margo-p2p-bw.c 16.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
/*
 * Copyright (c) 2017 UChicago Argonne, LLC
 *
 * See COPYRIGHT in top-level directory.
 */

/* Effective streaming bandwidth test, as measured by client including RPC
 * used to start and complete the streaming operation.
 *
 * NOTE: This test is not as clean as it could be.  Because it is set up as 
 * an MPI program, the server is able to make assumptions about the pattern; 
 * it assumes that it should set a fill pattern after the first RPC and shut
 * down after the second RPC.  It assumes it can read all params from argv.
 */


#include <unistd.h>
#include <stdio.h>
#include <string.h>
#include <assert.h>

#include <mpi.h>

#include <margo.h>
#include <mercury.h>
#include <abt.h>
#include <ssg.h>
#include <ssg-mpi.h>

struct options
{
    int xfer_size;
    int duration_seconds;
    int concurrency;
    int threads;
    unsigned int mercury_timeout_client;
    unsigned int mercury_timeout_server;
    char* diag_file_name;
    char* na_transport;
};

#define BW_TOTAL_MEM_SIZE 2147483648UL

44
static int parse_args(int argc, char **argv, struct options *opts);
45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
static void usage(void);

MERCURY_GEN_PROC(bw_rpc_in_t,
        ((hg_bulk_t)(bulk_handle))\
        ((int32_t)(op)))
MERCURY_GEN_PROC(bw_rpc_out_t,
        ((hg_size_t)(bytes_moved)))
DECLARE_MARGO_RPC_HANDLER(bw_ult);

static int run_benchmark(hg_id_t id, ssg_member_id_t target, 
    ssg_group_id_t gid, margo_instance_id mid);

struct bw_worker_arg
{
    double start_tm;
    margo_instance_id mid;
    ABT_mutex *cur_off_mutex;
    size_t *cur_off;
    hg_bulk_t *client_bulk_handle;
    const hg_addr_t *target_addr;
    hg_size_t bytes_moved;
    hg_bulk_op_t op;
};

static void bw_worker(void *_arg);

static hg_id_t g_bw_id;
static ABT_pool g_transfer_pool;
static ABT_eventual g_bw_done_eventual;
static struct options g_opts;
static char *g_buffer = NULL;
static hg_size_t g_buffer_size = BW_TOTAL_MEM_SIZE;
static hg_bulk_t g_bulk_handle = HG_BULK_NULL;

int main(int argc, char **argv) 
{
    margo_instance_id mid;
    int nranks;
    int ret;
    ssg_group_id_t gid;
    ssg_member_id_t self;
    int rank;
    int namelen;
    char processor_name[MPI_MAX_PROCESSOR_NAME];
    int i;
    ABT_xstream *bw_worker_xstreams = NULL;
    ABT_sched *bw_worker_scheds = NULL;
92
    struct hg_init_info hii;
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107

    MPI_Init(&argc, &argv);

    /* 2 process one-way bandwidth measurement only */
    MPI_Comm_size(MPI_COMM_WORLD, &nranks);
    if(nranks != 2)
    {
        usage();
        exit(EXIT_FAILURE);
    }
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Get_processor_name(processor_name,&namelen);
    printf("Process %d of %d is on %s\n",
	rank, nranks, processor_name);

108 109 110 111 112 113 114
    ret = parse_args(argc, argv, &g_opts);
    if(ret < 0)
    {
        if(rank == 0)
            usage();
        exit(EXIT_FAILURE);
    }
115 116 117 118 119 120 121 122 123 124

    /* allocate one big buffer for rdma transfers */
    g_buffer = calloc(g_buffer_size, 1);
    if(!g_buffer)
    {
        perror("calloc");
        fprintf(stderr, "Error: unable to allocate %lu byte buffer.\n", g_buffer_size);
        return(-1);
    }

125
    memset(&hii, 0, sizeof(hii));
126 127 128 129 130 131 132 133 134 135 136 137 138 139
    if((rank == 0 && g_opts.mercury_timeout_client == 0) ||
       (rank == 1 && g_opts.mercury_timeout_server == 0))
    {
        
        /* If mercury timeout of zero is requested, then set
         * init option to NO_BLOCK.  This allows some transports to go
         * faster because they do not have to set up or maintain the data
         * structures necessary for signaling completion on blocked
         * operations.
         */
        hii.na_init_info.progress_mode = NA_NO_BLOCK;
    }

    /* actually start margo */
140
    mid = margo_init_opt(g_opts.na_transport, MARGO_SERVER_MODE, &hii, 0, -1);
141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
    assert(mid);

    if(g_opts.diag_file_name)
        margo_diag_start(mid);

    /* adjust mercury timeout in Margo if requested */
    if(rank == 0 && g_opts.mercury_timeout_client != UINT_MAX)
        margo_set_param(mid, MARGO_PARAM_PROGRESS_TIMEOUT_UB, &g_opts.mercury_timeout_client);
    if(rank == 1 && g_opts.mercury_timeout_server != UINT_MAX)
        margo_set_param(mid, MARGO_PARAM_PROGRESS_TIMEOUT_UB, &g_opts.mercury_timeout_server);

    g_bw_id = MARGO_REGISTER(
        mid, 
        "bw_rpc", 
        bw_rpc_in_t,
        bw_rpc_out_t,
        bw_ult);

    /* set up group */
    ret = ssg_init(mid);
    assert(ret == 0);
    gid = ssg_group_create_mpi("margo-p2p-latency", MPI_COMM_WORLD, NULL, NULL);
    assert(gid != SSG_GROUP_ID_NULL);

    assert(ssg_get_group_size(gid) == 2);

    self = ssg_get_group_self_id(gid);

    if(self == 1)
    {
        /* server side: prep everything before letting the client initiate
         * benchmark
         */
        void* buffer = g_buffer;

        /* register memory for xfer */
        ret = margo_bulk_create(mid, 1, &buffer, &g_buffer_size, HG_BULK_READWRITE, &g_bulk_handle);
        assert(ret == 0);

        /* set up abt pool */
        if(g_opts.threads == 0)
        {
183 184 185
            ABT_pool pool;
            ABT_xstream xstream;
            
186
            /* run bulk transfers from primary pool on server */
187 188 189 190 191 192 193

            ret = ABT_xstream_self(&xstream);
            assert(ret == 0);

            ret = ABT_xstream_get_main_pools(xstream, 1, &pool);
            assert(ret == 0);

194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
            g_transfer_pool = pool;
        }
        else
        {
            /* run bulk transfers from a dedicated pool */
            bw_worker_xstreams = malloc(
                    g_opts.threads * sizeof(*bw_worker_xstreams));
            bw_worker_scheds = malloc(
                    g_opts.threads * sizeof(*bw_worker_scheds));
            assert(bw_worker_xstreams && bw_worker_scheds);

            ret = ABT_pool_create_basic(ABT_POOL_FIFO_WAIT, ABT_POOL_ACCESS_MPMC,
                    ABT_TRUE, &g_transfer_pool);
            assert(ret == ABT_SUCCESS);
            for(i = 0; i < g_opts.threads; i++)
            {
                ret = ABT_sched_create_basic(ABT_SCHED_BASIC_WAIT, 1, &g_transfer_pool,
                        ABT_SCHED_CONFIG_NULL, &bw_worker_scheds[i]);
                assert(ret == ABT_SUCCESS);
                ret = ABT_xstream_create(bw_worker_scheds[i], &bw_worker_xstreams[i]);
                assert(ret == ABT_SUCCESS);
            }
        }

        /* signaling mechanism for server to exit at conclusion of test */
        ret = ABT_eventual_create(0, &g_bw_done_eventual);
        assert(ret == 0);
    }

    MPI_Barrier(MPI_COMM_WORLD);

    if(self == 0)
    {
        /* ssg id 0 (client) initiates benchmark */
        ret = run_benchmark(g_bw_id, 1, gid, mid);
        assert(ret == 0);
    }
    else
    {
        /* ssg id 1 (server) waits for test RPC to complete */
        int i;

        ABT_eventual_wait(g_bw_done_eventual, NULL);

        /* cleanup dedicated pool if needed */
        for (i = 0; i < g_opts.threads; i++) {
            ABT_xstream_join(bw_worker_xstreams[i]);
            ABT_xstream_free(&bw_worker_xstreams[i]);
        }
        if(bw_worker_xstreams)
            free(bw_worker_xstreams);
        if(bw_worker_scheds)
            free(bw_worker_scheds);
    
        margo_bulk_free(g_bulk_handle);
    }

    ssg_group_destroy(gid);
    ssg_finalize();

    if(g_opts.diag_file_name)
        margo_diag_dump(mid, g_opts.diag_file_name, 1);

    free(g_buffer);

    margo_finalize(mid);
    MPI_Finalize();

    return 0;
}

265
static int parse_args(int argc, char **argv, struct options *opts)
266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286
{
    int opt;
    int ret;

    memset(opts, 0, sizeof(*opts));

    opts->concurrency = 1;

    /* default to using whatever the standard timeout is in margo */
    opts->mercury_timeout_client = UINT_MAX;
    opts->mercury_timeout_server = UINT_MAX; 

    while((opt = getopt(argc, argv, "n:x:c:T:d:t:D:")) != -1)
    {
        switch(opt)
        {
            case 'd':
                opts->diag_file_name = strdup(optarg);
                if(!opts->diag_file_name)
                {
                    perror("strdup");
287
                    return(-1);
288 289 290 291 292
                }
                break;
            case 'x':
                ret = sscanf(optarg, "%d", &opts->xfer_size);
                if(ret != 1)
293
                    return(-1);
294 295 296 297
                break;
            case 'c':
                ret = sscanf(optarg, "%d", &opts->concurrency);
                if(ret != 1)
298
                    return(-1);
299 300 301 302
                break;
            case 'T':
                ret = sscanf(optarg, "%d", &opts->threads);
                if(ret != 1)
303
                    return(-1);
304 305 306 307
                break;
            case 'D':
                ret = sscanf(optarg, "%d", &opts->duration_seconds);
                if(ret != 1)
308
                    return(-1);
309 310 311 312
                break;
            case 't':
                ret = sscanf(optarg, "%u,%u", &opts->mercury_timeout_client, &opts->mercury_timeout_server);
                if(ret != 2)
313
                    return(-1);
314 315 316 317 318 319
                break;
            case 'n':
                opts->na_transport = strdup(optarg);
                if(!opts->na_transport)
                {
                    perror("strdup");
320
                    return(-1);
321 322 323
                }
                break;
            default:
324
                return(-1);
325 326 327 328 329
        }
    }

    if(opts->xfer_size < 1 || opts->concurrency < 1 || opts->duration_seconds < 1 || !opts->na_transport)
    {
330
        return(-1);
331 332
    }

333
    return(0);
334 335 336 337 338 339 340 341 342 343 344 345
}

static void usage(void)
{
    fprintf(stderr,
        "Usage: "
        "margo-p2p-bw -x <xfer_size> -D <duration> -n <na>\n"
        "\t-x <xfer_size> - size of each bulk tranfer in bytes\n"
        "\t-D <duration> - duration of test in seconds\n"
        "\t-n <na> - na transport\n"
        "\t[-c concurrency] - number of concurrent operations to issue with ULTs\n"
        "\t[-T <os threads] - number of dedicated operating system threads to run ULTs on\n"
346 347
        "\t[-d filename] - enable diagnostics output\n"
        "\t[-t client_progress_timeout,server_progress_timeout] # use \"-t 0,0\" to busy spin\n"
348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445
        "\t\texample: mpiexec -n 2 ./margo-p2p-bw -x 4096 -D 30 -n verbs://\n"
        "\t\t(must be run with exactly 2 processes\n");
    
    return;
}

/* service an RPC that runs the bandwidth test */
static void bw_ult(hg_handle_t handle)
{
    int i;
    bw_rpc_in_t in;
    bw_rpc_out_t out;
    ABT_thread *tid_array;
    struct bw_worker_arg *arg_array;
    int ret;
    double start_time;
    margo_instance_id mid;
    const struct hg_info *hgi;
    size_t cur_off = 0;
    ABT_mutex cur_off_mutex;
    unsigned long bytes_to_check = 0;
    hg_size_t x;

    ABT_mutex_create(&cur_off_mutex);
    
    /* get handle info and margo instance */
    hgi = margo_get_info(handle);
    assert(hgi);
    mid = margo_hg_info_get_instance(hgi);
    assert(mid != MARGO_INSTANCE_NULL);

    ret = margo_get_input(handle, &in);
        assert(ret == HG_SUCCESS);

    tid_array = malloc(g_opts.concurrency * sizeof(*tid_array));
    assert(tid_array);
    arg_array = calloc(g_opts.concurrency, sizeof(*arg_array));
    assert(arg_array);

    start_time = ABT_get_wtime();
    /* create requested number of workers to run transfer */
    for(i=0; i<g_opts.concurrency; i++)
    {
        arg_array[i].start_tm = start_time;
        arg_array[i].mid = mid;
        arg_array[i].cur_off = &cur_off;
        arg_array[i].cur_off_mutex = &cur_off_mutex;
        arg_array[i].client_bulk_handle = &in.bulk_handle;
        arg_array[i].target_addr = &hgi->addr;
        arg_array[i].op = in.op;

        ret = ABT_thread_create(g_transfer_pool, bw_worker, &arg_array[i], ABT_THREAD_ATTR_NULL, &tid_array[i]);
        assert(ret == 0);
    }

    out.bytes_moved = 0;
    for(i=0; i<g_opts.concurrency; i++)
    {
        ABT_thread_join(tid_array[i]);
        ABT_thread_free(&tid_array[i]);
        
        out.bytes_moved += arg_array[i].bytes_moved;
    }
    
    margo_respond(handle, &out);

    if(in.op == HG_BULK_PULL)
    {
        /* calculate how many bytes of the buffer have been transferred */
        bytes_to_check = (g_buffer_size / g_opts.xfer_size) * g_opts.xfer_size;
        if(out.bytes_moved < bytes_to_check)
            bytes_to_check = out.bytes_moved;

        /* check integrity of fill pattern.  Note that this isn't as strong as
         * checking every RDMA transfer separately since we are looping around
         * and overwriting in a ring-buffer style.  We could corrupt early and
         * but then overwrite it with correct results on a later pass.
         */
        for(x=0; x<(bytes_to_check/sizeof(x)); x++)
        {
            assert(((hg_size_t*)g_buffer)[x] == x);
        }

        /* fill pattern for return trip, increment each value by 1 */
        for(x=0; x<(g_buffer_size/sizeof(x)); x++)
            ((hg_size_t*)g_buffer)[x] = x+1;
    }

    margo_free_input(handle, &in);
    margo_destroy(handle);

    free(tid_array);

    ABT_mutex_free(&cur_off_mutex);
    
    if(arg_array[0].op == HG_BULK_PUSH)
        ABT_eventual_set(g_bw_done_eventual, NULL, 0);

Philip Carns's avatar
Philip Carns committed
446 447
    free(arg_array);

448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570
    return;
}
DEFINE_MARGO_RPC_HANDLER(bw_ult)

static int run_benchmark(hg_id_t id, ssg_member_id_t target, 
    ssg_group_id_t gid, margo_instance_id mid)
{
    hg_handle_t handle;
    hg_addr_t target_addr;
    int ret;
    bw_rpc_in_t in;
    bw_rpc_out_t out;
    void* buffer = g_buffer;
    hg_size_t i;
    hg_size_t bytes_to_check;
    double start_ts, end_ts;

    /* fill pattern in origin buffer */
    for(i=0; i<(g_buffer_size/sizeof(i)); i++)
        ((hg_size_t*)buffer)[i] = i;

    target_addr = ssg_get_addr(gid, target);
    assert(target_addr != HG_ADDR_NULL);

    ret = margo_create(mid, target_addr, id, &handle);
    assert(ret == 0);

    ret = margo_bulk_create(mid, 1, &buffer, &g_buffer_size, HG_BULK_READWRITE, &in.bulk_handle);
    assert(ret == 0);
    in.op = HG_BULK_PULL;

    start_ts = ABT_get_wtime();
    ret = margo_forward(handle, &in);
    end_ts = ABT_get_wtime();
    assert(ret == 0);

    ret = margo_get_output(handle, &out);
    assert(ret == HG_SUCCESS);

    printf("<op>\t<concurrency>\t<threads>\t<xfer_size>\t<total_bytes>\t<seconds>\t<MiB/s>\n");
    printf("PULL\t%d\t%d\t%d\t%lu\t%f\t%f\n",
        g_opts.concurrency,
        g_opts.threads,
        g_opts.xfer_size,
        out.bytes_moved,
        (end_ts-start_ts),
        ((double)out.bytes_moved/(end_ts-start_ts))/(1024.0*1024.0));

    margo_free_output(handle, &out);

    /* pause a moment */
    margo_thread_sleep(mid, 100);

    in.op = HG_BULK_PUSH;

    start_ts = ABT_get_wtime();
    ret = margo_forward(handle, &in);
    end_ts = ABT_get_wtime();
    assert(ret == 0);

    ret = margo_get_output(handle, &out);
    assert(ret == HG_SUCCESS);

    printf("PUSH\t%d\t%d\t%d\t%lu\t%f\t%f\n",
        g_opts.concurrency,
        g_opts.threads,
        g_opts.xfer_size,
        out.bytes_moved,
        (end_ts-start_ts),
        ((double)out.bytes_moved/(end_ts-start_ts))/(1024.0*1024.0));

    /* calculate how many bytes of the buffer have been transferred */
    bytes_to_check = (g_buffer_size / g_opts.xfer_size) * g_opts.xfer_size;
    if(out.bytes_moved < bytes_to_check)
        bytes_to_check = out.bytes_moved;
    /* check fill pattern we got back; should be what we set plus one */
    for(i=0; i<(bytes_to_check/sizeof(i)); i++)
    {
        assert(((hg_size_t*)g_buffer)[i] == i+1);
    }

    margo_free_output(handle, &out);
    margo_bulk_free(in.bulk_handle);
    margo_destroy(handle);

    return(0);
}

/* function that assists in transferring data until end condition is met */
static void bw_worker(void *_arg)
{
    struct bw_worker_arg *arg = _arg;
    double now;
    size_t my_off;
    int ret;

    // printf("# DBG: worker started.\n");

    now = ABT_get_wtime();

    while((now - arg->start_tm) < g_opts.duration_seconds)
    {
        /* find the offset for this transfer and then increment for next
         * one
         */
        ABT_mutex_spinlock(*arg->cur_off_mutex);
        my_off = *arg->cur_off;
        (*arg->cur_off) += g_opts.xfer_size;
        if(((*arg->cur_off)+g_opts.xfer_size) > g_buffer_size)
            *arg->cur_off = 0;
        ABT_mutex_unlock(*arg->cur_off_mutex);

        ret = margo_bulk_transfer(arg->mid, arg->op,
                *arg->target_addr, *arg->client_bulk_handle, my_off, g_bulk_handle, my_off, g_opts.xfer_size);
        assert(ret == 0);

        arg->bytes_moved += g_opts.xfer_size;
        now = ABT_get_wtime();
    }

    // printf("# DBG: worker stopped.\n");
    return;
}