691 692
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
693
            if(s->gen_data - s->prev_switch >= perm_switch_thresh)
694
            {
695
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
696
                bf->c2 = 1;
697
                m->rc.saved_prev_switch = s->prev_switch;
698
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
699 700
                dest_svr[0] = tw_rand_integer(lp->rng, 0, num_clients - 1);
                if(dest_svr[0] == s->local_rank)
701
                    dest_svr[0] = (s->local_rank + num_clients/2) % num_clients;
702
                /* TODO: Fix random number generation code */
703 704
                m->rc.saved_perm = s->saved_perm_dest;
                s->saved_perm_dest = dest_svr[0];
705
                assert(s->saved_perm_dest != s->local_rank);
706 707
            }
            else
708
                dest_svr[0] = s->saved_perm_dest;
709 710

            assert(dest_svr[0] != s->local_rank);
711 712
        }
        break;
713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781
        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);
782
            m->event_rc = model_net_event(net_id, "medium", global_dest_id, payload_sz, 0.0, 
783 784 785 786
                    sizeof(nw_message), (const void*)&remote_m, 
                    0, NULL, lp);
            
            s->gen_data += payload_sz;
787
            s->num_bytes_sent += payload_sz;
788 789
            num_syn_bytes_sent += payload_sz; 
        }
790
    }
791 792
    s->num_sends++;

793
    /* New event after MEAN_INTERVAL */  
794
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
795 796 797
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
798
    m_new = (struct nw_message*)tw_event_data(e);
799 800
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
801
    
Neil McGlohon's avatar
Neil McGlohon committed
802
    if(s->gen_data >= max_gen_data)
803 804
    {
        bf->c5 = 1;
805
        s->is_finished = 1;
806
    }
807
    free(dest_svr);
808 809 810 811
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
812 813 814
    (void)bf;
    (void)m;
    (void)lp;
815
    s->num_recvs--;
816 817 818
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
819
    s->num_bytes_recvd -= data;
820
    s->send_time = m->rc.saved_send_time;
821 822 823
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
824 825 826
    (void)bf;
    (void)lp;

827 828 829
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    if(s->local_rank == 0)
     {
Neil McGlohon's avatar
Neil McGlohon committed
830
        printf("Data arrived %d total data %ld: time %lf\n", m->fwd.num_bytes, s->syn_data, tw_now(lp));
831
//    	printf("\n Data arrived %lld rank %llu total data %ld ", m->fwd.num_bytes, s->nw_id, s->syn_data);
832
/*	if(s->syn_data > upper_threshold)
833 834 835 836
    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)
837 838 839 840 841 842 843
	{ 
        	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;
	}*/
844
	}
845
    m->rc.saved_send_time = s->send_time;
846
    if((tw_now(lp) - m->fwd.sim_start_time) > s->max_time)
847 848
        s->max_time = tw_now(lp) - m->fwd.sim_start_time;

849 850
    s->send_time += (tw_now(lp) - m->fwd.sim_start_time);
    s->num_recvs++;
851 852
    int data = m->fwd.num_bytes;
    s->syn_data += data;
853
    s->num_bytes_recvd += data;
854 855
    num_syn_bytes_recvd += data;
}
856
/* Debugging functions, may generate unused function warning */
857
/*static void print_waiting_reqs(uint32_t * reqs, int count)
858
{
859
    lprintf("\n Waiting reqs: %d count", count);
860 861
    int i;
    for(i = 0; i < count; i++ )
862
        lprintf(" %d ", reqs[i]);
863
}*/
864 865 866 867 868 869 870 871 872 873 874 875
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);
876
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
877 878
       }
}
879
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
880
{
881
//    printf("\n Completed queue: ");
882 883
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
884
      tw_output(lp, "\n");
885 886 887
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
888
            tw_output(lp, " %llu ", current->req_id);
889
       }
890
}
891
static int clear_completed_reqs(nw_state * s,
892
        tw_lp * lp,
893
        unsigned int * reqs, int count)
894
{
895 896 897
    (void)s;
    (void)lp;

898
    int i, matched = 0;
899

900 901 902
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
903 904 905
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

906
      int index = 0;
907 908
      qlist_for_each(ent, &s->completed_reqs)
       {
909 910
           if(prev)
           {
911
              rc_stack_push(lp, prev, free, s->matched_reqs);
912 913
              prev = NULL;
           }
914
            
915 916
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
917 918
            if(current->req_id == reqs[i])
            {
919
                ++matched;
920
                qlist_del(&current->ql);
921
                prev = current;
922
            }
923
            ++index;
924
       }
925 926

      if(prev)
927 928 929 930
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
931
    }
932
    return matched;
933
}
934
static void add_completed_reqs(nw_state * s,
935 936
        tw_lp * lp,
        int count)
937
{
938
    (void)lp;
939
    for(int i = 0; i < count; i++)
940
    {
941 942
       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.
943
       qlist_add(&req->ql, &s->completed_reqs);
944
    }//end for
945
}
946

947 948 949 950 951 952
/* 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);
}

953
static int notify_posted_wait(nw_state* s,
954
        tw_bf * bf, nw_message * m, tw_lp * lp,
955
        unsigned int completed_req)
956
{
957 958
    (void)bf;

959 960
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
961

962
    m->fwd.wait_completed = 0;
963

964 965
    if(!wait_elem)
        return 0;
966

967
    int op_type = wait_elem->op_type;
968

969 970 971
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
972
            m->fwd.wait_completed = 1;
973 974
            wait_completed = 1;
    }
975 976
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
977 978 979 980 981 982
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
983
            {
984
                wait_elem->num_completed++;
985
                if(wait_elem->num_completed > wait_elem->count)
986
                    printf("\n Num completed %d count %d LP %llu ",
987 988
                            wait_elem->num_completed,
                            wait_elem->count,
989
                            LLU(lp->gid));
990 991
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
992

993
                if(wait_elem->num_completed >= wait_elem->count)
994
                {
995
                    if(enable_debug)
996
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
997
                    wait_completed = 1;
998
                }
999

1000
                m->fwd.wait_completed = 1;
1001
            }
1002
        }
1003
    }
1004
    return wait_completed;
1005
}
1006

1007
/* reverse handler of MPI wait operation */
1008
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
1009
{
1010
   if(bf->c1)
1011
    {
1012
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
1013 1014
        if(m->fwd.found_match == 0)
        {
1015
            qlist_add(&qi->ql, &s->completed_reqs);
1016
        }
1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029
        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++;
           }
1030
        }
1031
        codes_issue_next_event_rc(lp);
1032
        return;
1033
    }
1034 1035 1036
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
1037
}
1038

1039
/* execute MPI wait operation */
1040
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1041
{
1042
    /* check in the completed receives queue if the request ID has already been completed.*/
1043 1044
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
1045
    assert(!s->wait_op);
1046
    unsigned int req_id = mpi_op->u.wait.req_id;
1047

1048
    struct completed_requests* current = NULL;
1049

1050
    struct qlist_head * ent = NULL;
1051
    int index = 0;
1052 1053 1054 1055 1056
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
1057
            bf->c1=1;
1058
            qlist_del(&current->ql);
1059
            rc_stack_push(lp, current, free, s->processed_ops);
1060
            codes_issue_next_event(lp);
1061
            m->fwd.found_match = index;
1062
            if(s->nw_id == (tw_lpid)TRACK_LP)
1063 1064 1065
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
1066
            }
1067 1068
            return;
        }
1069
        ++index;
1070
    }
1071

1072 1073 1074 1075 1076
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
1077
    /* If not, add the wait operation in the pending 'waits' list. */
1078
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1079 1080 1081
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
1082 1083
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
1084
    s->wait_op = wait_op;
1085

1086
    return;
1087 1088
}

1089
static void codes_exec_mpi_wait_all_rc(
1090
        nw_state* s,
1091 1092
        tw_bf * bf,
        nw_message * m,
1093
        tw_lp* lp)
1094
{
1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105
  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--;
    }
  }
1106 1107 1108 1109 1110 1111 1112 1113
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1114
      add_completed_reqs(s, lp, m->fwd.num_matched);
1115 1116 1117
      codes_issue_next_event_rc(lp);
  }
  return;
1118
}
1119

1120
static void codes_exec_mpi_wait_all(
1121
        nw_state* s,
1122 1123
        tw_bf * bf,
        nw_message * m,
1124
        tw_lp* lp,
1125
        struct codes_workload_op * mpi_op)
1126
{
1127
  if(enable_debug)
1128
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1129

1130 1131 1132 1133 1134 1135 1136 1137
  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;
1138
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1139 1140 1141 1142 1143 1144
        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)
    {
1145
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1146 1147 1148 1149 1150 1151 1152 1153
        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++;
  }
1154
  int count = mpi_op->u.waits.count;
1155 1156
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1157

1158
  int i = 0, num_matched = 0;
1159
  m->fwd.num_matched = 0;
1160

1161
  /*if(lp->gid == TRACK_LP)
1162
  {
1163
      printf("\n MPI Wait all posted ");
1164
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1165 1166
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1167
      /* check number of completed irecvs in the completion queue */
1168 1169
  for(i = 0; i < count; i++)
  {
1170
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1171 1172 1173 1174
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1175
            current = qlist_entry(ent, struct completed_requests, ql);
1176 1177 1178 1179
            if(current->req_id == req_id)
                num_matched++;
       }
  }
1180

1181
  m->fwd.found_match = num_matched;
1182 1183 1184 1185
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
1186 1187 1188
      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);
1189 1190
      s->wait_op = NULL;
      codes_issue_next_event(lp);
1191 1192
  }
  else
1193 1194
  {
      /* If not, add the wait operation in the pending 'waits' list. */
1195
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1196 1197 1198 1199 1200 1201 1202 1203
	  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;
1204
	  wait_op->start_time = tw_now(lp);
1205
      s->wait_op = wait_op;
1206
  }
1207 1208
  return;
}
1209

1210 1211
/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
1212
 * Index is used for inserting the element once again in the queue for reverse computation. */
1213
static int rm_matching_rcv(nw_state * ns,
1214
        tw_bf * bf,
1215 1216
        nw_message * m,
        tw_lp * lp,
1217
        mpi_msgs_queue * qitem)
1218 1219
{
    int matched = 0;
1220
    int index = 0;
1221
    int is_rend = 0;
1222 1223
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;
1224

1225 1226
    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
1227 1228
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
1229
               ((qi->tag == qitem->tag) || qi->tag == -1)
1230
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
1231 1232
        {
            matched = 1;
1233
            qi->num_bytes = qitem->num_bytes;
1234 1235
            break;
        }
1236
        ++index;