codes-workload.c 10.3 KB
Newer Older
1 2 3 4 5 6
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

7
#include <assert.h>
8 9 10 11 12

#include "ross.h"
#include "codes/codes-workload.h"
#include "codes-workload-method.h"

13
/* list of available methods.  These are statically compiled for now, but we
14 15
 * could make generators optional via autoconf tests etc. if needed
 */
16
extern struct codes_workload_method test_workload_method;
17
extern struct codes_workload_method iolang_workload_method;
18 19 20
#ifdef USE_DUMPI
extern struct codes_workload_method dumpi_trace_workload_method;
#endif
21
#ifdef USE_DARSHAN
22
extern struct codes_workload_method darshan_io_workload_method;
23
#endif
24 25 26
#ifdef USE_RECORDER
extern struct codes_workload_method recorder_io_workload_method;
#endif
27
extern struct codes_workload_method checkpoint_workload_method;
28

29
static struct codes_workload_method *method_array[] =
30 31
{
    &test_workload_method,
32
    &iolang_workload_method,
33 34 35
#ifdef USE_DUMPI
    &dumpi_trace_workload_method,
#endif
36 37
#ifdef USE_DARSHAN
    &darshan_io_workload_method,
38 39 40
#endif
#ifdef USE_RECORDER
    &recorder_io_workload_method,
41
#endif
42
    &checkpoint_workload_method,
43
    NULL};
44

45 46 47 48 49
/* This shim layer is responsible for queueing up reversed operations and
 * re-issuing them so that the underlying workload generator method doesn't
 * have to worry about reverse events.
 *
 * NOTE: we could make this faster with a smarter data structure.  For now
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
 * we just have a linked list of rank_queue structs, one per rank that has
 * opened the workload.  We then have a linked list off of each of those
 * to hold a lifo queue of operations that have been reversed for that rank.
 */

/* holds an operation that has been reversed */
struct rc_op
{
    struct codes_workload_op op;
    struct rc_op* next;
};

/* tracks lifo queue of reversed operations for a given rank */
struct rank_queue
{
65
    int app;
66 67 68 69 70 71 72
    int rank;
    struct rc_op *lifo;
    struct rank_queue *next;
};

static struct rank_queue *ranks = NULL;

73 74 75 76 77
int codes_workload_load(
        const char* type,
        const char* params,
        int app_id,
        int rank)
78
{
79 80
    int i;
    int ret;
81
    struct rank_queue *tmp;
82 83 84 85 86

    for(i=0; method_array[i] != NULL; i++)
    {
        if(strcmp(method_array[i]->method_name, type) == 0)
        {
87
            /* load appropriate workload generator */
88
            ret = method_array[i]->codes_workload_load(params, app_id, rank);
89 90 91 92
            if(ret < 0)
            {
                return(-1);
            }
93 94 95 96 97

            /* are we tracking information for this rank yet? */
            tmp = ranks;
            while(tmp)
            {
98
                if(tmp->rank == rank && tmp->app == app_id)
99 100 101 102 103
                    break;
                tmp = tmp->next;
            }
            if(tmp == NULL)
            {
104
                tmp = (struct rank_queue*)malloc(sizeof(*tmp));
105
                assert(tmp);
106
                tmp->app  = app_id;
107 108 109 110 111 112
                tmp->rank = rank;
                tmp->lifo = NULL;
                tmp->next = ranks;
                ranks = tmp;
            }

113 114 115 116 117
            return(i);
        }
    }

    fprintf(stderr, "Error: failed to find workload generator %s\n", type);
118 119 120
    return(-1);
}

121 122 123 124 125
void codes_workload_get_next(
        int wkld_id,
        int app_id,
        int rank,
        struct codes_workload_op *op)
126
{
127 128 129 130 131 132 133 134 135
    struct rank_queue *tmp;
    struct rc_op *tmp_op;

    /* first look to see if we have a reversed operation that we can
     * re-issue
     */
    tmp = ranks;
    while(tmp)
    {
136
        if(tmp->rank == rank && tmp->app == app_id)
137 138 139
            break;
        tmp = tmp->next;
    }
yangxuserene's avatar
yangxuserene committed
140 141
    if(tmp==NULL)
        printf("tmp is NULL, rank=%d, app_id = %d", rank, app_id);
142 143 144 145
    assert(tmp);
    if(tmp->lifo)
    {
        tmp_op = tmp->lifo;
146
        tmp->lifo = tmp_op->next;
147 148 149 150 151 152 153

        *op = tmp_op->op;
        free(tmp_op);
        return;
    }

    /* ask generator for the next operation */
154
    method_array[wkld_id]->codes_workload_get_next(app_id, rank, op);
155

156 157 158
    return;
}

159 160 161 162 163
void codes_workload_get_next_rc(
        int wkld_id,
        int app_id,
        int rank,
        const struct codes_workload_op *op)
164
{
165 166 167 168 169 170
    struct rank_queue *tmp;
    struct rc_op *tmp_op;

    tmp = ranks;
    while(tmp)
    {
171
        if(tmp->rank == rank && tmp->app == app_id)
172 173 174 175 176
            break;
        tmp = tmp->next;
    }
    assert(tmp);

177
    tmp_op = (struct rc_op*)malloc(sizeof(*tmp_op));
178 179 180 181
    assert(tmp_op);
    tmp_op->op = *op;
    tmp_op->next = tmp->lifo;
    tmp->lifo = tmp_op;
182

183 184 185
    return;
}

186 187 188 189
int codes_workload_get_rank_cnt(
        const char* type,
        const char* params,
        int app_id)
190
{
191 192 193 194 195 196 197
    int i;
    int rank_cnt;

    for(i=0; method_array[i] != NULL; i++)
    {
        if(strcmp(method_array[i]->method_name, type) == 0)
        {
198 199
            rank_cnt =
                method_array[i]->codes_workload_get_rank_cnt(params, app_id);
200 201 202 203 204 205 206
            assert(rank_cnt > 0);
            return(rank_cnt);
        }
    }

    fprintf(stderr, "Error: failed to find workload generator %s\n", type);
    return(-1);
207 208
}

209 210 211 212 213 214
void codes_workload_print_op(
        FILE *f,
        struct codes_workload_op *op,
        int app_id,
        int rank)
{
215 216
    switch(op->op_type){
        case CODES_WK_END:
217
            fprintf(f, "op: app:%d rank:%d type:end\n", app_id, rank);
218 219
            break;
        case CODES_WK_DELAY:
220
            fprintf(f, "op: app:%d rank:%d type:delay seconds:%lf\n",
221
                    app_id, rank, op->u.delay.seconds);
222 223
            break;
        case CODES_WK_BARRIER:
224
            fprintf(f, "op: app:%d rank:%d type:barrier count:%d root:%d\n",
225
                    app_id, rank, op->u.barrier.count, op->u.barrier.root);
226 227
            break;
        case CODES_WK_OPEN:
228
            fprintf(f, "op: app:%d rank:%d type:open file_id:%lu flag:%d\n",
229
                    app_id, rank, op->u.open.file_id, op->u.open.create_flag);
230 231
            break;
        case CODES_WK_CLOSE:
232
            fprintf(f, "op: app:%d rank:%d type:close file_id:%lu\n",
233
                    app_id, rank, op->u.close.file_id);
234 235
            break;
        case CODES_WK_WRITE:
236
            fprintf(f, "op: app:%d rank:%d type:write "
237
                       "file_id:%lu off:%lu size:%lu\n",
238
                    app_id, rank, op->u.write.file_id, op->u.write.offset,
239 240 241
                    op->u.write.size);
            break;
        case CODES_WK_READ:
242
            fprintf(f, "op: app:%d rank:%d type:read "
243
                       "file_id:%lu off:%lu size:%lu\n",
244
                    app_id, rank, op->u.read.file_id, op->u.read.offset,
245 246
                    op->u.read.size);
            break;
247
        case CODES_WK_SEND:
248
            fprintf(f, "op: app:%d rank:%d type:send "
249
                    "src:%d dst:%d bytes:%d type:%d count:%d tag:%d "
250
                    "start:%.5e end:%.5e\n",
251
                    app_id, rank,
252 253
                    op->u.send.source_rank, op->u.send.dest_rank,
                    op->u.send.num_bytes, op->u.send.data_type,
254 255
                    op->u.send.count, op->u.send.tag,
                    op->start_time, op->end_time);
256 257
            break;
        case CODES_WK_RECV:
258
            fprintf(f, "op: app:%d rank:%d type:recv "
259
                    "src:%d dst:%d bytes:%d type:%d count:%d tag:%d\n"
260
                    "start:%.5e end:%.5e\n",
261
                    app_id, rank,
262 263
                    op->u.recv.source_rank, op->u.recv.dest_rank,
                    op->u.recv.num_bytes, op->u.recv.data_type,
264 265
                    op->u.recv.count, op->u.recv.tag,
                    op->start_time, op->end_time);
266 267
            break;
        case CODES_WK_ISEND:
268
            fprintf(f, "op: app:%d rank:%d type:isend "
269
                    "src:%d dst:%d bytes:%d type:%d count:%d tag:%d\n"
270
                    "start:%.5e end:%.5e\n",
271
                    app_id, rank,
272 273
                    op->u.send.source_rank, op->u.send.dest_rank,
                    op->u.send.num_bytes, op->u.send.data_type,
274 275
                    op->u.send.count, op->u.send.tag,
                    op->start_time, op->end_time);
276 277
            break;
        case CODES_WK_IRECV:
278
            fprintf(f, "op: app:%d rank:%d type:irecv "
279
                    "src:%d dst:%d bytes:%d type:%d count:%d tag:%d\n"
280
                    "start:%.5e end:%.5e\n",
281
                    app_id, rank,
282 283
                    op->u.recv.source_rank, op->u.recv.dest_rank,
                    op->u.recv.num_bytes, op->u.recv.data_type,
284 285
                    op->u.recv.count, op->u.recv.tag,
                    op->start_time, op->end_time);
286
            break;
287 288
#define PRINT_COL(_type_str) \
            fprintf(f, "op: app:%d rank:%d type:%s" \
289
                    " bytes:%d, start:%.5e, end:%.5e\n", app_id, rank, \
290 291
                    _type_str, op->u.collective.num_bytes, op->start_time, \
                    op->end_time)
292
        case CODES_WK_BCAST:
293
            PRINT_COL("bcast");
294 295
            break;
        case CODES_WK_ALLGATHER:
296
            PRINT_COL("allgather");
297 298
            break;
        case CODES_WK_ALLGATHERV:
299
            PRINT_COL("allgatherv");
300 301
            break;
        case CODES_WK_ALLTOALL:
302
            PRINT_COL("alltoall");
303 304
            break;
        case CODES_WK_ALLTOALLV:
305
            PRINT_COL("alltoallv");
306 307
            break;
        case CODES_WK_REDUCE:
308
            PRINT_COL("reduce");
309 310
            break;
        case CODES_WK_ALLREDUCE:
311
            PRINT_COL("allreduce");
312 313
            break;
        case CODES_WK_COL:
314
            PRINT_COL("collective");
315
            break;
316 317 318
#undef PRINT_COL
#define PRINT_WAIT(_type_str, _ct) \
            fprintf(f, "op: app:%d rank:%d type:%s" \
319
                    "num reqs:%d, start:%.5e, end:%.5e\n", \
320
                    app_id, rank, _type_str, _ct, op->start_time, op->end_time)
321
        case CODES_WK_WAITALL:
322
            PRINT_WAIT("waitall", op->u.waits.count);
323 324
            break;
        case CODES_WK_WAIT:
325
            PRINT_WAIT("wait", 1);
326 327
            break;
        case CODES_WK_WAITSOME:
328
            PRINT_WAIT("waitsome", op->u.waits.count);
329 330
            break;
        case CODES_WK_WAITANY:
331
            PRINT_WAIT("waitany", op->u.waits.count);
332
            break;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
333 334
        case CODES_WK_IGNORE:
            break;
335
        default:
Jonathan Jenkins's avatar
Jonathan Jenkins committed
336 337 338
            fprintf(stderr,
                    "%s:%d: codes_workload_print_op: unrecognized workload type "
                    "(op code %d)\n", __FILE__, __LINE__, op->op_type);
339 340 341
    }
}

342 343 344 345
/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
346
 *  indent-tabs-mode: nil
347 348 349 350
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */