codes-online-comm-wrkld.C 29 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
/*
 * Copyright (C) 2014 University of Chicago
 * See COPYRIGHT notice in top-level directory.
 *
 */

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <mpi.h>
#include <ross.h>
#include <assert.h>
#include <deque>
14
#include <iostream>
15
#include <inttypes.h>
16
#include <fstream>
17 18
#include <boost/property_tree/json_parser.hpp>
#include <boost/property_tree/ptree.hpp>
19 20 21 22 23 24
#include "codes/codes-workload.h"
#include "codes/quickhash.h"
#include "codes/codes-jobmap.h"
#include "codes_config.h"
#include "lammps.h"
#include "nekbone_swm_user_code.h"
25
#include "nearest_neighbor_swm_user_code.h"
26
// #include "all_to_one_swm_user_code.h"
27

28 29
#define ALLREDUCE_SHORT_MSG_SIZE 2048

30
//#define DBG_COMM 0
31 32 33 34 35 36 37

using namespace std;

static struct qhash_table *rank_tbl = NULL;
static int rank_tbl_pop = 0;
static int total_rank_cnt = 0;
ABT_thread global_prod_thread = NULL;
38
ABT_xstream self_es;
39 40 41 42 43 44 45 46 47 48
long cpu_freq = 1.0;
long num_allreduce = 0;
long num_isends = 0;
long num_irecvs = 0;
long num_barriers = 0;
long num_sends = 0;
long num_recvs = 0;
long num_sendrecv = 0;
long num_waitalls = 0;

49 50 51
//std::map<int64_t, int> send_count;
//std::map<int64_t, int> isend_count;
//std::map<int64_t, int> allreduce_count;
52 53

struct shared_context {
54
    int my_rank;
55
    uint32_t wait_id;
56 57 58 59 60
    int num_ranks;
    char workload_name[MAX_NAME_LENGTH_WKLD];
    void * swm_obj;
    ABT_thread      producer;
    std::deque<struct codes_workload_op*> fifo;
61 62 63
};

struct rank_mpi_context {
64 65 66
    struct qhash_head hash_link;
    int app_id;
    struct shared_context sctx;
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84
};

typedef struct rank_mpi_compare {
    int app_id;
    int rank;
} rank_mpi_compare;

/*
 * peer: the receiving peer id 
 * comm_id: the communicator id being used
 * tag: tag id 
 * reqvc: virtual channel being used by the message (to be ignored)
 * rspvc: virtual channel being used by the message (to be ignored)
 * buf: the address of sender's buffer in memory
 * bytes: number of bytes to be sent 
 * reqrt and rsprt: routing types (to be ignored) */

void SWM_Send(SWM_PEER peer,
85 86 87 88 89 90 91 92 93
        SWM_COMM_ID comm_id,
        SWM_TAG tag,
        SWM_VC reqvc,
        SWM_VC rspvc,
        SWM_BUF buf,
        SWM_BYTES bytes,
        SWM_BYTES pktrspbytes,
        SWM_ROUTING_TYPE reqrt,
        SWM_ROUTING_TYPE rsprt)
94 95
{
    /* add an event in the shared queue and then yield */
96
    //    printf("\n Sending to rank %d ", comm_id);
97 98 99 100 101 102 103 104
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_SEND;
    wrkld_per_rank.u.send.tag = tag;
    wrkld_per_rank.u.send.num_bytes = bytes;
    wrkld_per_rank.u.send.dest_rank = peer;

#ifdef DBG_COMM
105
/*    if(tag != 1235 && tag != 1234) 
106 107 108 109 110 111 112 113 114 115
    {
        auto it = send_count.find(bytes);
        if(it == send_count.end())
        {
            send_count.insert(std::make_pair(bytes, 1));
        }
        else
        {
            it->second = it->second + 1;
        }
116
    }*/
117 118 119 120 121 122 123 124 125 126 127 128 129
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    wrkld_per_rank.u.send.source_rank = sctx->my_rank;
    sctx->fifo.push_back(&wrkld_per_rank);

    ABT_thread_yield_to(global_prod_thread);
130
    num_sends++;
131 132
}

133 134 135 136 137 138 139
/*
 * @param comm_id: communicator ID (For now, MPI_COMM_WORLD)
 * reqvc and rspvc: virtual channel IDs for request and response (ignore for
 * our purpose)
 * buf: buffer location for the call (ignore for our purpose)
 * reqrt and rsprt: routing types, ignore and use routing from config file instead. 
 * */
140 141 142 143 144 145 146 147 148 149 150
void SWM_Barrier(
        SWM_COMM_ID comm_id,
        SWM_VC reqvc,
        SWM_VC rspvc,
        SWM_BUF buf, 
        SWM_UNKNOWN auto1,
        SWM_UNKNOWN2 auto2,
        SWM_ROUTING_TYPE reqrt, 
        SWM_ROUTING_TYPE rsprt)
{
    /* Add an event in the shared queue and then yield */
151
#if 0
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_DELAY;
    /* TODO: Check how to convert cycle count into delay? */
    wrkld_per_rank.u.delay.nsecs = 0.1;

#ifdef DBG_COMM
    printf("\n Barrier delay %lf ", wrkld_per_rank.u.delay.nsecs);
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    sctx->fifo.push_back(&wrkld_per_rank);

    ABT_thread_yield_to(global_prod_thread);
172 173
#endif
#ifdef DBG_COMM
174
//     printf("\n barrier ");
175 176 177 178 179
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err;
180
    int rank, size, src, dest, mask;
181 182 183 184 185 186 187 188 189 190 191 192

    err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);

    rank = sctx->my_rank;
    size = sctx->num_ranks;
    mask = 0x1;

    while(mask < size) {
193
        dest = (rank + mask) % size;
194 195
        src = (rank - mask + size) % size;

196 197
        SWM_Sendrecv(comm_id, dest, 1234, reqvc, rspvc, 0, 0, 0,
                src,  1234, 0,  reqrt, rsprt);
198 199
        mask <<= 1;
    }
200
    num_barriers++;
201 202 203
}

void SWM_Isend(SWM_PEER peer,
204 205 206 207 208 209 210 211 212 213
        SWM_COMM_ID comm_id,
        SWM_TAG tag,
        SWM_VC reqvc,
        SWM_VC rspvc,
        SWM_BUF buf,
        SWM_BYTES bytes,
        SWM_BYTES pktrspbytes,
        uint32_t * handle,
        SWM_ROUTING_TYPE reqrt,
        SWM_ROUTING_TYPE rsprt)
214 215
{
    /* add an event in the shared queue and then yield */
216
    //    printf("\n Sending to rank %d ", comm_id);
217 218 219 220 221 222 223 224
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_ISEND;
    wrkld_per_rank.u.send.tag = tag;
    wrkld_per_rank.u.send.num_bytes = bytes;
    wrkld_per_rank.u.send.dest_rank = peer;

#ifdef DBG_COMM
225
/*    if(tag != 1235 && tag != 1234) 
226 227 228 229 230 231 232 233 234 235
    {
        auto it = isend_count.find(bytes);
        if(it == isend_count.end())
        {
            isend_count.insert(std::make_pair(bytes, 1));
        }
        else
        {
            it->second = it->second + 1;
        }
236
    }*/
237 238 239 240 241 242 243 244 245 246 247 248
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    wrkld_per_rank.u.send.source_rank = sctx->my_rank;
    sctx->fifo.push_back(&wrkld_per_rank);

249 250 251 252
    *handle = sctx->wait_id;
    wrkld_per_rank.u.send.req_id = *handle;
    sctx->wait_id++;

253
    ABT_thread_yield_to(global_prod_thread);
254
    num_isends++;
255 256 257 258 259 260 261 262 263 264 265 266
}
void SWM_Recv(SWM_PEER peer,
        SWM_COMM_ID comm_id,
        SWM_TAG tag,
        SWM_BUF buf)
{
    /* Add an event in the shared queue and then yield */
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_RECV;
    wrkld_per_rank.u.recv.tag = tag;
    wrkld_per_rank.u.recv.source_rank = peer;
267
    wrkld_per_rank.u.recv.num_bytes = 0;
268 269

#ifdef DBG_COMM
270
    //printf("\n recv op tag: %d source: %d ", tag, peer);
271 272 273 274 275 276 277 278 279 280 281 282 283
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    wrkld_per_rank.u.recv.dest_rank = sctx->my_rank;
    sctx->fifo.push_back(&wrkld_per_rank);

    ABT_thread_yield_to(global_prod_thread);
284
    num_recvs++;
285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
}

/* handle is for the request ID */
void SWM_Irecv(SWM_PEER peer,
        SWM_COMM_ID comm_id,
        SWM_TAG tag,
        SWM_BUF buf, 
        uint32_t* handle)
{
    /* Add an event in the shared queue and then yield */
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_IRECV;
    wrkld_per_rank.u.recv.tag = tag;
    wrkld_per_rank.u.recv.source_rank = peer;
    wrkld_per_rank.u.recv.num_bytes = 0;

#ifdef DBG_COMM
303
//    printf("\n irecv op tag: %d source: %d ", tag, peer);
304 305 306 307 308 309 310 311 312 313 314 315
#endif

    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    wrkld_per_rank.u.recv.dest_rank = sctx->my_rank;
    sctx->fifo.push_back(&wrkld_per_rank);
316 317 318 319
    
    *handle = sctx->wait_id;
    wrkld_per_rank.u.recv.req_id = *handle;
    sctx->wait_id++;
320 321

    ABT_thread_yield_to(global_prod_thread);
322
    num_irecvs++;
323 324 325 326
}

void SWM_Compute(long cycle_count)
{
327
    if(!cpu_freq)
328
        cpu_freq = 2.0;
329 330 331 332 333
    /* Add an event in the shared queue and then yield */
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_DELAY;
    /* TODO: Check how to convert cycle count into delay? */
334 335
    wrkld_per_rank.u.delay.nsecs = cycle_count;
    wrkld_per_rank.u.delay.seconds = (cycle_count) / (1000.0 * 1000.0 * 1000.0);
336 337 338 339 340 341 342 343 344 345 346 347
#ifdef DBG_COMM
    printf("\n compute op delay: %ld ", cycle_count);
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    sctx->fifo.push_back(&wrkld_per_rank);
348
	
349
    ABT_thread_yield_to(global_prod_thread);
350

351 352 353 354 355 356 357 358 359 360 361 362
}

void SWM_Wait(uint32_t req_id)
{
    /* Add an event in the shared queue and then yield */
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_WAIT;
    /* TODO: Check how to convert cycle count into delay? */
    wrkld_per_rank.u.wait.req_id = req_id;

#ifdef DBG_COMM
363
//    printf("\n wait op req_id: %"PRIu32"\n", req_id);
364
//      printf("\n wait ");
365 366 367 368 369 370 371 372 373 374
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    sctx->fifo.push_back(&wrkld_per_rank);
375

376 377 378 379 380
    ABT_thread_yield_to(global_prod_thread);
}

void SWM_Waitall(int len, uint32_t * req_ids)
{
381
    num_waitalls++;
382 383 384 385 386 387 388 389 390 391 392 393
    /* Add an event in the shared queue and then yield */
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_WAITALL;
    /* TODO: Check how to convert cycle count into delay? */
    wrkld_per_rank.u.waits.count = len;
    wrkld_per_rank.u.waits.req_ids = (unsigned int*)calloc(len, sizeof(int));    

    for(int i = 0; i < len; i++)
        wrkld_per_rank.u.waits.req_ids[i] = req_ids[i];

#ifdef DBG_COMM
394
//    for(int i = 0; i < len; i++)
395
//        printf("\n wait op len %d req_id: %"PRIu32"\n", len, req_ids[i]);
396 397 398 399 400 401 402 403 404 405
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    sctx->fifo.push_back(&wrkld_per_rank);
406

407 408 409 410
    ABT_thread_yield_to(global_prod_thread);
}

void SWM_Sendrecv(
411 412 413 414 415 416 417 418 419 420 421 422 423
        SWM_COMM_ID comm_id,
        SWM_PEER sendpeer,
        SWM_TAG sendtag,
        SWM_VC sendreqvc,
        SWM_VC sendrspvc,
        SWM_BUF sendbuf,
        SWM_BYTES sendbytes,
        SWM_BYTES pktrspbytes,
        SWM_PEER recvpeer,
        SWM_TAG recvtag,
        SWM_BUF recvbuf,
        SWM_ROUTING_TYPE reqrt,
        SWM_ROUTING_TYPE rsprt)
424
{
425
    //    printf("\n Sending to %d receiving from %d ", sendpeer, recvpeer);
426 427 428 429 430 431 432 433 434 435 436 437 438
    struct codes_workload_op send_op;

    send_op.op_type = CODES_WK_SEND;
    send_op.u.send.tag = sendtag;
    send_op.u.send.num_bytes = sendbytes;
    send_op.u.send.dest_rank = sendpeer;

    /* Add an event in the shared queue and then yield */
    struct codes_workload_op recv_op;

    recv_op.op_type = CODES_WK_RECV;
    recv_op.u.recv.tag = recvtag;
    recv_op.u.recv.source_rank = recvpeer;
439
    recv_op.u.recv.num_bytes = 0;
440

441
#ifdef DBG_COMM
442
/*    if(sendtag != 1235 && sendtag != 1234) 
443 444 445 446 447 448 449 450 451 452
    {
        auto it = send_count.find(sendbytes);
        if(it == send_count.end())
        {
            send_count.insert(std::make_pair(sendbytes, 1));
        }
        else
        {
            it->second = it->second + 1;
        }
453
    }*/
454 455 456 457 458 459 460 461 462 463 464 465 466 467 468
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    recv_op.u.recv.dest_rank = sctx->my_rank;
    send_op.u.send.source_rank = sctx->my_rank;
    sctx->fifo.push_back(&send_op);
    sctx->fifo.push_back(&recv_op);

    ABT_thread_yield_to(global_prod_thread);
469
    num_sendrecv++;
470 471
}

472
/* @param count: number of bytes in Allreduce
473 474 475 476 477 478 479 480 481
 * @param respbytes: number of bytes to be sent in response (ignore for our
 * purpose)
 * $params comm_id: communicator ID (MPI_COMM_WORLD for our case)
 * @param sendreqvc: virtual channel of the sender request (ignore for our
 * purpose)
 * @param sendrspvc: virtual channel of the response request (ignore for our
 * purpose)
 * @param sendbuf and rcvbuf: buffers for send and receive calls (ignore for
 * our purpose) */
482
void SWM_Allreduce(
483
        SWM_BYTES count,
484 485 486 487 488 489 490
        SWM_BYTES respbytes,
        SWM_COMM_ID comm_id,
        SWM_VC sendreqvc,
        SWM_VC sendrspvc,
        SWM_BUF sendbuf,
        SWM_BUF rcvbuf)
{
491
#if 0
492
    /* TODO: For now, simulate a constant delay for ALlreduce*/
493
    //    printf("\n Allreduce bytes %d ", bytes);
494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514
    /* Add an event in the shared queue and then yield */
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_DELAY;
    /* TODO: Check how to convert cycle count into delay? */
    wrkld_per_rank.u.delay.nsecs = bytes + 0.1;

#ifdef DBG_COMM
    printf("\n Allreduce delay %lf ", wrkld_per_rank.u.delay.nsecs);
#endif
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    sctx->fifo.push_back(&wrkld_per_rank);

    ABT_thread_yield_to(global_prod_thread);
515 516
#endif

517 518 519 520 521 522 523 524 525 526 527
#ifdef DBG_COMM
        auto it = allreduce_count.find(count);
        if(it == allreduce_count.end())
        {
            allreduce_count.insert(std::make_pair(count, 1));
        }
        else
        {
            it->second = it->second + 1;
        }
#endif
528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556
    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);

    int comm_size, i, send_idx, recv_idx, last_idx, send_cnt, recv_cnt;
    int pof2, mask, rem, newrank, newdst, dst, *cnts, *disps;
    int rank = sctx->my_rank;
    comm_size = sctx->num_ranks;

    cnts = disps = NULL;

    pof2 = 1;
    while (pof2 <= comm_size) pof2 <<= 1;
    pof2 >>=1;

    rem = comm_size - pof2;

    /* In the non-power-of-two case, all even-numbered
       processes of rank < 2*rem send their data to
       (rank+1). These even-numbered processes no longer
       participate in the algorithm until the very end. The
       remaining processes form a nice power-of-two. */
    if (rank < 2*rem) {
        if (rank % 2 == 0) { /* even */
557
            SWM_Send(rank+1, comm_id, 1235, sendreqvc, sendrspvc, 0, count, 1, 0, 0);
558 559
            newrank = -1;
        } else { /* odd */
560
            SWM_Recv(rank-1, comm_id, 1235, 0);
561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582
            newrank = rank / 2;
        }
    } else {
        newrank = rank - rem;
    }

    /* If op is user-defined or count is less than pof2, use
       recursive doubling algorithm. Otherwise do a reduce-scatter
       followed by allgather. (If op is user-defined,
       derived datatypes are allowed and the user could pass basic
       datatypes on one process and derived on another as long as
       the type maps are the same. Breaking up derived
       datatypes to do the reduce-scatter is tricky, therefore
       using recursive doubling in that case.) */
    if (newrank != -1) {
        if ((count <= ALLREDUCE_SHORT_MSG_SIZE) || (count < pof2)) {

            mask = 0x1;
            while (mask < pof2) {
                newdst = newrank ^ mask;
                dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

583 584
                SWM_Sendrecv(comm_id, dst, 1235, sendreqvc, sendrspvc, 0,
                        count, 1, dst, 1235, 0, 0, 0);
585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625

                mask <<= 1;
            }
        } else {
            /* do a reduce-scatter followed by allgather */
            /* for the reduce-scatter, calculate the count that
               each process receives and the displacement within
               the buffer */

            cnts = (int*)malloc(pof2*sizeof(int));
            disps = (int*)malloc(pof2*sizeof(int));

            for (i=0; i<(pof2-1); i++)
                cnts[i] = count/pof2;
            cnts[pof2-1] = count - (count/pof2)*(pof2-1);

            disps[0] = 0;
            for (i=1; i<pof2; i++)
                disps[i] = disps[i-1] + cnts[i-1];

            mask = 0x1;
            send_idx = recv_idx = 0;
            last_idx = pof2;
            while (mask < pof2) {
                newdst = newrank ^ mask;
                dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
                send_cnt = recv_cnt = 0;
                if (newrank < newdst) {
                    send_idx = recv_idx + pof2/(mask*2);
                    for (i=send_idx; i<last_idx; i++)
                        send_cnt += cnts[i];
                    for (i=recv_idx; i<send_idx; i++)
                        recv_cnt += cnts[i];
                } else {
                    recv_idx = send_idx + pof2/(mask*2);
                    for (i=send_idx; i<recv_idx; i++)
                        send_cnt += cnts[i];
                    for (i=recv_idx; i<last_idx; i++)
                        recv_cnt += cnts[i];
                }

626 627
                SWM_Sendrecv(comm_id, dst, 1235, sendreqvc, sendrspvc, 0,
                        send_cnt, 1, dst, 1235, 0, 0, 0);
628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660

                send_idx = recv_idx;
                mask <<= 1;

                if(mask < pof2)
                    last_idx = recv_idx + pof2/mask;
            }

            /* now do the allgather */
            mask >>= 1;
            while (mask > 0) {
                newdst = newrank ^ mask;
                /* find real rank of dest */
                dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

                send_cnt = recv_cnt = 0;
                if (newrank < newdst) {
                    if (mask != pof2/2)
                        last_idx = last_idx + pof2/(mask*2);

                    recv_idx = send_idx + pof2/(mask*2);
                    for (i=send_idx; i<recv_idx; i++)
                        send_cnt += cnts[i];
                    for (i=recv_idx; i<last_idx; i++)
                        recv_cnt += cnts[i];
                } else {
                    recv_idx = send_idx - pof2/(mask*2);
                    for (i=send_idx; i<last_idx; i++)
                        send_cnt += cnts[i];
                    for (i=recv_idx; i<send_idx; i++)
                        recv_cnt += cnts[i];
                }

661 662
                SWM_Sendrecv(comm_id, dst, 1235, sendreqvc, sendrspvc, 0,
                        send_cnt, 1, dst, 1235, 0, 0, 0);
663 664 665 666 667 668 669 670 671 672

                if (newrank > newdst) send_idx = recv_idx;

                mask >>= 1;
            }
        }
    }

    if(rank < 2*rem) {
        if(rank % 2) {/* odd */
673
            SWM_Send(rank-1, comm_id, 1235, sendreqvc, sendrspvc, 0, count, 1, 0, 0);
674
        } else {
675
            SWM_Recv(rank+1, comm_id, 1235, 0);
676 677 678 679 680
        }
    }

    if(cnts) free(cnts);
    if(disps) free(disps);
681 682

    num_allreduce++;
683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717
}

void SWM_Allreduce(
        SWM_BYTES bytes,
        SWM_BYTES respbytes,
        SWM_COMM_ID comm_id,
        SWM_VC sendreqvc,
        SWM_VC sendrspvc,
        SWM_BUF sendbuf,
        SWM_BUF rcvbuf,
        SWM_UNKNOWN auto1,
        SWM_UNKNOWN2 auto2,
        SWM_ROUTING_TYPE reqrt,
        SWM_ROUTING_TYPE rsprt)
{
    SWM_Allreduce(bytes, respbytes, comm_id, sendreqvc, sendrspvc, sendbuf, rcvbuf);
}

void SWM_Finalize()
{
    /* Add an event in the shared queue and then yield */
    struct codes_workload_op wrkld_per_rank;

    wrkld_per_rank.op_type = CODES_WK_END;

    /* Retreive the shared context state */
    ABT_thread prod;
    void * arg;
    int err = ABT_thread_self(&prod);
    assert(err == ABT_SUCCESS);
    err =  ABT_thread_get_arg(prod, &arg);
    assert(err == ABT_SUCCESS);
    struct shared_context * sctx = static_cast<shared_context*>(arg);
    sctx->fifo.push_back(&wrkld_per_rank);

718
#ifdef DBG_COMM 
719
/*    auto it = allreduce_count.begin();
720 721 722 723 724 725 726 727 728 729 730 731 732 733 734
    for(; it != allreduce_count.end(); it++)
    {
        cout << "\n Allreduce " << it->first << " " << it->second;
    }
    
    it = send_count.begin();
    for(; it != send_count.end(); it++)
    {
        cout << "\n Send " << it->first << " " << it->second;
    }
    
    it = isend_count.begin();
    for(; it != isend_count.end(); it++)
    {
        cout << "\n isend " << it->first << " " << it->second;
735
    }*/
736
#endif
737 738
//#ifdef DBG_COMM
//    printf("\n finalize workload for rank %d ", sctx->my_rank);
739
//    printf("\n finalize workload for rank %d num_sends %d num_recvs %d num_isends %d num_irecvs %d num_allreduce %d num_barrier %d num_waitalls %d", sctx->my_rank, num_sends, num_recvs, num_isends, num_irecvs, num_allreduce, num_barriers, num_waitalls);
740
//#endif
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
    ABT_thread_yield_to(global_prod_thread);
}

static int hash_rank_compare(void *key, struct qhash_head *link)
{
    rank_mpi_compare *in = (rank_mpi_compare*)key;
    rank_mpi_context *tmp;

    tmp = qhash_entry(link, rank_mpi_context, hash_link);
    if (tmp->sctx.my_rank == in->rank && tmp->app_id == in->app_id)
        return 1;
    return 0;
}
static void workload_caller(void * arg)
{
    shared_context* sctx = static_cast<shared_context*>(arg);

    if(strcmp(sctx->workload_name, "lammps") == 0)
    {
        LAMMPS_SWM * lammps_swm = static_cast<LAMMPS_SWM*>(sctx->swm_obj);
        lammps_swm->call();
    }
    else if(strcmp(sctx->workload_name, "nekbone") == 0) 
    {
        NEKBONESWMUserCode * nekbone_swm = static_cast<NEKBONESWMUserCode*>(sctx->swm_obj);
        nekbone_swm->call();
    }
768 769 770 771 772
    else if(strcmp(sctx->workload_name, "nearest_neighbor") == 0)
    {
       NearestNeighborSWMUserCode * nn_swm = static_cast<NearestNeighborSWMUserCode*>(sctx->swm_obj);
       nn_swm->call();
    }
773 774 775 776 777
    // else if(strcmp(sctx->workload_name, "incast") == 0 || strcmp(sctx->workload_name, "incast1") == 0 || strcmp(sctx->workload_name, "incast2") == 0)
    // {
    //    AllToOneSWMUserCode * incast_swm = static_cast<AllToOneSWMUserCode*>(sctx->swm_obj);
    //    incast_swm->call();
    // }
778 779 780 781 782 783
}
static int comm_online_workload_load(const char * params, int app_id, int rank)
{
    /* LOAD parameters from JSON file*/
    online_comm_params * o_params = (online_comm_params*)params;
    int nprocs = o_params->nprocs;
784

785 786
    rank_mpi_context *my_ctx = new rank_mpi_context;
    //my_ctx = (rank_mpi_context*)caloc(1, sizeof(rank_mpi_context));  
787 788
    assert(my_ctx); 
    my_ctx->sctx.my_rank = rank; 
789
    my_ctx->sctx.num_ranks = nprocs;
790
    my_ctx->sctx.wait_id = 0;
791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810
    my_ctx->app_id = app_id;

    void** generic_ptrs;
    int array_len = 1;
    generic_ptrs = (void**)calloc(array_len,  sizeof(void*));
    generic_ptrs[0] = (void*)&rank;

    strcpy(my_ctx->sctx.workload_name, o_params->workload_name);
    boost::property_tree::ptree root;
    string path;
    path.append(SWM_DATAROOTDIR);

    if(strcmp(o_params->workload_name, "lammps") == 0)
    {
        path.append("/lammps_workload.json");
    }
    else if(strcmp(o_params->workload_name, "nekbone") == 0)
    {
        path.append("/workload.json"); 
    }
811 812 813 814
    else if(strcmp(o_params->workload_name, "nearest_neighbor") == 0)
    {
        path.append("/skeleton.json"); 
    }
815 816 817 818 819 820 821 822 823 824 825 826
    // else if(strcmp(o_params->workload_name, "incast") == 0)
    // {
    //     path.append("/incast.json"); 
    // }
    // else if(strcmp(o_params->workload_name, "incast1") == 0)
    // {
    //     path.append("/incast1.json"); 
    // }
    // else if(strcmp(o_params->workload_name, "incast2") == 0)
    // {
    //     path.append("/incast2.json"); 
    // }
827 828 829
    else
        tw_error(TW_LOC, "\n Undefined workload type %s ", o_params->workload_name);

830
    try {
831
        std::ifstream jsonFile(path.c_str());
832 833
        boost::property_tree::json_parser::read_json(jsonFile, root);
        uint32_t process_cnt = root.get<uint32_t>("jobs.size", 1);
834
        cpu_freq = root.get<double>("jobs.cfg.cpu_freq") / 1e9; 
835 836 837 838 839 840
    }
    catch(std::exception & e)
    {
        printf("%s \n", e.what());
        return -1;
    }
841 842 843 844 845 846 847
    if(strcmp(o_params->workload_name, "lammps") == 0)
    {
        LAMMPS_SWM * lammps_swm = new LAMMPS_SWM(root, generic_ptrs);
        my_ctx->sctx.swm_obj = (void*)lammps_swm;
    }
    else if(strcmp(o_params->workload_name, "nekbone") == 0)
    {
848 849
        NEKBONESWMUserCode * nekbone_swm = new NEKBONESWMUserCode(root, generic_ptrs);
        my_ctx->sctx.swm_obj = (void*)nekbone_swm;
850
    }
851 852 853 854 855
    else if(strcmp(o_params->workload_name, "nearest_neighbor") == 0)
    {
        NearestNeighborSWMUserCode * nn_swm = new NearestNeighborSWMUserCode(root, generic_ptrs);
        my_ctx->sctx.swm_obj = (void*)nn_swm;
    }
856 857 858 859 860
    // else if(strcmp(o_params->workload_name, "incast") == 0 || strcmp(o_params->workload_name, "incast1") == 0 || strcmp(o_params->workload_name, "incast2") == 0)
    // {
    //     AllToOneSWMUserCode * incast_swm = new AllToOneSWMUserCode(root, generic_ptrs);
    //     my_ctx->sctx.swm_obj = (void*)incast_swm;
    // }
861 862

    if(global_prod_thread == NULL)
863 864
    {
        ABT_xstream_self(&self_es);
865
        ABT_thread_self(&global_prod_thread);
866
    }
867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888
    ABT_thread_create_on_xstream(self_es, 
            &workload_caller, (void*)&(my_ctx->sctx),
            ABT_THREAD_ATTR_NULL, &(my_ctx->sctx.producer));

    rank_mpi_compare cmp;
    cmp.app_id = app_id;
    cmp.rank = rank;

    if(!rank_tbl)
    {
        rank_tbl = qhash_init(hash_rank_compare, quickhash_64bit_hash, nprocs);
        if(!rank_tbl)
            return -1;
    }
    qhash_add(rank_tbl, &cmp, &(my_ctx->hash_link));
    rank_tbl_pop++;

    return 0;
}

static void comm_online_workload_get_next(int app_id, int rank, struct codes_workload_op * op)
{
889 890 891
    /* At this point, we will use the "call" function. The send/receive/wait
     * definitions will be replaced by our own function definitions that will do a
     * yield to argobots if an event is not available. */
892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912
    /* if shared queue is empty then yield */

    rank_mpi_context * temp_data;
    struct qhash_head * hash_link = NULL;
    rank_mpi_compare cmp;
    cmp.rank = rank;
    cmp.app_id = app_id;
    hash_link = qhash_search(rank_tbl, &cmp);
    if(!hash_link)
    {
        printf("\n not found for rank id %d , %d", rank, app_id);
        op->op_type = CODES_WK_END;
        return;
    }
    temp_data = qhash_entry(hash_link, rank_mpi_context, hash_link);
    assert(temp_data);
    while(temp_data->sctx.fifo.empty())
    {
        ABT_thread_yield_to(temp_data->sctx.producer); 
    }
    struct codes_workload_op * front_op = temp_data->sctx.fifo.front();
913
    assert(front_op);
914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944
    *op = *front_op;
    temp_data->sctx.fifo.pop_front();
    return;
}
static int comm_online_workload_get_rank_cnt(const char *params, int app_id)
{
    online_comm_params * o_params = (online_comm_params*)params;
    int nprocs = o_params->nprocs;
    return nprocs;
}

static int comm_online_workload_finalize(const char* params, int app_id, int rank)
{
    rank_mpi_context * temp_data;
    struct qhash_head * hash_link = NULL;
    rank_mpi_compare cmp;
    cmp.rank = rank;
    cmp.app_id = app_id;
    hash_link = qhash_search(rank_tbl, &cmp);
    if(!hash_link)
    {
        printf("\n not found for rank id %d , %d", rank, app_id);
        return -1;
    }
    temp_data = qhash_entry(hash_link, rank_mpi_context, hash_link);
    assert(temp_data);

    ABT_thread_join(temp_data->sctx.producer);    
    ABT_thread_free(&(temp_data->sctx.producer));
    return 0;
}
945
extern "C" {
946 947 948
/* workload method name and function pointers for the CODES workload API */
struct codes_workload_method online_comm_workload_method =
{
949 950 951 952 953 954 955 956 957 958 959 960 961 962
    //.method_name =
    (char*)"online_comm_workload",
    //.codes_workload_read_config = 
    NULL,
    //.codes_workload_load = 
    comm_online_workload_load,
    //.codes_workload_get_next = 
    comm_online_workload_get_next,
    // .codes_workload_get_next_rc2 = 
    NULL,
    // .codes_workload_get_rank_cnt
    comm_online_workload_get_rank_cnt,
    // .codes_workload_finalize = 
    comm_online_workload_finalize
963
};
964
} // closing brace for extern "C"
965