model-net-mpi-replay.c 98.5 KB
Newer Older
1
/*
2
* Copyright (C) 2014 University of Chicago.
3 4 5 6 7
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
8
#include <sys/stat.h>
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 27
#define RANK_HASH_TABLE_SZ 2000
#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
#define COL_TAG 1235
#define BAR_TAG 1234
33

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

37
static unsigned long perm_switch_thresh = 8388608;
38

39
/* NOTE: Message tracking works in sequential mode only! */
40
static int debug_cols = 0;
41
static int synthetic_pattern = 1;
42 43 44 45 46
/* 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;
static int enable_msg_tracking = 0;
static int is_synthetic = 0;
47
static unsigned long long max_gen_data = 1310720;
48
static int num_qos_levels;
49
tw_lpid TRACK_LP = -1;
50
int nprocs = 0;
51
static double total_syn_data = 0;
52
static int unmatched = 0;
53
char workload_type[128];
54
char workload_name[128];
55 56 57 58
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
59
static int priority_type = 0;
60
static int num_dumpi_traces = 0;
61
static int64_t EAGER_THRESHOLD = 8192;
62

63
static long num_ops = 0;
64
static int upper_threshold = 1048576;
65
static int alloc_spec = 0;
66 67
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;
68
static int payload_sz = 1024;
69 70

/* Doing LP IO*/
71
static char * params = NULL;
72
static char lp_io_dir[256] = {'\0'};
73
static char sampling_dir[32] = {'\0'};
74
static char mpi_msg_dir[32] = {'\0'};
75 76 77 78
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

79 80 81 82
/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
83 84 85
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
86 87 88 89 90
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

91 92
/* Variables for Cortex Support */
/* Matthieu's additions start */
93
#ifdef ENABLE_CORTEX_PYTHON
94 95 96
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
97
#endif
98 99
/* Matthieu's additions end */

100 101
typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
102
typedef unsigned int dumpi_req_id;
103 104

static int net_id = 0;
105
static float noise = 1.0;
106 107 108
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;
109
static int syn_type = 0;
110

111
FILE * workload_log = NULL;
112
FILE * msg_size_log = NULL;
113 114 115 116
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;
117

118 119
unsigned long long num_bytes_sent=0;
unsigned long long num_bytes_recvd=0;
120

121 122
unsigned long long num_syn_bytes_sent = 0;
unsigned long long num_syn_bytes_recvd = 0;
123

124 125 126 127 128 129
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;
130 131 132
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
133
static int enable_debug = 0;
134

135
/* set group context */
136 137 138 139 140 141 142 143 144 145 146
struct codes_mctx mapping_context;
enum MAPPING_CONTEXTS
{
    GROUP_RATIO=1,
    GROUP_RATIO_REVERSE,
    GROUP_DIRECT,
    GROUP_MODULO,
    GROUP_MODULO_REVERSE,
    UNKNOWN
};
static int map_ctxt = GROUP_MODULO;
147

148
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
149
* 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.
150 151 152 153 154 155 156
* 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,
157 158 159 160 161 162
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
163 164
};

165 166 167 168 169 170
/* 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 */
171 172
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
173
};
174 175 176 177
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
178
    int app_id;
179 180 181 182 183
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
184 185 186 187 188 189 190
/* 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;
191
    int64_t num_bytes;
192
    int64_t seq_id;
193 194 195 196 197 198 199 200
    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
{
201
	unsigned int req_id;
202
    struct qlist_head ql;
203
    int index;
204 205 206 207 208 209
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
210
    unsigned int req_ids[MAX_WAIT_REQS];
211
	int num_completed;
212 213
	int count;
    tw_stime start_time;
214 215 216
    struct qlist_head ql;
};

217 218 219 220 221 222
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
223
    struct qhash_head  hash_link;
224 225
    struct qlist_head ql; 
};
226 227 228 229 230 231 232 233 234 235 236 237 238 239

struct ross_model_sample
{
    tw_lpid nw_id;
    int app_id;
    int local_rank;
    unsigned long num_sends;
    unsigned long num_recvs;
    unsigned long long num_bytes_sent;
    unsigned long long num_bytes_recvd;
    double send_time;
    double recv_time;
    double wait_time;
    double compute_time;
240 241 242
    double comm_time;
    double max_time;
    double avg_msg_time;
243 244
};

245 246 247 248 249 250 251 252 253 254
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;
255 256
    int app_id;
    int local_rank;
257

258
    int synthetic_pattern;
259 260 261
    int is_finished;
    int neighbor_completed;

262
    struct rc_stack * processed_ops;
263
    struct rc_stack * processed_wait_op;
264
    struct rc_stack * matched_reqs;
265
//    struct rc_stack * indices;
266 267 268 269 270 271 272 273 274 275

    /* 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;

276

277 278
	/* time spent by the LP in executing the app trace*/
	double start_time;
279 280 281 282 283 284 285 286 287

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

288 289 290 291 292
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
293 294
    /* max time for synthetic traffic message */
    double max_time;
295 296 297 298 299 300 301 302 303 304
	/* 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;
305

306
    tw_stime cur_interval_end;
307
    
308 309
    /* Pending wait operation */
    struct pending_waits * wait_op;
310

311 312 313 314 315 316
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

317 318
    unsigned long long num_bytes_sent;
    unsigned long long num_bytes_recvd;
319

320 321
    unsigned long long syn_data;
    unsigned long long gen_data;
322 323
  
    unsigned long prev_switch;
324
    int saved_perm_dest;
325 326
    unsigned long rc_perm;

327 328 329 330
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
331
    char output_buf[512];
332
    char col_stats[64];
333
    struct ross_model_sample ross_sample;
334 335 336 337
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
338
* 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.
339 340 341
* 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
{
342
   // forward message handler
343
   int msg_type;
344
   int op_type;
345
   model_net_event_return event_rc;
346
   struct codes_workload_op * mpi_op;
347

348 349 350
   struct
   {
       tw_lpid src_rank;
351
       int dest_rank;
352
       int64_t num_bytes;
353 354 355 356 357
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
358
       unsigned int req_id;
359
       int matched_req;
360
       int tag;
361
       int app_id;
362 363
       int found_match;
       short wait_completed;
364
       short rend_send;
365 366 367
   } fwd;
   struct
   {
368
       int saved_perm;
369
       double saved_send_time;
370
       double saved_send_time_sample;
371
       double saved_recv_time;
372
       double saved_recv_time_sample;
373
       double saved_wait_time;
374
       double saved_wait_time_sample;
375
       double saved_delay;
376
       double saved_delay_sample;
377
       int64_t saved_num_bytes;
378
       int saved_syn_length;
379
       unsigned long saved_prev_switch;
380
       double saved_prev_max_time;
381
   } rc;
382 383
};

384
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);
385 386

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
387 388
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
389
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
390 391
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
392
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
393 394
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
395
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
396 397
/* execute the computational delay */
static void codes_exec_comp_delay(
398
        nw_state* s, tw_bf *bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413
/* 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 */
414 415 416 417 418 419 420 421
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);
422 423 424 425 426 427 428 429 430 431 432 433 434 435 436
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);

437
/*static void update_message_size_rc(
438 439 440 441
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
442
{*/
443
/*TODO: Complete reverse handler */
444
/*    (void)ns;
445 446 447
    (void)lp;
    (void)bf;
    (void)m;
448
}*/
449 450 451 452 453 454 455 456 457 458
/* 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)
{
459 460 461
            (void)bf;
            (void)is_eager;

462 463 464
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
465
            if(ns->msg_sz_table == NULL)
466 467 468 469 470 471 472 473 474 475
                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)
            {
476
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
477 478
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
479
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
480
                msg_info->avg_latency = msg_info->agg_latency;
481 482
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
483 484 485 486 487 488 489
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
490
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
491
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
492
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
493 494 495 496 497 498 499 500
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
501 502 503
    (void)ns;
    (void)bf;
    (void)m;
504 505 506 507 508
        
    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); 
509 510 511 512 513 514 515 516
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
517 518 519
        (void)bf;
        (void)m;

520 521 522 523
        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); 
524 525
        
        for(int other_id = 0; other_id < num_jobs; other_id++)
526
        {
527 528
            if(other_id == jid.job)
                continue;
529 530 531 532 533 534

            struct codes_jobmap_id other_jid;
            other_jid.job = other_id;

            int num_other_ranks = codes_jobmap_get_num_ranks(other_id, jobmap_ctx);

535
            lprintf("\n Other ranks %d ", num_other_ranks);
536
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
537 538 539 540 541 542 543 544 545 546 547
            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);
548
                m_new = (struct nw_message*)tw_event_data(e);
549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577
                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)
{
578
    if(ns->local_rank == num_dumpi_traces - 1 
579 580 581
            && ns->is_finished == 1
            && ns->neighbor_completed == 1)
    {
582 583 584
//        printf("\n All workloads completed, notifying background traffic ");
        bf->c0 = 1;
        notify_background_traffic(ns, lp, bf, m);
585 586 587 588 589 590 591 592 593 594 595
        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;

596
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
597
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
598 599 600 601 602 603 604 605 606
        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);
607
        m_new = (struct nw_message*)tw_event_data(e); 
608 609 610 611 612 613 614 615 616 617
        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)
{
618 619 620 621
        (void)b;
        (void)msg;
        (void)lp;

622 623 624 625 626 627 628 629 630
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
631 632
        (void)b;
        (void)msg;
633
        ns->is_finished = 1;
634
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
635
        
636 637 638 639 640 641 642 643 644 645
        return;
}

void finish_nbr_wkld_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 0;
646
    
647 648 649 650 651 652 653 654 655 656
    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;
657

658 659 660 661 662 663 664
    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;

665 666 667 668 669 670
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
671
        s->prev_switch = m->rc.saved_prev_switch;
672
        s->saved_perm_dest = m->rc.saved_perm;
673 674
        tw_rand_reverse_unif(lp->rng);
    }
675 676 677 678 679
    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;
680
        s->num_bytes_sent -= payload_sz;
681
        s->ross_sample.num_bytes_sent -= payload_sz;
682 683
    }
        tw_rand_reverse_unif(lp->rng);
684
        s->num_sends--;
685
        s->ross_sample.num_sends--;
686

687 688
     if(bf->c5)
         s->is_finished = 0;
689 690 691 692 693 694 695 696 697 698 699 700 701
}

/* 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;
702 703
    int intm_dest_id;
    nw_message remote_m;
704 705 706 707 708 709

    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);

710 711 712 713
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
714
    {
715 716
        case UNIFORM:
        {
717
            bf->c1 = 1;
718 719 720 721 722 723 724
            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;
725 726 727

        case PERMUTATION:
        {
728 729
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

730 731
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
732
            if(s->gen_data - s->prev_switch >= perm_switch_thresh)
733
            {
734
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
735
                bf->c2 = 1;
736
                m->rc.saved_prev_switch = s->prev_switch;
737
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
738 739 740
                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;
741 742
                /* TODO: Fix random number generation code */
                m->rc.saved_perm = s->saved_perm_dest;
743
                s->saved_perm_dest = dest_svr[0];
744
                assert(s->saved_perm_dest != s->local_rank);
745 746 747
            }
            else
                dest_svr[0] = s->saved_perm_dest;
748 749

            assert(dest_svr[0] != s->local_rank);
750 751
        }
        break;
752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800
        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;
801

802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820
    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);
821
            m->event_rc = model_net_event(net_id, "medium", global_dest_id, payload_sz, 0.0, 
822 823 824 825
                    sizeof(nw_message), (const void*)&remote_m, 
                    0, NULL, lp);
            
            s->gen_data += payload_sz;
826
            s->num_bytes_sent += payload_sz;
827
            s->ross_sample.num_bytes_sent += payload_sz;
828 829
            num_syn_bytes_sent += payload_sz; 
        }
830
    }
831
    s->num_sends++;
832
    s->ross_sample.num_sends++;
833

834
    /* New event after MEAN_INTERVAL */  
835
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
836 837 838
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
839
    m_new = (struct nw_message*)tw_event_data(e);
840 841
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
842
    
843 844 845 846 847
    if(num_qos_levels > 1) { //instead of using notify neighbor with QoS, ranks constantly check to see if they have exceeded the hard limit on data - if they have, then they're finished.
        if(s->gen_data >= max_gen_data) {
            bf->c5 = 1;
            s->is_finished = 1;
        }
848
    }
849 850

    free(dest_svr);
851 852 853 854
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
855 856 857
    (void)bf;
    (void)m;
    (void)lp;
858
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
859
    s->num_recvs--;
860
    s->ross_sample.num_recvs--;
861 862 863
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
864
    s->num_bytes_recvd -= data;
865
    s->ross_sample.num_bytes_recvd -= data;
866
    s->send_time = m->rc.saved_send_time;
867
    s->ross_sample.send_time = m->rc.saved_send_time_sample;
868 869 870 871 872
    if((tw_now(lp) - m->fwd.sim_start_time) > s->max_time)
    {
        s->max_time = m->rc.saved_prev_max_time;
        s->ross_sample.max_time = m->rc.saved_prev_max_time;
    }
873 874 875
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
876 877 878
    (void)bf;
    (void)lp;

879
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
880 881 882 883
    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)
884 885 886 887
    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)
888 889 890 891 892 893 894
	{ 
        	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;
	}*/
895
	}
896
    m->rc.saved_send_time = s->send_time;
897
    m->rc.saved_send_time_sample = s->ross_sample.send_time;
898
    if((tw_now(lp) - m->fwd.sim_start_time) > s->max_time)
899 900
    {
        m->rc.saved_prev_max_time = s->max_time;
901
        s->max_time = tw_now(lp) - m->fwd.sim_start_time;
902 903
        s->ross_sample.max_time = tw_now(lp) - m->fwd.sim_start_time;
    }
904 905

    s->send_time += (tw_now(lp) - m->fwd.sim_start_time);
906
    s->ross_sample.send_time += (tw_now(lp) - m->fwd.sim_start_time);
907
    s->num_recvs++;
908
    s->ross_sample.num_recvs++;
909 910
    int data = m->fwd.num_bytes;
    s->syn_data += data;
911
    s->num_bytes_recvd += data;
912
    s->ross_sample.num_bytes_recvd += data;
913 914
    num_syn_bytes_recvd += data;
}
915
/* Debugging functions, may generate unused function warning */
916
/*static void print_waiting_reqs(uint32_t * reqs, int count)
917
{
918
    lprintf("\n Waiting reqs: %d count", count);
919 920
    int i;
    for(i = 0; i < count; i++ )
921
        lprintf(" %d ", reqs[i]);
922
}*/
923 924 925 926 927 928 929 930 931 932 933 934
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);
935
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
936 937
       }
}
938
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
939
{
940
//    printf("\n Completed queue: ");
941 942
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
943
      tw_output(lp, "\n");
944 945 946
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
947
            tw_output(lp, " %llu ", current->req_id);
948
       }
949
}
950
static int clear_completed_reqs(nw_state * s,
951
        tw_lp * lp,
952
        unsigned int * reqs, int count)
953
{
954 955 956
    (void)s;
    (void)lp;

957
    int i, matched = 0;
958

959 960 961
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
962 963 964
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

965
      int index = 0;
966 967
      qlist_for_each(ent, &s->completed_reqs)
       {
968 969
           if(prev)
           {
970
              rc_stack_push(lp, prev, free, s->matched_reqs);
971 972
              prev = NULL;
           }
973
            
974 975
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
976 977
            if(current->req_id == reqs[i])
            {
978
                ++matched;
979
                qlist_del(&current->ql);
980
                prev = current;
981
            }
982
            ++index;
983
       }
984 985

      if(prev)
986 987 988 989
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
990
    }
991
    return matched;
992
}
993
static void add_completed_reqs(nw_state * s,
994 995
        tw_lp * lp,
        int count)
996
{
997
    (void)lp;
998
    for(int i = 0; i < count; i++)
999
    {
1000 1001
       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.
1002
       qlist_add(&req->ql, &s->completed_reqs);
1003
    }//end for
1004
}
1005

1006 1007 1008 1009 1010 1011
/* 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);
}

1012
static int notify_posted_wait(nw_state* s,
1013
        tw_bf * bf, nw_message * m, tw_lp * lp,
1014
        unsigned int completed_req)
1015
{
1016 1017
    (void)bf;

1018 1019
    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;
1020

1021
    m->fwd.wait_completed = 0;
1022

1023 1024
    if(!wait_elem)
        return 0;
1025

1026
    int op_type = wait_elem->op_type;
1027

1028 1029 1030
    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
1031
            m->fwd.wait_completed = 1;
1032 1033
            wait_completed = 1;
    }
1034 1035
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
1036 1037 1038 1039 1040 1041
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
1042
            {
1043
                wait_elem->num_completed++;
1044
                if(wait_elem->num_completed > wait_elem->count)
1045
                    printf("\n Num completed %d count %d LP %llu ",
1046 1047
                            wait_elem->num_completed,
                            wait_elem->count,
1048
                            LLU(lp->gid));
1049 1050
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);
1051

1052
                if(wait_elem->num_completed >= wait_elem->count)
1053
                {
1054
                    if(enable_debug)
1055
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
1056
                    wait_completed = 1;
1057
                }
1058

1059
                m->fwd.wait_completed = 1;
1060
            }
1061
        }
1062
    }
1063
    return wait_completed;
1064
}
1065

1066
/* reverse handler of MPI wait operation */
1067
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
1068
{
1069
   if(bf->c1)
1070
    {
1071
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
1072 1073
        if(m->fwd.found_match == 0)
        {
1074
            qlist_add(&qi->ql, &s->completed_reqs);
1075
        }
1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088
        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++;
           }
1089
        }
1090
        codes_issue_next_event_rc(lp);
1091
        return;
1092
    }
1093 1094 1095
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
1096
}
1097

1098
/* execute MPI wait operation */
1099
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
1100
{
1101
    /* check in the completed receives queue if the request ID has already been completed.*/
1102 1103
                
//    printf("\n Wait posted rank id %d ", s->nw_id);
1104
    assert(!s->wait_op);
1105
    unsigned int req_id = mpi_op->u.wait.req_id;
1106

1107
    struct completed_requests* current = NULL;
1108

1109
    struct qlist_head * ent = NULL;
1110
    int index = 0;
1111 1112 1113 1114 1115
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
1116
            bf->c1=1;
1117
            qlist_del(&current->ql);
1118
            rc_stack_push(lp, current, free, s->processed_ops);
1119
            codes_issue_next_event(lp);
1120
            m->fwd.found_match = index;
1121
            if(s->nw_id == (tw_lpid)TRACK_LP)
1122 1123 1124
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
1125
            }
1126 1127
            return;
        }
1128
        ++index;
1129
    }
1130

1131 1132 1133 1134 1135
    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
1136
    /* If not, add the wait operation in the pending 'waits' list. */
1137
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
1138 1139 1140
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
1141 1142
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
1143
    s->wait_op = wait_op;
1144

1145
    return;
1146 1147
}

1148
static void codes_exec_mpi_wait_all_rc(
1149
        nw_state* s,
1150 1151
        tw_bf * bf,
        nw_message * m,
1152
        tw_lp* lp)
1153
{
1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164
  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--;
    }
  }
1165 1166 1167 1168 1169 1170 1171 1172
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
1173
      add_completed_reqs(s, lp, m->fwd.num_matched);
1174 1175 1176
      codes_issue_next_event_rc(lp);
  }
  return;
1177
}
1178

1179
static void codes_exec_mpi_wait_all(
1180
        nw_state* s,
1181 1182
        tw_bf * bf,
        nw_message * m,
1183
        tw_lp* lp,
1184
        struct codes_workload_op * mpi_op)
1185
{
1186
  if(enable_debug)
1187
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));
1188

1189 1190 1191 1192 1193 1194 1195 1196
  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;
1197
        s->mpi_wkld_samples[indx].app_id = s->app_id;
1198 1199 1200 1201 1202 1203
        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)
    {
1204
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
1205 1206 1207 1208 1209 1210 1211 1212
        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++;
  }
1213
  int count = mpi_op->u.waits.count;
1214 1215
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);
1216

1217
  int i = 0, num_matched = 0;
1218
  m->fwd.num_matched = 0;
1219

1220
  /*if(lp->gid == TRACK_LP)
1221
  {
1222
      printf("\n MPI Wait all posted ");
1223
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
1224 1225
      print_completed_queue(lp, &s->completed_reqs);
  }*/
1226
      /* check number of completed irecvs in the completion queue */
1227 1228
  for(i = 0; i < count; i++)
  {
1229
      unsigned int req_id = mpi_op->u.waits.req_ids[i];
1230 1231 1232 1233
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
1234
            current = qlist_entry(ent, struct completed_requests, ql);
1235 1236 1237 1238
            if(current->req_id == req_id)
                num_matched++;
       }
  }