OpenSHMEM Bundles

bundle_example.c

/**********************************************************
  Bundle API summary:
  
  int shmem_bundle_start(IN shmem_bundle_config_t config,
                         IN long config_mask,
                         OUT shmem_bundle_t *bundle);
  void shmem_bundle_stop(IN shmem_bundle_t bundle);

  typedef struct shmem_bundle {
      shmem_ctx_t ctx;
  } shmem_bundle_config_t;

***********************************************************/


shmem_ctx_t my_ctx;
shmem_bundle_t my_bundle;
shmem_bundle_config_t my_bundle_config;

/* Starting a bundle is a hint...
   all RMA/AMO semantics are preserved */
int ret = shmem_bundle_start(my_bundle_config, 0, &my_bundle);
if (ret < 0) {
  /* User knows not to expect an optimization */
}

/* Any sequence of RMA/AMO operations on my_ctx could be bundled: */
for (int i = 0; i < num_ops; i++) {
  
  /* blocking RMA would *not* be bundled by the implementation */
  shmem_put(my_ctx, ...);
  shmem_atomic_add(my_ctx, ...);
  
  /* non-blocking routines could be chained
     (and maybe aggregated/coalesced?) */
  shmem_put_nbi(my_ctx, ...);
  shmem_atomic_fetch_add_nbi(my_ctx, ...);
}

/* Stoping a bundle would not affect memory ordering,
   just informs the runtime a bundle is done */
shmem_bundle_stop(my_bundle);

/* Can nbi source buffers be reused at this point?
   Likely not if "stop" has no affect on memory model...
   Maybe a shmem_local_complete() could be useful here? */

/* Quiet required on context for remote completion */
shmem_ctx_quiet(my_ctx);