model-net.h 14.2 KB
Newer Older
1
/*
Philip Carns's avatar
Philip Carns committed
2
 * Copyright (C) 2013 University of Chicago.
3
 * See COPYRIGHT notice in top-level directory.
Philip Carns's avatar
Philip Carns committed
4
 *
5 6 7 8 9
 */

#ifndef MODELNET_H
#define MODELNET_H

10 11 12 13 14
#include <ross.h>
#include <codes/lp-type-lookup.h>
#include <codes/configuration.h>
#include <codes/lp-io.h>
#include <codes/codes-mapping-context.h>
15
#include <stdint.h>
16

17 18
#define PULL_MSG_SIZE 128

19
#define MAX_NAME_LENGTH 256
20 21
#define CATEGORY_NAME_MAX 16
#define CATEGORY_MAX 12
22

23
// simple deprecation attribute hacking
24 25 26 27 28 29
#if !defined(DEPRECATED)
#  if defined(__GNUC__) || defined(__GNUG__) || defined(__clang__)
#    define DEPRECATED __attribute__((deprecated))
#  else
#    define DEPRECATED
#  endif
30 31
#endif

32 33 34 35 36
/* HACK: there is currently no scheduling fidelity across multiple
 * model_net_event calls. Hence, problems arise when some LP sends multiple
 * messages as part of an event and expects FCFS ordering. A proper fix which
 * involves model-net LP-level scheduling of requests is ideal, but not 
 * feasible for now (would basically have to redesign model-net), so expose
37 38
 * explicit start-sequence and stop-sequence markers as a workaround
 */
39
extern int mn_in_sequence;
40 41
extern tw_stime mn_msg_offset;
#define MN_START_SEQ() do {\
42
    mn_in_sequence = 1; \
43 44 45
    mn_msg_offset = 0.0; \
} while (0)
#define MN_END_SEQ() do {\
46
    mn_in_sequence = 0;\
47 48 49
} while (0)


50
typedef struct mn_stats mn_stats;
51

52 53 54 55 56
// use the X-macro to get types and names rolled up into one structure
// format: { enum vals, config name, internal lp name, lp method struct}
// last value is sentinel
#define NETWORK_DEF \
    X(SIMPLENET, "modelnet_simplenet", "simplenet", &simplenet_method)\
Jonathan Jenkins's avatar
Jonathan Jenkins committed
57
    X(SIMPLEP2P, "modelnet_simplep2p", "simplep2p", &simplep2p_method)\
58 59 60 61 62 63
    X(TORUS,     "modelnet_torus",     "torus",     &torus_method)\
    X(DRAGONFLY, "modelnet_dragonfly", "dragonfly", &dragonfly_method)\
    X(LOGGP,     "modelnet_loggp",     "loggp",     &loggp_method)\
    X(MAX_NETS,  NULL,                 NULL,        NULL)

#define X(a,b,c,d) a,
64 65
enum NETWORKS
{
66
    NETWORK_DEF
67
};
68 69
#undef X

70 71
// message parameter types
enum msg_param_type {
72
    // parameters for modelnet-scheduler (priorities and such)
73
    MN_MSG_PARAM_SCHED,
74 75 76
    // parameter allowing the explicit setting of a messages start time (useful
    // for routing scenarios in which a single message is implemented in terms
    // of multiple model-net events, such as for model_net_pull_*event's)
77
    MN_MSG_PARAM_START_TIME,
78 79 80
    MAX_MN_MSG_PARAM_TYPES
};

81 82 83 84 85
// MN_MSG_PARAM_MSG_START_TIME parameter types (only one for the time being)
enum msg_param_start_time {
    MN_MSG_PARAM_START_TIME_VAL
};

86 87 88 89 90
// return type for model_net_*event calls, to be passed into RC
// currently is just an int, but in the future may indicate whether lp-io was
// called etc.
typedef int model_net_event_return;

91 92 93 94
// network identifiers (both the config lp names and the model-net internal
// names)
extern char * model_net_lp_config_names[];
extern char * model_net_method_names[];
95

96 97 98 99
// request structure that gets passed around (by the model-net implementation,
// not the user)
typedef struct model_net_request {
    tw_lpid  final_dest_lp;
100
    tw_lpid  dest_mn_lp; // destination modelnet lp
101
    tw_lpid  src_lp;
102 103
    // time the source event was called
    tw_stime msg_start_time;
104
    uint64_t msg_size;
105
    uint64_t pull_size;
106
    uint64_t packet_size;
107 108 109
    // unique message id set, *at the modelnet LP*, for each message the
    // modelnet LP processes
    uint64_t msg_id;
110 111 112 113 114 115 116
    int      net_id;
    int      is_pull;
    int      remote_event_size;
    int      self_event_size;
    char     category[CATEGORY_NAME_MAX];
} model_net_request;

117 118 119 120 121 122 123 124 125 126 127 128 129
/* data structure for tracking network statistics */
struct mn_stats
{
    char category[CATEGORY_NAME_MAX];
    long send_count;
    long send_bytes;
    tw_stime send_time;
    long recv_count;
    long recv_bytes;
    tw_stime recv_time;
    long max_event_size;
};

130 131 132
/* Registers all model-net LPs in ROSS. Should be called after 
 * configuration_load, but before codes_mapping_setup */
void model_net_register();
133

134 135 136 137 138 139 140 141
/* Configures all model-net LPs based on the CODES configuration, and returns
 * ids to address the different types by.
 *
 * id_count - the output number of networks
 *
 * return - the set of network IDs, indexed in the order given by the
 * modelnet_order configuration parameter */
int* model_net_configure(int *id_count);
142

143 144 145
/* Sets up a sampling loop for model net events. The sampling data provided by
 * each modelnet lp is model-defined. This is a PE-wide setting. Data is sent
 * to LP-IO with the category modelnet-samples */
146
void model_net_enable_sampling(tw_stime interval, tw_stime end);
147 148 149 150

/* Returns 1 if modelnet is performing sampling, 0 otherwise */
int model_net_sampling_enabled(void);

151 152 153 154 155
/* Initialize/configure the network(s) based on the CODES configuration.
 * returns an array of the network ids, indexed in the order given by the 
 * modelnet_order configuration parameter 
 * OUTPUT id_count - the output number of networks */
int* model_net_set_params(int *id_count);
156

157 158
/* utility function to get the modelnet ID post-setup */
int model_net_get_id(char *net_name);
159 160 161 162

/* This event does a collective operation call for model-net */
void model_net_event_collective(
    int net_id,
163
    char const * category,
164 165
    int message_size,
    int remote_event_size,
166
    void const * remote_event,
167 168 169 170 171 172 173 174
    tw_lp *sender);

/* reverse event of the collective operation call */
void model_net_event_collective_rc(
        int net_id,
        int message_size,
        tw_lp *sender);

175 176 177
/* allocate and transmit a new event that will pass through model_net to 
 * arrive at its destination:
 *
178 179
 * - net_id: the type of network to send this message through. The set of
 *   net_id's is given by model_net_configure.
180 181 182 183 184 185 186 187 188 189 190 191
 * - category: category name to associate with this communication
 *   - OPTIONAL: callers can set this to NULL if they don't want to use it,
 *     and model_net methods can ignore it if they don't support it
 * - final_dest_lp: the LP that the message should be delivered to.
 *   - NOTE: this is _not_ the LP of an underlying network method (for
 *     example, it is not a torus or dragonfly LP), but rather the LP of an
 *     MPI process or storage server that you are transmitting to.
 * - message_size: this is the size of the message (in bytes) that modelnet
 *     will simulate transmitting to the final_dest_lp.  It can be any size
 *     (i.e. it is not constrained by transport packet size).
 * - remote_event_size: this is the size of the ROSS event structure that
 *     will be delivered to the final_dest_lp.
192 193 194
 * - remote_event: pointer to data to be used as the remote event message. When
 *   the message payload (of size message_size) has been fully delivered to the
 *   destination (given by final_dest_lp), this event will be issued.
195 196 197 198 199 200 201
 * - self_event_size: this is the size of the ROSS event structure that will
 *     be delivered to the calling LP once local completion has occurred for
 *     the network transmission.
 *     - NOTE: "local completion" in this sense means that model_net has
 *       transmitted the data off of the local node, but it does not mean that
 *       the data has been (or even will be) delivered.  Once this event is
 *       delivered the caller is free to re-use its buffer.
202 203 204 205
 * - self_event: pointer to data to be used as the self event message. When the
 *   message payload (of size message_size) has been fully sent from the
 *   sender's point of view (e.g. the corresponding NIC has sent out all
 *   packets for this message), the event will be issued to the sender.
206 207
 * - sender: pointer to the tw_lp structure of the API caller.  This is
 *     identical to the sender argument to tw_event_new().
208
 *
209 210 211 212
 * The modelnet LP used for communication is determined by the default CODES
 * map context (see codes-base, codes/codes-mapping-context.h), using net_id
 * to differentiate different model types. Note that the map context is used
 * when calculating *both* sender and receiver modelnet LPs
213 214
 */
// first argument becomes the network ID
215
model_net_event_return model_net_event(
216
    int net_id,
217
    char const * category, 
218
    tw_lpid final_dest_lp, 
219
    uint64_t message_size, 
220
    tw_stime offset,
221
    int remote_event_size,
222
    void const * remote_event,
223
    int self_event_size,
224
    void const * self_event,
225
    tw_lp *sender);
226 227 228 229 230
/*
 * See model_net_event for a general description.
 *
 * Unlike model_net_event, this function uses the annotation to differentiate
 * multiple modelnet LPs with the same type but different annotation. The caller
231 232
 * annotation is not consulted here. The corresponding CODES map context is
 * CODES_MCTX_GROUP_MODULO with the supplied annotation arguments.
233
 */
234
model_net_event_return model_net_event_annotated(
235
        int net_id,
236 237
        char const * annotation,
        char const * category, 
238 239 240 241
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        tw_stime offset,
        int remote_event_size,
242
        void const * remote_event,
243
        int self_event_size,
244
        void const * self_event,
245
        tw_lp *sender);
246 247 248 249 250 251
/*
 * See model_net_event for a general description.
 *
 * This variant uses CODES map contexts to calculate the sender and receiver
 * modelnet LPs
 */
252
model_net_event_return model_net_event_mctx(
253 254 255 256 257 258 259 260 261 262 263 264 265
        int net_id,
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
        char const * category, 
        tw_lpid final_dest_lp, 
        uint64_t message_size, 
        tw_stime offset,
        int remote_event_size,
        void const * remote_event,
        int self_event_size,
        void const * self_event,
        tw_lp *sender);

266

267 268
/* model_net_find_local_device()
 *
269 270 271
 * returns the LP id of the network card attached to the calling LP using the
 * default CODES mapping context if ignore_annotations is true, and
 * CODES_MCTX_GROUP_MODULO with the supplied annotation parameters otherwise
272
 */
273 274 275 276
tw_lpid model_net_find_local_device(
        int          net_id,
        const char * annotation,
        int          ignore_annotations,
277
        tw_lpid      sender_gid);
278

279 280 281 282 283 284
/* same as ^, except use the supplied mapping context */
tw_lpid model_net_find_local_device_mctx(
        int net_id,
        struct codes_mctx const * map_ctx,
        tw_lpid sender_gid);

285 286
int model_net_get_msg_sz(int net_id);

287 288 289 290 291 292 293 294 295 296
/* model_net_event_rc()
 *
 * This function does reverse computation for the model_net_event_new()
 * function.
 * - sender: pointer to the tw_lp structure of the API caller.  This is
 *   identical to the sender argument to tw_event_new().
 */
/* NOTE: we may end up needing additoinal arguments here to track state for
 * reverse computation; add as needed 
 */
297
DEPRECATED
298 299 300
void model_net_event_rc(
    int net_id,
    tw_lp *sender,
301
    uint64_t message_size);
302

303 304 305 306 307 308
/* This function replaces model_net_event_rc, and will replace the name later
 * on. The num_rng_calls argument is the return value of model_net_*event*
 * calls */
void model_net_event_rc2(
        tw_lp *sender,
        model_net_event_return const * ret);
309 310 311 312 313 314 315 316 317 318 319 320 321

/* Issue a 'pull' from the memory of the destination LP, without
 * requiring the destination LP to do event processing. This is meant as a
 * simulation-based abstraction of RDMA. A control packet will be sent to the
 * destination LP, the payload will be sent back to the requesting LP, and the
 * requesting LP will be issued it's given completion event.
 *
 * Parameters are largely the same as model_net_event, with the following
 * exceptions:
 * - final_dest_lp is the lp to pull data from
 * - self_event_size, self_event are applied at the requester upon receipt of 
 *   the payload from the dest
 */
322
model_net_event_return model_net_pull_event(
323
        int net_id,
324
        char const *category,
325 326 327 328
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
329
        void const *self_event,
330
        tw_lp *sender);
331
model_net_event_return model_net_pull_event_annotated(
332
        int net_id,
333 334
        char const * annotation,
        char const *category,
335 336 337 338
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
339
        void const *self_event,
340
        tw_lp *sender);
341
model_net_event_return model_net_pull_event_mctx(
342
        int net_id,
343 344
        struct codes_mctx const * send_map_ctx,
        struct codes_mctx const * recv_map_ctx,
345 346 347 348 349 350 351
        char const *category,
        tw_lpid final_dest_lp,
        uint64_t message_size,
        tw_stime offset,
        int self_event_size,
        void const *self_event,
        tw_lp *sender);
352

353
DEPRECATED
354 355 356 357
void model_net_pull_event_rc(
        int net_id,
        tw_lp *sender);

358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379
/*
 * Set message-specific parameters
 * type     - overall type (see msg_param_type)
 * sub_type - type of parameter specific to type. This is intended to be
 *            an enum for each of msg_param_type's values
 * params   - the parameter payload
 *
 * This function works by setting up a temporary parameter context within the
 * model-net implementation (currently implemented as a set of translation-unit
 * globals). Upon a subsequent model_net_*event* call, the context is consumed
 * and reset to an unused state.
 * 
 * NOTE: this call MUST be placed in the same calling context as the subsequent
 * model_net_*event* call. Otherwise, the parameters are not guaranteed to work
 * on the intended event, and may possibly be consumed by another, unrelated
 * event.
 */
void model_net_set_msg_param(
        enum msg_param_type type,
        int sub_type,
        const void * params);

380 381 382
/* returns pointer to LP information for simplenet module */
const tw_lptype* model_net_get_lp_type(int net_id);

383
DEPRECATED
384
uint64_t model_net_get_packet_size(int net_id);
385

386 387 388
/* used for reporting overall network statistics for e.g. average latency ,
 * maximum latency, total number of packets finished during the entire
 * simulation etc. */
389
void model_net_report_stats(int net_id);
390

391
/* writing model-net statistics on a per LP basis */
392 393
void model_net_write_stats(tw_lpid lpid, mn_stats* stat);

394
/* printing model-net statistics on a per LP basis */
395 396 397
void model_net_print_stats(tw_lpid lpid, mn_stats mn_stats_array[]);

/* find model-net statistics */
398
mn_stats* model_net_find_stats(char const * category, mn_stats mn_stats_array[]);
399 400 401 402 403 404 405 406 407 408
#endif /* MODELNET_H */

/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ft=c ts=8 sts=4 sw=4 expandtab
 */