margo-p2p-latency.c 8.7 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
/*
 * Copyright (c) 2017 UChicago Argonne, LLC
 *
 * See COPYRIGHT in top-level directory.
 */

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

#include <mpi.h>

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

struct options
{
    int iterations;
Philip Carns's avatar
Philip Carns committed
24
    char* diag_file_name;
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
    char* na_transport;
};

static void parse_args(int argc, char **argv, struct options *opts);
static void usage(void);
static int run_benchmark(int iterations, hg_id_t id, ssg_member_id_t target, 
    ssg_group_id_t gid, margo_instance_id mid, hg_context_t *hg_context, 
    double *measurement_array);
static void bench_routine_print(const char* op, int size, int iterations, 
    double* measurement_array);
static int measurement_cmp(const void* a, const void *b);
DECLARE_MARGO_RPC_HANDLER(noop_ult);

static hg_id_t noop_id;
static int rpcs_serviced = 0;
static ABT_eventual rpcs_serviced_eventual;
static struct options g_opts;

int main(int argc, char **argv) 
{
    margo_instance_id mid;
    int nranks;
    hg_context_t *hg_context;
    hg_class_t *hg_class;
    int ret;
    ssg_group_id_t gid;
    ssg_member_id_t self;
    int rank;
    hg_bool_t flag;
    double *measurement_array;
55 56
    int namelen;
    char processor_name[MPI_MAX_PROCESSOR_NAME];
57 58 59 60 61 62 63 64 65 66 67 68

    ABT_init(argc, argv);
    MPI_Init(&argc, &argv);

    /* 2 process rtt measurements only */
    MPI_Comm_size(MPI_COMM_WORLD, &nranks);
    if(nranks != 2)
    {
        usage();
        exit(EXIT_FAILURE);
    }
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
69 70 71
    MPI_Get_processor_name(processor_name,&namelen);
    printf("Process %d of %d is on %s\n",
	rank, nranks, processor_name);
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102

    parse_args(argc, argv, &g_opts);

    /* boilerplate HG initialization steps */
    /***************************************/
    hg_class = HG_Init(g_opts.na_transport, HG_TRUE);
    if(!hg_class)
    {
        fprintf(stderr, "Error: HG_Init()\n");
        return(-1);
    }
    hg_context = HG_Context_create(hg_class);
    if(!hg_context)
    {
        fprintf(stderr, "Error: HG_Context_create()\n");
        HG_Finalize(hg_class);
        return(-1);
    }

    /* set primary ES to idle without polling */
    ret = ABT_snoozer_xstream_self_set();
    if(ret != 0)
    {
        fprintf(stderr, "Error: ABT_snoozer_xstream_self_set()\n");
        return(-1);
    }

    /* actually start margo */
    mid = margo_init(0, 0, hg_context);
    assert(mid);

Philip Carns's avatar
Philip Carns committed
103 104 105
    if(g_opts.diag_file_name)
        margo_diag_start(mid);

106
    MARGO_REGISTER_MPLEX(
107 108 109 110
        mid, 
        "noop_rpc", 
        void,
        void,
111
        noop_ult,
112
        MARGO_DEFAULT_MPLEX_ID,
113 114
        NULL, 
        MARGO_RPC_ID_IGNORE);
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157

    /* 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 0
    printf("MPI rank %d has SSG ID %lu\n", rank, self);
#endif

    /* TODO: there should be a cleaner way to get ID from MARGO_REGISTER */
    ret = HG_Registered_name(hg_class, "noop_rpc", &noop_id, &flag);
    assert(ret == 0 && flag);

    if(self == 0)
    {
        /* ssg id 0 runs benchmark */

        measurement_array = calloc(g_opts.iterations, sizeof(*measurement_array));
        assert(measurement_array);

        ret = run_benchmark(g_opts.iterations, noop_id, 1, gid, mid, hg_context, measurement_array);
        assert(ret == 0);

        printf("# <op> <iterations> <size> <min> <q1> <med> <avg> <q3> <max>\n");
        bench_routine_print("noop", 0, g_opts.iterations, measurement_array);
        free(measurement_array);
    }
    else
    {
        /* ssg id 1 acts as server, waiting until iterations have been
         * completed
         */

        ret = ABT_eventual_create(0, &rpcs_serviced_eventual);
        assert(ret == 0);

        ABT_eventual_wait(rpcs_serviced_eventual, NULL);
        assert(rpcs_serviced == g_opts.iterations);
Philip Carns's avatar
Philip Carns committed
158
        sleep(3);
159 160 161 162
    }

    ssg_group_destroy(gid);
    ssg_finalize();
Philip Carns's avatar
Philip Carns committed
163 164 165 166

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

167 168
    margo_finalize(mid);
    HG_Context_destroy(hg_context);
169 170 171 172
/* TODO: debug this further later; HG_Finalize() always hangs at this point
 * when using CCI/verbs right now.
 */
#if 0
173
    HG_Finalize(hg_class);
174
#endif
175 176 177 178 179 180 181 182 183 184 185 186 187
    MPI_Finalize();
    ABT_finalize();

    return 0;
}

static void parse_args(int argc, char **argv, struct options *opts)
{
    int opt;
    int ret;

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

Philip Carns's avatar
Philip Carns committed
188
    while((opt = getopt(argc, argv, "n:i:d:")) != -1)
189 190 191
    {
        switch(opt)
        {
Philip Carns's avatar
Philip Carns committed
192 193 194 195 196 197 198 199
            case 'd':
                opts->diag_file_name = strdup(optarg);
                if(!opts->diag_file_name)
                {
                    perror("strdup");
                    exit(EXIT_FAILURE);
                }
                break;
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
            case 'i':
                ret = sscanf(optarg, "%d", &opts->iterations);
                if(ret != 1)
                {
                    usage();
                    exit(EXIT_FAILURE);
                }
                break;
            case 'n':
                opts->na_transport = strdup(optarg);
                if(!opts->na_transport)
                {
                    perror("strdup");
                    exit(EXIT_FAILURE);
                }
                break;
            default:
                usage();
                exit(EXIT_FAILURE);
        }
    }

    if(opts->iterations < 1 || !opts->na_transport)
    {
        usage();
        exit(EXIT_FAILURE);
    }

    return;
}

static void usage(void)
{
    fprintf(stderr,
        "Usage: "
        "margo-p2p-latency -i <iterations> -n <na>\n"
        "\t-i <iterations> - number of RPC iterations\n"
        "\t-n <na> - na transport\n"
        "\t\texample: mpiexec -n 2 ./margo-p2p-latency -i 10000 -n verbs://\n"
        "\t\t(must be run with exactly 2 processes\n");
    
    return;
}


/* service a remote RPC for a no-op */
static void noop_ult(hg_handle_t handle)
{
    margo_instance_id mid;

250
    mid = margo_hg_handle_get_instance(handle);
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363

    margo_respond(mid, handle, NULL);
    HG_Destroy(handle);

    rpcs_serviced++;
    if(rpcs_serviced == g_opts.iterations)
    {
        ABT_eventual_set(rpcs_serviced_eventual, NULL, 0);
    }

    return;
}
DEFINE_MARGO_RPC_HANDLER(noop_ult)

static int run_benchmark(int iterations, hg_id_t id, ssg_member_id_t target, 
    ssg_group_id_t gid, margo_instance_id mid, hg_context_t *hg_context, 
    double *measurement_array)
{
    hg_handle_t handle;
    hg_addr_t target_addr;
    int i;
    int ret;
    double tm1, tm2;

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

    ret = HG_Create(hg_context, target_addr, id, &handle);
    assert(ret == 0);

    /* TODO: have command line option to toggle whether we reuse one handle
     * or create/release on every cycle
     */
    for(i=0; i<iterations; i++)
    {
        tm1 = ABT_get_wtime();
        ret = margo_forward(mid, handle, NULL);
        tm2 = ABT_get_wtime();
        assert(ret == 0);
        measurement_array[i] = tm2-tm1;
    }

    HG_Destroy(handle);

    return(0);
}

static void bench_routine_print(const char* op, int size, int iterations, double* measurement_array)
{
    double min, max, q1, q3, med, avg, sum;
    int bracket1, bracket2;
    int i;

    qsort(measurement_array, iterations, sizeof(double), measurement_cmp);

    min = measurement_array[0];
    max = measurement_array[iterations-1];

    sum = 0;
    for(i=0; i<iterations; i++)
    {
        sum += measurement_array[i];
    }
    avg = sum/(double)iterations;

    bracket1 = iterations/2;
    if(iterations%2)
        bracket2 = bracket1 + 1;
    else
        bracket2 = bracket1;
    med = (measurement_array[bracket1] + measurement_array[bracket2])/(double)2;

    bracket1 = iterations/4;
    if(iterations%4)
        bracket2 = bracket1 + 1;
    else
        bracket2 = bracket1;
    q1 = (measurement_array[bracket1] + measurement_array[bracket2])/(double)2;

    bracket1 *= 3;
    if(iterations%4)
        bracket2 = bracket1 + 1;
    else
        bracket2 = bracket1;
    q3 = (measurement_array[bracket1] + measurement_array[bracket2])/(double)2;

    printf("%s\t%d\t%d\t%.9f\t%.9f\t%.9f\t%.9f\t%.9f\t%.9f\n", op, iterations, size, min, q1, med, avg, q3, max);
#if 0
    for(i=0; i<iterations; i++)
    {
        printf("\t%.9f", measurement_array[i]);
    }
    printf("\n");
#endif
    fflush(NULL);

    return;
}

static int measurement_cmp(const void* a, const void *b)
{
    const double *d_a = a;
    const double *d_b = b;

    if(*d_a < *d_b)
        return(-1);
    else if(*d_a > *d_b)
        return(1);
    else
        return(0);
}