margo-p2p-latency.c 8.73 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
/*
 * 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 <mercury.h>
#include <abt.h>
#include <ssg.h>
#include <ssg-mpi.h>

struct options
{
    int iterations;
    unsigned int mercury_timeout_client;
    unsigned int mercury_timeout_server;
    char* diag_file_name;
    char* na_transport;
};

29
static int parse_args(int argc, char **argv, struct options *opts);
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53
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, 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;
    int ret;
    ssg_group_id_t gid;
    ssg_member_id_t self;
    int rank;
    double *measurement_array;
    int namelen;
    char processor_name[MPI_MAX_PROCESSOR_NAME];
54
    struct hg_init_info hii;
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

    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);
    MPI_Get_processor_name(processor_name,&namelen);
    printf("Process %d of %d is on %s\n",
	rank, nranks, processor_name);

70 71 72 73 74 75 76
    ret = parse_args(argc, argv, &g_opts);
    if(ret < 0)
    {
        if(rank == 0)
            usage();
        exit(EXIT_FAILURE);
    }
77

78
    memset(&hii, 0, sizeof(hii));
79 80 81 82 83 84 85 86 87 88 89 90 91 92
    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 */
93
    mid = margo_init_opt(g_opts.na_transport, MARGO_SERVER_MODE, &hii, 0, -1);
94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164
    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);

    noop_id = MARGO_REGISTER(
        mid, 
        "noop_rpc", 
        void,
        void,
        noop_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 0
    printf("MPI rank %d has SSG ID %lu\n", rank, self);
#endif

    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, 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);
        sleep(3);
    }

    ssg_group_destroy(gid);
    ssg_finalize();

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

    margo_finalize(mid);
    MPI_Finalize();

    return 0;
}

165
static int parse_args(int argc, char **argv, struct options *opts)
166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184
{
    int opt;
    int ret;

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

    /* 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:i:d:t:")) != -1)
    {
        switch(opt)
        {
            case 'd':
                opts->diag_file_name = strdup(optarg);
                if(!opts->diag_file_name)
                {
                    perror("strdup");
185
                    return(-1);
186 187 188 189 190
                }
                break;
            case 'i':
                ret = sscanf(optarg, "%d", &opts->iterations);
                if(ret != 1)
191
                    return(-1);
192 193 194 195
                break;
            case 't':
                ret = sscanf(optarg, "%u,%u", &opts->mercury_timeout_client, &opts->mercury_timeout_server);
                if(ret != 2)
196
                    return(-1);
197 198 199 200 201 202
                break;
            case 'n':
                opts->na_transport = strdup(optarg);
                if(!opts->na_transport)
                {
                    perror("strdup");
203
                    return(-1);
204 205 206
                }
                break;
            default:
207
                return(-1);
208 209 210 211
        }
    }

    if(opts->iterations < 1 || !opts->na_transport)
212
        return(-1);
213

214
    return(0);
215 216 217 218 219 220 221 222 223
}

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"
224 225
        "\t[-d filename] - enable diagnostics output\n"
        "\t[-t client_progress_timeout,server_progress_timeout] # use \"-t 0,0\" to busy spin\n"
226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
        "\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_respond(handle, NULL);
    margo_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, 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 = margo_create(mid, target_addr, id, &handle);
    assert(ret == 0);

    for(i=0; i<iterations; i++)
    {
        tm1 = ABT_get_wtime();
        ret = margo_forward(handle, NULL);
        tm2 = ABT_get_wtime();
        assert(ret == 0);
        measurement_array[i] = tm2-tm1;
    }

    margo_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);
}