codes-workload.h 8.47 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

#ifndef CODES_WORKLOAD_H
#define CODES_WORKLOAD_H

15
16
17
#include <ross.h>
#include "configuration.h"

Jonathan Jenkins's avatar
Jonathan Jenkins committed
18
#define MAX_NAME_LENGTH_WKLD 512
19

20
typedef struct iolang_params iolang_params;
21
typedef struct darshan_params darshan_params;
22
typedef struct recorder_params recorder_params;
23

24
25
/* struct to hold the actual data from a single MPI event*/
typedef struct dumpi_trace_params dumpi_trace_params;
26
typedef struct checkpoint_wrkld_params checkpoint_wrkld_params;
27

28
struct iolang_params
29
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
30
    /* the rank count is defined in the workload config file */
31
    int num_cns;
32
33
    /* flag - use path to find kernel files relative to the metafile */
    int use_relpath;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
34
    char io_kernel_meta_path[MAX_NAME_LENGTH_WKLD];
35
    /* set by config in the metadata path */
Jonathan Jenkins's avatar
Jonathan Jenkins committed
36
    char io_kernel_path[MAX_NAME_LENGTH_WKLD];
37
38
};

39
40
41
struct darshan_params
{
    char log_file_path[MAX_NAME_LENGTH_WKLD];
42
    int64_t aggregator_cnt;
43
44
};

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

51
52
struct dumpi_trace_params {
   char file_name[MAX_NAME_LENGTH_WKLD];
53
   int num_net_traces;
54
55
};

56
57
58
59
60
61
62
63
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 */
};
64

65
66
67
68
/* supported I/O operations */
enum codes_workload_op_type
{
    /* terminator; there are no more operations for this rank */
69
    CODES_WK_END = 1,
70
71
72
73
    /* sleep/delay to simulate computation or other activity */
    CODES_WK_DELAY,
    /* block until specified ranks have reached the same point */
    CODES_WK_BARRIER,
74
75

    /* IO operations */
76
77
    /* open */
    CODES_WK_OPEN,
78
    /* close */
79
80
81
82
    CODES_WK_CLOSE,
    /* write */
    CODES_WK_WRITE,
    /* read */
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
    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,
110
    /* Waitall operation */
111
    CODES_WK_WAITALL,
112
    /* Wait operation */
113
    CODES_WK_WAIT,
114
    /* Waitsome operation */
115
    CODES_WK_WAITSOME,
116
    /* Waitany operation */
117
    CODES_WK_WAITANY,
118
    /* Testall operation */
119
    CODES_WK_TESTALL,
Jonathan Jenkins's avatar
Jonathan Jenkins committed
120
121
122
123

    /* for workloads that have events not yet handled
     * (eg the workload language) */
    CODES_WK_IGNORE
124
125
126
127
128
129
130
131
132
};

/* 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
133
    /* what type of operation this is */
134
    enum codes_workload_op_type op_type;
135
136
137
    /* currently only used by network workloads */
    double start_time;
    double end_time;
138
    double sim_start_time;
139

Philip Carns's avatar
Philip Carns committed
140
    /* parameters for each operation type */
141
142
143
144
    union
    {
        struct {
            double seconds;
145
	    double nsecs;
146
147
        } delay;
        struct {
Philip Carns's avatar
Philip Carns committed
148
            int count;  /* num ranks in barrier, -1 means "all" */
149
150
151
            int root;   /* root rank */
        } barrier;
        struct {
152
            uint64_t file_id;      /* integer identifier for the file */
153
154
155
            int create_flag;  /* file must be created, not just opened */
        } open;
        struct {
156
            uint64_t file_id;  /* file to operate on */
157
158
159
160
            off_t offset; /* offset and size */
            size_t size;
        } write;
        struct {
161
            uint64_t file_id;  /* file to operate on */
162
163
164
165
            off_t offset; /* offset and size */
            size_t size;
        } read;
        struct {
166
            uint64_t file_id;  /* file to operate on */
167
        } close;
168
169
170
171
172
173
174
175
        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 */
176
            int16_t req_id;
177
178
179
180
181
182
183
184
185
        } 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 */
186
            int16_t req_id;
187
188
189
190
191
        } recv;
        /* TODO: non-stub for other collectives */
        struct {
            int num_bytes;
        } collective;
192
193
194
195
196
197
198
        struct {
            int count;
            int16_t* req_ids;
        } waits;
        struct {
            int16_t req_id;
        } wait;
199
    }u;
200
201
};

202
203
204
205
206
207
208
209
210
/* read workload configuration from a CODES configuration file and return the
 * workload name and parameters, which can then be passed to
 * codes_workload_load */
typedef struct
{
    char const * type;
    void * params;
} codes_workload_config_return;

211
212
213
// NOTE: some workloads (iolang, checkpoint) require information about the
// total number of ranks to correctly process traces/config files, etc. Other
// workload generators (darshan) ignore it
214
215
codes_workload_config_return codes_workload_read_config(
        ConfigHandle * handle,
216
217
218
        char const * section_name,
        char const * annotation,
        int num_ranks);
219
220
221

void codes_workload_free_config_return(codes_workload_config_return *c);

222
223
/* load and initialize workload of of type "type" with parameters specified by
 * "params".  The rank is the caller's relative rank within the collection
224
225
226
 * 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
227
228
229
230
 *
 * 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.
231
 *
232
233
234
 * Returns and identifier that can be used to retrieve operations later.
 * Returns -1 on failure.
 */
235
236
237
238
239
int codes_workload_load(
        const char* type,
        const char* params,
        int app_id,
        int rank);
240
241
242
243
244

/* 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.
 */
245
246
247
248
249
void codes_workload_get_next(
        int wkld_id,
        int app_id,
        int rank,
        struct codes_workload_op *op);
250
251

/* Reverse of the above function. */
252
253
254
255
256
void codes_workload_get_next_rc(
        int wkld_id,
        int app_id,
        int rank,
        const struct codes_workload_op *op);
257

258
/* Retrieve the number of ranks contained in a workload */
259
260
261
262
int codes_workload_get_rank_cnt(
        const char* type,
        const char* params,
        int app_id);
263

264
/* for debugging/logging: print an individual operation to the specified file */
265
266
267
268
269
void codes_workload_print_op(
        FILE *f,
        struct codes_workload_op *op,
        int app_id,
        int rank);
270

271
272
273
274
275
276
277
278
279
280
281
282
/* 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
283
 *  indent-tabs-mode: nil
284
285
286
287
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */