1
1

Merge pull request #5755 from gpaulsen/osc_rdma_cleanup

osc/rdma: clean out stale aggregation code
Этот коммит содержится в:
Geoff Paulsen 2018-09-22 15:00:21 -05:00 коммит произвёл GitHub
родитель 4462396df3 56e31f8206
Коммит bc798b6135
Не найден ключ, соответствующий данной подписи
Идентификатор ключа GPG: 4AEE18F83AFDEB23
7 изменённых файлов: 2 добавлений и 301 удалений

Просмотреть файл

@ -77,9 +77,6 @@ struct ompi_osc_rdma_component_t {
/** RDMA component buffer size */
unsigned int buffer_size;
/** aggregation limit */
unsigned int aggregation_limit;
/** List of requests that need to be freed */
opal_list_t request_gc;
@ -104,9 +101,6 @@ struct ompi_osc_rdma_component_t {
/** Priority of the osc/rdma component */
unsigned int priority;
/** aggregation free list */
opal_free_list_t aggregate;
/** directory where to place backing files */
char *backing_directory;
};
@ -569,16 +563,6 @@ static inline void ompi_osc_rdma_sync_rdma_dec (ompi_osc_rdma_sync_t *rdma_sync)
*/
static inline void ompi_osc_rdma_sync_rdma_complete (ompi_osc_rdma_sync_t *sync)
{
if (opal_list_get_size (&sync->aggregations)) {
ompi_osc_rdma_aggregation_t *aggregation, *next;
OPAL_THREAD_SCOPED_LOCK(&sync->lock,
OPAL_LIST_FOREACH_SAFE(aggregation, next, &sync->aggregations, ompi_osc_rdma_aggregation_t) {
fprintf (stderr, "Flushing aggregation %p, peer %p\n", (void*)aggregation, (void*)aggregation->peer);
ompi_osc_rdma_peer_aggregate_flush (aggregation->peer);
});
}
#if !defined(BTL_VERSION) || (BTL_VERSION < 310)
do {
opal_progress ();
@ -611,16 +595,6 @@ static inline bool ompi_osc_rdma_access_epoch_active (ompi_osc_rdma_module_t *mo
return (module->all_sync.epoch_active || ompi_osc_rdma_in_passive_epoch (module));
}
static inline void ompi_osc_rdma_aggregation_return (ompi_osc_rdma_aggregation_t *aggregation)
{
if (aggregation->sync) {
opal_list_remove_item (&aggregation->sync->aggregations, (opal_list_item_t *) aggregation);
}
opal_free_list_return(&mca_osc_rdma_component.aggregate, (opal_free_list_item_t *) aggregation);
}
__opal_attribute_always_inline__
static inline bool ompi_osc_rdma_oor (int rc)
{

Просмотреть файл

@ -434,28 +434,6 @@ static void ompi_osc_rdma_put_complete_flush (struct mca_btl_base_module_t *btl,
}
}
static void ompi_osc_rdma_aggregate_put_complete (struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint,
void *local_address, mca_btl_base_registration_handle_t *local_handle,
void *context, void *data, int status)
{
ompi_osc_rdma_aggregation_t *aggregation = (ompi_osc_rdma_aggregation_t *) context;
ompi_osc_rdma_sync_t *sync = aggregation->sync;
ompi_osc_rdma_frag_t *frag = aggregation->frag;
assert (OPAL_SUCCESS == status);
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "aggregate put complete %p on sync %p. local address %p. status %d",
(void *) aggregation, (void *) sync, local_address, status);
ompi_osc_rdma_frag_complete (frag);
ompi_osc_rdma_aggregation_return (aggregation);
/* make sure the aggregation is returned before marking the operation as complete */
opal_atomic_wmb ();
ompi_osc_rdma_sync_rdma_dec (sync);
}
static int ompi_osc_rdma_put_real (ompi_osc_rdma_sync_t *sync, ompi_osc_rdma_peer_t *peer, uint64_t target_address,
mca_btl_base_registration_handle_t *target_handle, void *ptr,
mca_btl_base_registration_handle_t *local_handle, size_t size,
@ -492,75 +470,11 @@ static int ompi_osc_rdma_put_real (ompi_osc_rdma_sync_t *sync, ompi_osc_rdma_pee
return ret;
}
#if 0
static void ompi_osc_rdma_aggregate_append (ompi_osc_rdma_aggregation_t *aggregation, ompi_osc_rdma_request_t *request,
void *source_buffer, size_t size)
{
size_t offset = aggregation->buffer_used;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "appending %lu bytes of data from %p to aggregate fragment %p with start "
"address 0x%lx", (unsigned long) size, source_buffer, (void *) aggregation,
(unsigned long) aggregation->target_address);
memcpy (aggregation->buffer + offset, source_buffer, size);
aggregation->buffer_used += size;
if (request) {
/* the local buffer is now available */
ompi_osc_rdma_request_complete (request, 0);
}
}
static int ompi_osc_rdma_aggregate_alloc (ompi_osc_rdma_sync_t *sync, ompi_osc_rdma_peer_t *peer, uint64_t target_address,
mca_btl_base_registration_handle_t *target_handle, void *source_buffer, size_t size,
ompi_osc_rdma_request_t *request, int type)
{
ompi_osc_rdma_module_t *module = sync->module;
ompi_osc_rdma_aggregation_t *aggregation;
int ret;
aggregation = (ompi_osc_rdma_aggregation_t *) opal_free_list_get (&mca_osc_rdma_component.aggregate);
if (OPAL_UNLIKELY(NULL == aggregation)) {
return OPAL_ERR_OUT_OF_RESOURCE;
}
ret = ompi_osc_rdma_frag_alloc (module, mca_osc_rdma_component.aggregation_limit, &aggregation->frag,
&aggregation->buffer);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
opal_free_list_return(&mca_osc_rdma_component.aggregate, (opal_free_list_item_t *) aggregation);
return ret;
}
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "allocated new aggregate fragment %p for target %d", (void *) aggregation,
peer->rank);
peer->aggregate = aggregation;
aggregation->target_address = target_address;
aggregation->target_handle = target_handle;
aggregation->buffer_size = mca_osc_rdma_component.aggregation_limit;
aggregation->sync = sync;
aggregation->peer = peer;
aggregation->type = type;
aggregation->buffer_used = 0;
ompi_osc_rdma_aggregate_append (aggregation, request, source_buffer, size);
opal_list_append (&sync->aggregations, (opal_list_item_t *) aggregation);
return OMPI_SUCCESS;
}
#endif
int ompi_osc_rdma_put_contig (ompi_osc_rdma_sync_t *sync, ompi_osc_rdma_peer_t *peer, uint64_t target_address,
mca_btl_base_registration_handle_t *target_handle, void *source_buffer, size_t size,
ompi_osc_rdma_request_t *request)
{
ompi_osc_rdma_module_t *module = sync->module;
#if 0
ompi_osc_rdma_aggregation_t *aggregation = peer->aggregate;
#endif
mca_btl_base_registration_handle_t *local_handle = NULL;
mca_btl_base_rdma_completion_fn_t cbfunc = NULL;
ompi_osc_rdma_frag_t *frag = NULL;
@ -568,34 +482,6 @@ int ompi_osc_rdma_put_contig (ompi_osc_rdma_sync_t *sync, ompi_osc_rdma_peer_t *
void *cbcontext;
int ret;
#if 0
if (aggregation) {
if (size <= (aggregation->buffer_size - aggregation->buffer_used) && (target_handle == aggregation->target_handle) &&
(target_address == aggregation->target_address + aggregation->buffer_used)) {
assert (OMPI_OSC_RDMA_TYPE_PUT == aggregation->type);
ompi_osc_rdma_aggregate_append (aggregation, request, source_buffer, size);
return OMPI_SUCCESS;
}
/* can't aggregate this operation. flush the previous segment */
ret = ompi_osc_rdma_peer_aggregate_flush (peer);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
}
if (size <= (mca_osc_rdma_component.aggregation_limit >> 2)) {
ret = ompi_osc_rdma_aggregate_alloc (sync, peer, target_address, target_handle, source_buffer, size, request,
OMPI_OSC_RDMA_TYPE_PUT);
if (OPAL_LIKELY(OMPI_SUCCESS == ret)) {
if (request) {
}
return ret;
}
}
#endif
if (module->selected_btl->btl_register_mem && size > module->selected_btl->btl_put_local_registration_threshold) {
ret = ompi_osc_rdma_frag_alloc (module, size, &frag, &ptr);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
@ -680,37 +566,6 @@ static void ompi_osc_rdma_get_complete (struct mca_btl_base_module_t *btl, struc
ompi_osc_rdma_request_complete (request, status);
}
int ompi_osc_rdma_peer_aggregate_flush (ompi_osc_rdma_peer_t *peer)
{
ompi_osc_rdma_aggregation_t *aggregation = peer->aggregate;
int ret;
if (NULL == aggregation) {
return OMPI_SUCCESS;
}
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "flusing aggregate fragment %p", (void *) aggregation);
assert (OMPI_OSC_RDMA_TYPE_PUT == aggregation->type);
ret = ompi_osc_rdma_put_real (aggregation->sync, peer, aggregation->target_address, aggregation->target_handle,
aggregation->buffer, aggregation->frag->handle, aggregation->buffer_used,
ompi_osc_rdma_aggregate_put_complete, (void *) aggregation, NULL);
peer->aggregate = NULL;
if (OPAL_UNLIKELY(OMPI_SUCCESS == ret)) {
return OMPI_SUCCESS;
}
ompi_osc_rdma_cleanup_rdma (aggregation->sync, false, aggregation->frag, NULL, NULL);
ompi_osc_rdma_aggregation_return (aggregation);
return ret;
}
static int ompi_osc_rdma_get_partial (ompi_osc_rdma_sync_t *sync, ompi_osc_rdma_peer_t *peer, uint64_t source_address,
mca_btl_base_registration_handle_t *source_handle, void *target_buffer, size_t size,
ompi_osc_rdma_request_t *request) {

Просмотреть файл

@ -223,16 +223,6 @@ static int ompi_osc_rdma_component_register (void)
MCA_BASE_VAR_SCOPE_GROUP, &mca_osc_rdma_component.max_attach);
free(description_str);
mca_osc_rdma_component.aggregation_limit = 1024;
asprintf(&description_str, "Maximum size of an aggregated put/get. Messages are aggregated for consecutive"
"put and get operations. In some cases this may lead to higher latency but "
"should also lead to higher bandwidth utilization. Set to 0 to disable (default: %d)",
mca_osc_rdma_component.aggregation_limit);
(void) mca_base_component_var_register (&mca_osc_rdma_component.super.osc_version, "aggregation_limit",
description_str, MCA_BASE_VAR_TYPE_UNSIGNED_INT, NULL, 0, 0, OPAL_INFO_LVL_3,
MCA_BASE_VAR_SCOPE_GROUP, &mca_osc_rdma_component.aggregation_limit);
free(description_str);
mca_osc_rdma_component.priority = 101;
asprintf(&description_str, "Priority of the osc/rdma component (default: %d)",
mca_osc_rdma_component.priority);
@ -336,24 +326,6 @@ static int ompi_osc_rdma_component_init (bool enable_progress_threads,
__FILE__, __LINE__, ret);
}
OBJ_CONSTRUCT(&mca_osc_rdma_component.aggregate, opal_free_list_t);
if (!enable_mpi_threads && mca_osc_rdma_component.aggregation_limit) {
ret = opal_free_list_init (&mca_osc_rdma_component.aggregate,
sizeof(ompi_osc_rdma_aggregation_t), 8,
OBJ_CLASS(ompi_osc_rdma_aggregation_t), 0, 0,
32, 128, 32, NULL, 0, NULL, NULL, NULL);
if (OPAL_SUCCESS != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: opal_free_list_init failed: %d\n",
__FILE__, __LINE__, ret);
}
} else {
/* only enable put aggregation when not using threads */
mca_osc_rdma_component.aggregation_limit = 0;
}
return ret;
}
@ -373,7 +345,6 @@ int ompi_osc_rdma_component_finalize (void)
OBJ_DESTRUCT(&mca_osc_rdma_component.requests);
OBJ_DESTRUCT(&mca_osc_rdma_component.request_gc);
OBJ_DESTRUCT(&mca_osc_rdma_component.buffer_gc);
OBJ_DESTRUCT(&mca_osc_rdma_component.aggregate);
return OMPI_SUCCESS;
}
@ -1355,53 +1326,3 @@ static char* ompi_osc_rdma_set_no_lock_info(opal_infosubscriber_t *obj, char *ke
*/
return module->no_locks ? "true" : "false";
}
#if 0 // stale code?
static int ompi_osc_rdma_set_info (struct ompi_win_t *win, struct opal_info_t *info)
{
ompi_osc_rdma_module_t *module = GET_MODULE(win);
bool temp;
temp = check_config_value_bool ("no_locks", info);
if (temp && !module->no_locks) {
/* clean up the lock hash. it is up to the user to ensure no lock is
* outstanding from this process when setting the info key */
OBJ_DESTRUCT(&module->outstanding_locks);
OBJ_CONSTRUCT(&module->outstanding_locks, opal_hash_table_t);
module->no_locks = true;
win->w_flags |= OMPI_WIN_NO_LOCKS;
} else if (!temp && module->no_locks) {
int world_size = ompi_comm_size (module->comm);
int init_limit = world_size > 256 ? 256 : world_size;
int ret;
ret = opal_hash_table_init (&module->outstanding_locks, init_limit);
if (OPAL_SUCCESS != ret) {
return ret;
}
module->no_locks = false;
win->w_flags &= ~OMPI_WIN_NO_LOCKS;
}
/* enforce collectiveness... */
return module->comm->c_coll->coll_barrier(module->comm,
module->comm->c_coll->coll_barrier_module);
}
static int ompi_osc_rdma_get_info (struct ompi_win_t *win, struct opal_info_t **info_used)
{
opal_info_t *info = OBJ_NEW(opal_info_t);
if (NULL == info) {
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
*info_used = info;
return OMPI_SUCCESS;
}
#endif
OBJ_CLASS_INSTANCE(ompi_osc_rdma_aggregation_t, opal_list_item_t, NULL, NULL);

Просмотреть файл

@ -1,6 +1,6 @@
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2014-2017 Los Alamos National Security, LLC. All rights
* Copyright (c) 2014-2018 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
@ -44,9 +44,6 @@ struct ompi_osc_rdma_peer_t {
/** peer flags */
volatile int32_t flags;
/** aggregation support */
ompi_osc_rdma_aggregation_t *aggregate;
};
typedef struct ompi_osc_rdma_peer_t ompi_osc_rdma_peer_t;
@ -164,13 +161,6 @@ int ompi_osc_rdma_new_peer (struct ompi_osc_rdma_module_t *module, int peer_id,
*/
struct ompi_osc_rdma_peer_t *ompi_osc_rdma_peer_lookup (struct ompi_osc_rdma_module_t *module, int peer_id);
/**
* @brief flush queued aggregated operation
*
* @param[in] peer osc rdma peer
*/
int ompi_osc_rdma_peer_aggregate_flush (ompi_osc_rdma_peer_t *peer);
/**
* @brief lookup the btl endpoint for a peer
*

Просмотреть файл

@ -17,14 +17,12 @@ static void ompi_osc_rdma_sync_constructor (ompi_osc_rdma_sync_t *rdma_sync)
rdma_sync->type = OMPI_OSC_RDMA_SYNC_TYPE_NONE;
rdma_sync->epoch_active = false;
rdma_sync->outstanding_rdma.counter = 0;
OBJ_CONSTRUCT(&rdma_sync->aggregations, opal_list_t);
OBJ_CONSTRUCT(&rdma_sync->lock, opal_mutex_t);
OBJ_CONSTRUCT(&rdma_sync->demand_locked_peers, opal_list_t);
}
static void ompi_osc_rdma_sync_destructor (ompi_osc_rdma_sync_t *rdma_sync)
{
OBJ_DESTRUCT(&rdma_sync->aggregations);
OBJ_DESTRUCT(&rdma_sync->lock);
OBJ_DESTRUCT(&rdma_sync->demand_locked_peers);
}

Просмотреть файл

@ -97,9 +97,6 @@ struct ompi_osc_rdma_sync_t {
/** outstanding rdma operations on epoch */
ompi_osc_rdma_sync_aligned_counter_t outstanding_rdma __opal_attribute_aligned__(64);
/** aggregated operations in this epoch */
opal_list_t aggregations;
/** lock to protect sync structure members */
opal_mutex_t lock;
};

Просмотреть файл

@ -1,6 +1,6 @@
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2014-2017 Los Alamos National Security, LLC. All rights
* Copyright (c) 2014-2018 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
@ -171,40 +171,6 @@ struct ompi_osc_rdma_state_t {
};
typedef struct ompi_osc_rdma_state_t ompi_osc_rdma_state_t;
struct ompi_osc_rdma_aggregation_t {
opal_list_item_t super;
/** associated peer */
struct ompi_osc_rdma_peer_t *peer;
/** aggregation buffer frag */
struct ompi_osc_rdma_frag_t *frag;
/** synchronization object */
struct ompi_osc_rdma_sync_t *sync;
/** aggregation buffer */
char *buffer;
/** target for the operation */
osc_rdma_base_t target_address;
/** handle for target memory address */
mca_btl_base_registration_handle_t *target_handle;
/** buffer size */
size_t buffer_size;
/** buffer used */
size_t buffer_used;
/** type */
int type;
};
typedef struct ompi_osc_rdma_aggregation_t ompi_osc_rdma_aggregation_t;
OBJ_CLASS_DECLARATION(ompi_osc_rdma_aggregation_t);
typedef void (*ompi_osc_rdma_pending_op_cb_fn_t) (void *, void *, int);
struct ompi_osc_rdma_pending_op_t {