model-net-mpi-replay.c 97.6 KB
Newer Older
1
/*
2
* Copyright (C) 2014 University of Chicago.
3 4 5 6 7
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
8
#include <sys/stat.h>
mubarak's avatar
mubarak committed
9
#include <sys/resource.h>
10 11 12 13 14 15 16
#include "codes/codes-workload.h"
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
#include "codes/rc-stack.h"
#include "codes/quicklist.h"
17
#include "codes/quickhash.h"
18
#include "codes/codes-jobmap.h"
19

20
/* turning on track lp will generate a lot of output messages */
21
#define MN_LP_NM "modelnet_dragonfly_custom"
22
#define CONTROL_MSG_SZ 64
23
#define TRACE -1
24
#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

mubarak's avatar
mubarak committed
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;
Misbah Mubarak's avatar
Misbah Mubarak committed
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

135 136 137
/* set group context */
struct codes_mctx group_ratio;

138
/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
139
* 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.
140 141 142 143 144 145 146
* 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,
147 148 149 150 151 152
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
153 154
};

155 156 157 158 159 160
/* 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 */
161 162
    STENCIL = 4, /* sends message to 4 nearby neighbors */
    PERMUTATION = 5
163
};
164 165 166 167
struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
168
    int app_id;
169 170 171 172 173
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
174 175 176 177 178 179 180
/* 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;
181
    int64_t num_bytes;
182
    int64_t seq_id;
183 184 185 186 187 188 189 190
    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
191
	unsigned int req_id;
192
    struct qlist_head ql;
193
    int index;
194 195 196 197 198 199
};

/* 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
200
    unsigned int req_ids[MAX_WAIT_REQS];
201
	int num_completed;
202 203
	int count;
    tw_stime start_time;
204 205 206
    struct qlist_head ql;
};

207 208 209 210 211 212
struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
213
    struct qhash_head  hash_link;
214 215
    struct qlist_head ql; 
};
216 217 218 219 220 221 222 223 224 225 226 227 228 229

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;
230 231 232
    double comm_time;
    double max_time;
    double avg_msg_time;
233 234
};

235 236 237 238 239 240 241 242 243 244
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;
245 246
    int app_id;
    int local_rank;
247

248
    int synthetic_pattern;
249 250 251
    int is_finished;
    int neighbor_completed;

252
    struct rc_stack * processed_ops;
253
    struct rc_stack * processed_wait_op;
254
    struct rc_stack * matched_reqs;
255
//    struct rc_stack * indices;
256 257 258 259 260 261 262 263 264 265

    /* 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;

266

267 268
	/* time spent by the LP in executing the app trace*/
	double start_time;
269 270 271 272 273 274 275 276 277

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

278 279 280 281 282
	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
283 284
    /* max time for synthetic traffic message */
    double max_time;
285 286 287 288 289 290 291 292 293 294
	/* 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;
295

296
    tw_stime cur_interval_end;
297
    
298 299
    /* Pending wait operation */
    struct pending_waits * wait_op;
300

301 302 303 304 305 306
    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

307 308
    unsigned long long num_bytes_sent;
    unsigned long long num_bytes_recvd;
309

310 311
    unsigned long long syn_data;
    unsigned long long gen_data;
312 313
  
    unsigned long prev_switch;
314
    int saved_perm_dest;
315 316
    unsigned long rc_perm;

317 318 319 320
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
321
    char output_buf[512];
322
    char col_stats[64];
323
    struct ross_model_sample ross_sample;
324 325 326 327
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
328
* 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.
329 330 331
* 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
{
332
   // forward message handler
333
   int msg_type;
334
   int op_type;
335
   model_net_event_return event_rc;
336
   struct codes_workload_op * mpi_op;
337

338 339 340
   struct
   {
       tw_lpid src_rank;
341
       int dest_rank;
342
       int64_t num_bytes;
343 344 345 346 347
       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
348
       unsigned int req_id;
349
       int matched_req;
350
       int tag;
351
       int app_id;
352 353
       int found_match;
       short wait_completed;
354
       short rend_send;
355 356 357
   } fwd;
   struct
   {
358
       int saved_perm;
359
       double saved_send_time;
360
       double saved_send_time_sample;
361
       double saved_recv_time;
362
       double saved_recv_time_sample;
363
       double saved_wait_time;
364
       double saved_wait_time_sample;
365
       double saved_delay;
366
       double saved_delay_sample;
367
       int64_t saved_num_bytes;
368
       int saved_syn_length;
369
       unsigned long saved_prev_switch;
370
       double saved_prev_max_time;
371
   } rc;
372 373
};

374
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);
375 376

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
377 378
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
379
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
380 381
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
382
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
383 384
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
385
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
386 387
/* execute the computational delay */
static void codes_exec_comp_delay(
388
        nw_state* s, tw_bf *bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
389 390 391 392 393 394 395 396 397 398 399 400 401 402 403
/* 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 */
404 405 406 407 408 409 410 411
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);
412 413 414 415 416 417 418 419 420 421 422 423 424 425 426
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);

427
/*static void update_message_size_rc(
428 429 430 431
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
432
{*/
433
/*TODO: Complete reverse handler */
434
/*    (void)ns;
435 436 437
    (void)lp;
    (void)bf;
    (void)m;
438
}*/
439 440 441 442 443 444 445 446 447 448
/* 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)
{
449 450 451
            (void)bf;
            (void)is_eager;

452 453 454
            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
455
            if(ns->msg_sz_table == NULL)
456 457 458 459 460 461 462 463 464 465
                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)
            {
466
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
467 468
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
469
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
470
                msg_info->avg_latency = msg_info->agg_latency;
471 472
                assert(ns->msg_sz_table);
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), &(msg_info->hash_link));
473 474 475 476 477 478 479
                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++;
480
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
481
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
482
//                printf("\n Msg size %lld aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
483 484 485 486 487 488 489 490
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
491 492 493
    (void)ns;
    (void)bf;
    (void)m;
494 495 496 497 498
        
    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); 
499 500 501 502 503 504 505 506
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
507 508 509
        (void)bf;
        (void)m;

510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
        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);

525
            lprintf("\n Other ranks %d ", num_other_ranks);
526
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
527 528 529 530 531 532 533 534 535 536 537
            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);
538
                m_new = (struct nw_message*)tw_event_data(e);
539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571
                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)
    {
572
//        printf("\n All workloads completed, notifying background traffic ");
573 574 575 576 577 578 579 580 581 582 583 584 585
        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;

586
//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
587
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, noise);
588 589 590 591 592 593 594 595 596
        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);
597
        m_new = (struct nw_message*)tw_event_data(e); 
598 599 600 601 602 603 604 605 606 607
        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)
{
608 609 610 611
        (void)b;
        (void)msg;
        (void)lp;

612 613 614 615 616 617 618 619 620
        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
621 622
        (void)b;
        (void)msg;
623
        ns->is_finished = 1;
624
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
625
        
626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654
        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;

655 656 657 658 659 660
    if(bf->c1)
    {
        tw_rand_reverse_unif(lp->rng);
    }
    if(bf->c2)
    {
661
        s->prev_switch = m->rc.saved_prev_switch;
662
        s->saved_perm_dest = m->rc.saved_perm;
663 664
        tw_rand_reverse_unif(lp->rng);
    }
665 666 667 668 669
    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;
670
        s->num_bytes_sent -= payload_sz;
671
        s->ross_sample.num_bytes_sent -= payload_sz;
672 673
    }
        tw_rand_reverse_unif(lp->rng);
674
        s->num_sends--;
675
        s->ross_sample.num_sends--;
676

677 678
     if(bf->c5)
         s->is_finished = 0;
679 680 681 682 683 684 685 686 687 688 689 690 691
}

/* 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;
692 693
    int intm_dest_id;
    nw_message remote_m;
694 695 696 697 698 699

    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);

700 701 702 703
    /* Find destination */
    int* dest_svr = NULL; 
    int i, length=0;
    switch(s->synthetic_pattern)
704
    {
705 706
        case UNIFORM:
        {
707
            bf->c1 = 1;
708 709 710 711 712 713 714
            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;
715 716 717

        case PERMUTATION:
        {
718 719
            m->rc.saved_prev_switch = s->prev_switch; //for reverse computation

720 721
            length = 1;
            dest_svr = (int*) calloc(1, sizeof(int));
722
            if(s->gen_data - s->prev_switch >= perm_switch_thresh)
723
            {
724
                // printf("%d - %d >= %d\n",s->gen_data,s->prev_switch,perm_switch_thresh);
725
                bf->c2 = 1;
726
                m->rc.saved_prev_switch = s->prev_switch;
727
                s->prev_switch = s->gen_data; //Amount of data pushed at time when switch initiated
728 729 730
                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;
731 732
                /* TODO: Fix random number generation code */
                m->rc.saved_perm = s->saved_perm_dest;
733
                s->saved_perm_dest = dest_svr[0];
734
                assert(s->saved_perm_dest != s->local_rank);
735 736 737
            }
            else
                dest_svr[0] = s->saved_perm_dest;
738 739

            assert(dest_svr[0] != s->local_rank);
740 741
        }
        break;
742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790
        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;
791

792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810
    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);
811
            m->event_rc = model_net_event(net_id, "medium", global_dest_id, payload_sz, 0.0, 
812 813 814 815
                    sizeof(nw_message), (const void*)&remote_m, 
                    0, NULL, lp);
            
            s->gen_data += payload_sz;
816
            s->num_bytes_sent += payload_sz;
817
            s->ross_sample.num_bytes_sent += payload_sz;
818 819
            num_syn_bytes_sent += payload_sz; 
        }
820
    }
821
    s->num_sends++;
822
    s->ross_sample.num_sends++;
823

824
    /* New event after MEAN_INTERVAL */  
825
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
826 827 828
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
829
    m_new = (struct nw_message*)tw_event_data(e);
830 831
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
832
    
833 834 835 836 837
    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;
        }
838
    }
839

840
    free(dest_svr);
841 842 843 844
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
845 846 847
    (void)bf;
    (void)m;
    (void)lp;
848
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
849
    s->num_recvs--;
850
    s->ross_sample.num_recvs--;
851 852 853
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
854
    s->num_bytes_recvd -= data;
855
    s->ross_sample.num_bytes_recvd -= data;
856
    s->send_time = m->rc.saved_send_time;
857
    s->ross_sample.send_time = m->rc.saved_send_time_sample;
858 859 860 861 862
    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;
    }
863 864 865
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
866 867 868
    (void)bf;
    (void)lp;

869
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
870 871 872 873
    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)
mubarak's avatar
mubarak committed
874 875 876 877
    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)
878 879 880 881 882 883 884
	{ 
        	struct rusage mem_usage;
		int who = RUSAGE_SELF;
		int err = getrusage(who, &mem_usage);
		printf("\n Memory usage %lf gigabytes", ((double)mem_usage.ru_maxrss / (1024.0 * 1024.0)));
		upper_threshold += 1048576;
	}*/
mubarak's avatar
mubarak committed
885
	}
886
    m->rc.saved_send_time = s->send_time;
887
    m->rc.saved_send_time_sample = s->ross_sample.send_time;
888
    if((tw_now(lp) - m->fwd.sim_start_time) > s->max_time)
889 890
    {
        m->rc.saved_prev_max_time = s->max_time;
891
        s->max_time = tw_now(lp) - m->fwd.sim_start_time;
892 893
        s->ross_sample.max_time = tw_now(lp) - m->fwd.sim_start_time;
    }
894 895

    s->send_time += (tw_now(lp) - m->fwd.sim_start_time);
896
    s->ross_sample.send_time += (tw_now(lp) - m->fwd.sim_start_time);
897
    s->num_recvs++;
898
    s->ross_sample.num_recvs++;
899 900
    int data = m->fwd.num_bytes;
    s->syn_data += data;
901
    s->num_bytes_recvd += data;
902
    s->ross_sample.num_bytes_recvd += data;
903 904
    num_syn_bytes_recvd += data;
}
905
/* Debugging functions, may generate unused function warning */
906
/*static void print_waiting_reqs(uint32_t * reqs, int count)
907
{
908
    lprintf("\n Waiting reqs: %d count", count);
909 910
    int i;
    for(i = 0; i < count; i++ )
911
        lprintf(" %d ", reqs[i]);
912
}*/
913 914 915 916 917 918 919 920 921 922 923 924
static void print_msgs_queue(struct qlist_head * head, int is_send)
{
    if(is_send)
        printf("\n Send msgs queue: ");
    else
        printf("\n Recv msgs queue: ");

    struct qlist_head * ent = NULL;
    mpi_msgs_queue * current = NULL;
    qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, mpi_msgs_queue, ql);
mubarak's avatar
mubarak committed
925
            //printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
926 927
       }
}
Misbah Mubarak's avatar
Misbah Mubarak committed
928
static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
929
{
930
//    printf("\n Completed queue: ");
931 932
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
933
      tw_output(lp, "\n");
934 935 936
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
937
            tw_output(lp, " %llu ", current->req_id);
938
       }
Misbah Mubarak's avatar
Misbah Mubarak committed
939
}
940
static int clear_completed_reqs(nw_state * s,
941
        tw_lp * lp,
Misbah Mubarak's avatar
Misbah Mubarak committed
942
        unsigned int * reqs, int count)
943
{
944 945 946
    (void)s;
    (void)lp;

947
    int i, matched = 0;
948

949 950 951
    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
952 953 954
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

955
      int index = 0;
956 957
      qlist_for_each(ent, &s->