codes-workload.h 7.93 KB
Newer Older
1 2 3 4 5 6
/*
 * Copyright (C) 2013 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

7
/* I/O workload generator API to be used for reading I/O operations into
Philip Carns's avatar
Philip Carns committed
8 9 10
 * storage system simulations.  This API just describes the operations to be
 * executed; it does not service the operations.
 */
11 12 13 14 15

#ifndef CODES_WORKLOAD_H
#define CODES_WORKLOAD_H

#include "ross.h"
16
#define MAX_NAME_LENGTH_WKLD 512
17

18
typedef struct iolang_params iolang_params;
19
typedef struct darshan_params darshan_params;
20
typedef struct recorder_params recorder_params;
21

22 23 24
/* struct to hold the actual data from a single MPI event*/
typedef struct scala_trace_params scala_trace_params;
typedef struct dumpi_trace_params dumpi_trace_params;
25
typedef struct checkpoint_wrkld_params checkpoint_wrkld_params;
26

27
struct iolang_params
28
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
29
    /* the rank count is defined in the workload config file */
30
    int num_cns;
31 32
    /* flag - use path to find kernel files relative to the metafile */
    int use_relpath;
33
    char io_kernel_meta_path[MAX_NAME_LENGTH_WKLD];
34
    /* set by config in the metadata path */
35
    char io_kernel_path[MAX_NAME_LENGTH_WKLD];
36 37
};

38 39 40
struct darshan_params
{
    char log_file_path[MAX_NAME_LENGTH_WKLD];
41
    int64_t aggregator_cnt;
42 43
};

44 45 46
struct recorder_params
{
    char trace_dir_path[MAX_NAME_LENGTH_WKLD];
47
    int64_t nprocs;
48 49 50
};


51 52 53 54 55 56 57
struct scala_trace_params {
   char offset_file_name[MAX_NAME_LENGTH_WKLD];
   char nw_wrkld_file_name[MAX_NAME_LENGTH_WKLD];
};

struct dumpi_trace_params {
   char file_name[MAX_NAME_LENGTH_WKLD];
58
   int num_net_traces;
59 60
};

61 62 63 64 65 66 67 68
struct checkpoint_wrkld_params
{
    int nprocs; /* number of workload processes */
    double checkpoint_sz; /* size of checkpoint, in TiB */
    double checkpoint_wr_bw; /* checkpoint write b/w, in GiB/s */
    double app_runtime; /* app runtime, in hours */
    double mtti; /* mean time to interrupt, in hours */
};
69

70 71 72 73
/* supported I/O operations */
enum codes_workload_op_type
{
    /* terminator; there are no more operations for this rank */
74
    CODES_WK_END = 1,
75 76 77 78
    /* sleep/delay to simulate computation or other activity */
    CODES_WK_DELAY,
    /* block until specified ranks have reached the same point */
    CODES_WK_BARRIER,
79 80

    /* IO operations */
81 82
    /* open */
    CODES_WK_OPEN,
83
    /* close */
84 85 86 87
    CODES_WK_CLOSE,
    /* write */
    CODES_WK_WRITE,
    /* read */
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114
    CODES_WK_READ,

    /* network operations (modelled after MPI operations) */
    /* blocking send operation */
    CODES_WK_SEND,
    /* blocking recv operation */
    CODES_WK_RECV,
    /* non-blocking send operation */
    CODES_WK_ISEND,
    /* non-blocking receive operation */
    CODES_WK_IRECV,
    /* broadcast operation */
    CODES_WK_BCAST,
    /* Allgather operation */
    CODES_WK_ALLGATHER,
    /* Allgatherv operation */
    CODES_WK_ALLGATHERV,
    /* Alltoall operation */
    CODES_WK_ALLTOALL,
    /* Alltoallv operation */
    CODES_WK_ALLTOALLV,
    /* Reduce operation */
    CODES_WK_REDUCE,
    /* Allreduce operation */
    CODES_WK_ALLREDUCE,
    /* Generic collective operation */
    CODES_WK_COL,
115
    /* Waitall operation */
116
    CODES_WK_WAITALL,
117
    /* Wait operation */
118
    CODES_WK_WAIT,
119
    /* Waitsome operation */
120
    CODES_WK_WAITSOME,
121
    /* Waitany operation */
122
    CODES_WK_WAITANY,
123
    /* Testall operation */
124
    CODES_WK_TESTALL,
Jonathan Jenkins's avatar
Jonathan Jenkins committed
125 126 127 128

    /* for workloads that have events not yet handled
     * (eg the workload language) */
    CODES_WK_IGNORE
129 130 131 132 133 134 135 136 137
};

/* I/O operation paramaters */
struct codes_workload_op
{
    /* TODO: do we need different "classes" of operations to differentiate
     * between different APIs?
     */

Philip Carns's avatar
Philip Carns committed
138
    /* what type of operation this is */
139
    enum codes_workload_op_type op_type;
140 141 142
    /* currently only used by network workloads */
    double start_time;
    double end_time;
143
    double sim_start_time;
144

Philip Carns's avatar
Philip Carns committed
145
    /* parameters for each operation type */
146 147 148 149
    union
    {
        struct {
            double seconds;
150
	    double nsecs;
151 152
        } delay;
        struct {
Philip Carns's avatar
Philip Carns committed
153
            int count;  /* num ranks in barrier, -1 means "all" */
154 155 156
            int root;   /* root rank */
        } barrier;
        struct {
157
            uint64_t file_id;      /* integer identifier for the file */
158 159 160
            int create_flag;  /* file must be created, not just opened */
        } open;
        struct {
161
            uint64_t file_id;  /* file to operate on */
162 163 164 165
            off_t offset; /* offset and size */
            size_t size;
        } write;
        struct {
166
            uint64_t file_id;  /* file to operate on */
167 168 169 170
            off_t offset; /* offset and size */
            size_t size;
        } read;
        struct {
171
            uint64_t file_id;  /* file to operate on */
172
        } close;
173 174 175 176 177 178 179 180
        struct {
            /* TODO: not sure why source rank is here */
            int source_rank;/* source rank of MPI send message */
            int dest_rank; /* dest rank of MPI send message */
            int num_bytes; /* number of bytes to be transferred over the network */
            int data_type; /* MPI data type to be matched with the recv */
            int count; /* number of elements to be received */
            int tag; /* tag of the message */
181
            int16_t req_id;
182 183 184 185 186 187 188 189 190
        } send;
        struct {
            /* TODO: not sure why source rank is here */
            int source_rank;/* source rank of MPI recv message */
            int dest_rank;/* dest rank of MPI recv message */
            int num_bytes; /* number of bytes to be transferred over the network */
            int data_type; /* MPI data type to be matched with the send */
            int count; /* number of elements to be sent */
            int tag; /* tag of the message */
191
            int16_t req_id;
192 193 194 195 196
        } recv;
        /* TODO: non-stub for other collectives */
        struct {
            int num_bytes;
        } collective;
197 198 199 200 201 202 203
        struct {
            int count;
            int16_t* req_ids;
        } waits;
        struct {
            int16_t req_id;
        } wait;
204
    }u;
205 206 207 208
};

/* load and initialize workload of of type "type" with parameters specified by
 * "params".  The rank is the caller's relative rank within the collection
209 210 211
 * of processes that will participate in this workload. The app_id is the
 * "application" that the rank is participating in, used to differentiate
 * between multiple, concurrent workloads
212 213 214 215
 *
 * This function is intended to be called by a compute node LP in a model
 * and may be called multiple times over the course of a
 * simulation in order to execute different application workloads.
216
 *
217 218 219
 * Returns and identifier that can be used to retrieve operations later.
 * Returns -1 on failure.
 */
220 221 222 223 224
int codes_workload_load(
        const char* type,
        const char* params,
        int app_id,
        int rank);
225 226 227 228 229

/* Retrieves the next I/O operation to execute.  the wkld_id is the
 * identifier returned by the init() function.  The op argument is a pointer
 * to a structure to be filled in with I/O operation information.
 */
230 231 232 233 234
void codes_workload_get_next(
        int wkld_id,
        int app_id,
        int rank,
        struct codes_workload_op *op);
235 236

/* Reverse of the above function. */
237 238 239 240 241
void codes_workload_get_next_rc(
        int wkld_id,
        int app_id,
        int rank,
        const struct codes_workload_op *op);
242

243
/* Retrieve the number of ranks contained in a workload */
244 245 246 247
int codes_workload_get_rank_cnt(
        const char* type,
        const char* params,
        int app_id);
248

249
/* for debugging/logging: print an individual operation to the specified file */
250 251 252 253 254
void codes_workload_print_op(
        FILE *f,
        struct codes_workload_op *op,
        int app_id,
        int rank);
255

256 257 258 259 260 261 262 263 264 265 266 267
/* NOTE: there is deliberately no finalize function; we don't have any
 * reliable way to tell when a workload is truly done and will not
 * participate in further reverse computation.   The underlying generators
 * will shut down automatically once they have issued their last event.
 */

#endif /* CODES_WORKLOAD_H */

/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
268
 *  indent-tabs-mode: nil
269 270 271 272
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */