1
1
openmpi/ompi/mca/osc/rdma/osc_rdma_comm.c

918 строки
40 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
* Copyright (c) 2014-2018 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2016-2018 Intel, Inc. All rights reserved.
* Copyright (c) 2017 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2017 IBM Corporation. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "osc_rdma_comm.h"
#include "osc_rdma_sync.h"
#include "osc_rdma_request.h"
#include "osc_rdma_dynamic.h"
#include "ompi/mca/osc/base/osc_base_obj_convert.h"
#include "opal/align.h"
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
/* helper functions */
static inline void ompi_osc_rdma_cleanup_rdma (ompi_osc_rdma_sync_t *sync, bool dec_always, ompi_osc_rdma_frag_t *frag,
mca_btl_base_registration_handle_t *handle, ompi_osc_rdma_request_t *request)
{
if (frag) {
ompi_osc_rdma_frag_complete (frag);
} else {
ompi_osc_rdma_deregister (sync->module, handle);
}
if (request) {
(void) OPAL_THREAD_ADD_FETCH32 (&request->outstanding_requests, -1);
}
if (dec_always) {
ompi_osc_rdma_sync_rdma_dec_always (sync);
} else {
ompi_osc_rdma_sync_rdma_dec (sync);
}
}
static int ompi_osc_rdma_get_contig (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);
static void ompi_osc_get_data_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)
{
assert (OPAL_SUCCESS == status);
((bool *) context)[0] = true;
}
int ompi_osc_get_data_blocking (ompi_osc_rdma_module_t *module, struct mca_btl_base_endpoint_t *endpoint,
uint64_t source_address, mca_btl_base_registration_handle_t *source_handle,
void *data, size_t len)
{
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
const size_t btl_alignment_mask = ALIGNMENT_MASK(module->selected_btl->btl_get_alignment);
mca_btl_base_registration_handle_t *local_handle = NULL;
ompi_osc_rdma_frag_t *frag = NULL;
volatile bool read_complete = false;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
size_t aligned_len, offset;
uint64_t aligned_addr = source_address & ~btl_alignment_mask;
char *ptr = data;
int ret;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
offset = source_address & btl_alignment_mask;
aligned_len = (len + offset + btl_alignment_mask) & ~btl_alignment_mask;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "reading data from endpoint %p. source: 0x%" PRIx64 " (aligned: 0x%" PRIx64
"), len: %lu (aligned: %lu)", (void *) endpoint, source_address, aligned_addr, (unsigned long) len,
(unsigned long) aligned_len);
if (module->selected_btl->btl_register_mem && len >= module->selected_btl->btl_get_local_registration_threshold) {
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
do {
ret = ompi_osc_rdma_frag_alloc (module, aligned_len, &frag, &ptr);
if (OPAL_UNLIKELY(OMPI_ERR_OUT_OF_RESOURCE == ret)) {
ompi_osc_rdma_progress (module);
}
} while (OMPI_ERR_OUT_OF_RESOURCE == ret);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_ERROR, "error allocating temporary buffer");
return ret;
}
local_handle = frag->handle;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "allocated temporary buffer %p in fragment %p", (void*)ptr,
(void *) frag);
}
assert (!(source_address & ALIGNMENT_MASK(module->selected_btl->btl_get_alignment)));
do {
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ret = module->selected_btl->btl_get (module->selected_btl, endpoint, ptr, aligned_addr,
local_handle, source_handle, aligned_len, 0, MCA_BTL_NO_ORDER,
ompi_osc_get_data_complete, (void *) &read_complete, NULL);
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (!ompi_osc_rdma_oor (ret)) {
break;
}
ompi_osc_rdma_progress (module);
} while (1);
if (OPAL_UNLIKELY(OMPI_SUCCESS > ret)) {
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_ERROR, "btl get failed with opal error code %d", ret);
if (frag) {
ompi_osc_rdma_frag_complete (frag);
}
return ret;
}
/* block until the callback is called */
while (!read_complete) {
ompi_osc_rdma_progress (module);
}
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "finished reading state data from endpoint %p", (void *) endpoint);
opal_memchecker_base_mem_defined (ptr, len);
if (frag) {
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
memcpy (data, ptr + offset, len);
/* done with the fragment */
ompi_osc_rdma_frag_complete (frag);
}
return OMPI_SUCCESS;
}
/**
* @brief function signature for the rdma transfer function used by ompi_osc_rdma_master_noncontig()
*
* @param[in] peer peer object for remote peer
* @param[in] remote_address base of remote region (destination for put, source for get)
* @param[in] remote_handle btl registration handle for remote region (must be valid for the entire region)
* @param[in] local_address base of local region (source for put, destination for get)
* @param[in] size number of bytes to transfer
* @param[in] module osc rdma module
* @param[in] request osc rdma request if used (can be NULL)
*
* @returns OMPI_SUCCESS on success
* @returns OMPI_ERR_OUT_OF_RESOURCE on temporary error
* @returns other OMPI error on fatal error
*
* This function does the work of scheduling a contiguous transfer between the local and remote regions.
*/
typedef int (*ompi_osc_rdma_fn_t) (ompi_osc_rdma_sync_t *sync, ompi_osc_rdma_peer_t *peer, uint64_t remote_address,
mca_btl_base_registration_handle_t *remote_handle, void *local_address, size_t size,
ompi_osc_rdma_request_t *request);
/**
* @brief break down rdma transaction into contiguous regions
*
* @param[in] local_address base of local region (source for put, destination for get)
* @param[in] local_count number of elements in local region
* @param[in] local_datatype datatype of local region
* @param[in] peer peer object for remote peer
* @param[in] remote_address base of remote region (destination for put, source for get)
* @param[in] remote_handle btl registration handle for remote region (must be valid for the entire region)
* @param[in] remote_count number of elements in remote region
* @param[in] remote_datatype datatype of remote region
* @param[in] module osc rdma module
* @param[in] request osc rdma request if used (can be NULL)
* @param[in] max_rdma_len maximum length of an rdma request (usually btl limitation)
* @param[in] rdma_fn function to use for contiguous rdma operations
* @param[in] alloc_reqs true if rdma_fn requires a valid request object (any allocated objects will be marked internal)
*
* This function does the work of breaking a non-contiguous rdma transfer into contiguous components. It will
* continue to submit rdma transfers until the entire region is transferred or a fatal error occurs.
*/
static int ompi_osc_rdma_master_noncontig (ompi_osc_rdma_sync_t *sync, void *local_address, int local_count, ompi_datatype_t *local_datatype,
ompi_osc_rdma_peer_t *peer, uint64_t remote_address,
mca_btl_base_registration_handle_t *remote_handle, int remote_count,
ompi_datatype_t *remote_datatype, ompi_osc_rdma_request_t *request, const size_t max_rdma_len,
const ompi_osc_rdma_fn_t rdma_fn, const bool alloc_reqs)
{
ompi_osc_rdma_module_t *module = sync->module;
struct iovec local_iovec[OMPI_OSC_RDMA_DECODE_MAX], remote_iovec[OMPI_OSC_RDMA_DECODE_MAX];
opal_convertor_t local_convertor, remote_convertor;
uint32_t local_iov_count, remote_iov_count;
uint32_t local_iov_index, remote_iov_index;
/* needed for opal_convertor_raw but not used */
size_t local_size, remote_size, rdma_len;
ompi_osc_rdma_request_t *subreq;
int ret;
bool done;
subreq = NULL;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "scheduling rdma on non-contiguous datatype(s) or large region");
/* prepare convertors for the source and target. these convertors will be used to determine the
* contiguous segments within the source and target. */
OBJ_CONSTRUCT(&remote_convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &remote_datatype->super, remote_count,
(void *) (intptr_t) remote_address, 0, &remote_convertor);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
OBJ_CONSTRUCT(&local_convertor, opal_convertor_t);
ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &local_datatype->super, local_count,
local_address, 0, &local_convertor);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
if (request) {
/* keep the request from completing until all the transfers have started */
request->outstanding_requests = 1;
}
local_iov_index = 0;
local_iov_count = 0;
do {
/* decode segments of the remote data */
remote_iov_count = OMPI_OSC_RDMA_DECODE_MAX;
remote_iov_index = 0;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
/* opal_convertor_raw returns true when it has reached the end of the data */
done = opal_convertor_raw (&remote_convertor, remote_iovec, &remote_iov_count, &remote_size);
/* loop on the target segments until we have exhaused the decoded source data */
while (remote_iov_index != remote_iov_count) {
if (local_iov_index == local_iov_count) {
/* decode segments of the target buffer */
local_iov_count = OMPI_OSC_RDMA_DECODE_MAX;
local_iov_index = 0;
(void) opal_convertor_raw (&local_convertor, local_iovec, &local_iov_count, &local_size);
}
/* we already checked that the target was large enough. this should be impossible */
assert (0 != local_iov_count);
/* determine how much to transfer in this operation */
rdma_len = min(min(local_iovec[local_iov_index].iov_len, remote_iovec[remote_iov_index].iov_len), max_rdma_len);
/* execute the get */
if (!subreq && alloc_reqs) {
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, subreq);
subreq->internal = true;
subreq->type = OMPI_OSC_RDMA_TYPE_RDMA;
subreq->parent_request = request;
if (request) {
(void) OPAL_THREAD_ADD_FETCH32 (&request->outstanding_requests, 1);
}
} else if (!alloc_reqs) {
subreq = request;
}
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "performing rdma on contiguous region. local: %p, remote: %p, len: %lu",
local_iovec[local_iov_index].iov_base, remote_iovec[remote_iov_index].iov_base,
(unsigned long) remote_iovec[remote_iov_index].iov_len);
ret = rdma_fn (sync, peer, (uint64_t) (intptr_t) remote_iovec[remote_iov_index].iov_base, remote_handle,
local_iovec[local_iov_index].iov_base, rdma_len, subreq);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
if (OPAL_UNLIKELY(OMPI_ERR_OUT_OF_RESOURCE != ret)) {
if (request) {
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ompi_osc_rdma_request_deref (request);
}
if (alloc_reqs) {
OMPI_OSC_RDMA_REQUEST_RETURN(subreq);
}
/* something bad happened. need to figure out best way to handle rma errors */
return ret;
}
/* progress and try again */
ompi_osc_rdma_progress (module);
continue;
}
subreq = NULL;
/* adjust io vectors */
local_iovec[local_iov_index].iov_len -= rdma_len;
remote_iovec[remote_iov_index].iov_len -= rdma_len;
local_iovec[local_iov_index].iov_base = (void *)((intptr_t) local_iovec[local_iov_index].iov_base + rdma_len);
remote_iovec[remote_iov_index].iov_base = (void *)((intptr_t) remote_iovec[remote_iov_index].iov_base + rdma_len);
local_iov_index += (0 == local_iovec[local_iov_index].iov_len);
remote_iov_index += (0 == remote_iovec[remote_iov_index].iov_len);
}
} while (!done);
if (request) {
/* release our reference so the request can complete */
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ompi_osc_rdma_request_deref (request);
}
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "finished scheduling rdma on non-contiguous datatype(s)");
/* clean up convertors */
opal_convertor_cleanup (&local_convertor);
OBJ_DESTRUCT(&local_convertor);
opal_convertor_cleanup (&remote_convertor);
OBJ_DESTRUCT(&remote_convertor);
return OMPI_SUCCESS;
}
static inline int ompi_osc_rdma_master (ompi_osc_rdma_sync_t *sync, void *local_address, int local_count,
ompi_datatype_t *local_datatype, ompi_osc_rdma_peer_t *peer,
uint64_t remote_address, mca_btl_base_registration_handle_t *remote_handle,
int remote_count, ompi_datatype_t *remote_datatype,
ompi_osc_rdma_request_t *request, const size_t max_rdma_len,
const ompi_osc_rdma_fn_t rdma_fn, const bool alloc_reqs)
{
size_t rdma_len;
ptrdiff_t lb, extent;
int ret;
rdma_len = local_datatype->super.size * local_count;
/* fast path for contiguous rdma */
if (OPAL_LIKELY(ompi_datatype_is_contiguous_memory_layout (local_datatype, local_count) &&
ompi_datatype_is_contiguous_memory_layout (remote_datatype, remote_count) &&
rdma_len <= max_rdma_len)) {
if (NULL == request && alloc_reqs) {
ompi_osc_rdma_module_t *module = sync->module;
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, request);
request->internal = true;
request->type = OMPI_OSC_RDMA_TYPE_RDMA;
}
/* ignore failure here */
(void) ompi_datatype_get_true_extent (local_datatype, &lb, &extent);
local_address = (void *)((intptr_t) local_address + lb);
(void) ompi_datatype_get_true_extent (remote_datatype, &lb, &extent);
remote_address += lb;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "performing rdma on contiguous region. local: %p, "
"remote: 0x%lx, length: %lu", local_address, (unsigned long) remote_address,
rdma_len);
do {
ret = rdma_fn (sync, peer, remote_address, remote_handle, local_address, rdma_len, request);
if (OPAL_LIKELY(OPAL_SUCCESS == ret)) {
return OMPI_SUCCESS;
}
ompi_osc_rdma_progress (sync->module);
} while (1);
}
return ompi_osc_rdma_master_noncontig (sync, local_address, local_count, local_datatype, peer, remote_address,
remote_handle, remote_count, remote_datatype, request,
max_rdma_len, rdma_fn, alloc_reqs);
}
static int ompi_osc_rdma_copy_local (const void *source, int source_count, ompi_datatype_t *source_datatype,
void *target, int target_count, ompi_datatype_t *target_datatype,
ompi_osc_rdma_request_t *request)
{
int ret;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "performing local copy from %p -> %p", source, target);
opal_atomic_mb ();
ret = ompi_datatype_sndrcv (source, source_count, source_datatype, target, target_count, target_datatype);
if (request) {
ompi_osc_rdma_request_complete (request, ret);
}
return ret;
}
static void ompi_osc_rdma_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_sync_t *sync = (ompi_osc_rdma_sync_t *) context;
assert (OPAL_SUCCESS == status);
/* the lowest bit is used as a flag indicating this put operation has a request */
if ((intptr_t) context & 0x1) {
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ompi_osc_rdma_request_t *request = request = (ompi_osc_rdma_request_t *) ((intptr_t) context & ~1);
sync = request->sync;
/* NTH -- TODO: better error handling */
ompi_osc_rdma_request_complete (request, status);
}
OSC_RDMA_VERBOSE(status ? MCA_BASE_VERBOSE_ERROR : MCA_BASE_VERBOSE_TRACE, "btl put complete on sync %p. local "
"address %p. opal status %d", (void *) sync, local_address, status);
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (data) {
ompi_osc_rdma_frag_complete ((ompi_osc_rdma_frag_t *) data);
} else if (local_handle) {
ompi_osc_rdma_deregister (sync->module, local_handle);
}
ompi_osc_rdma_sync_rdma_dec (sync);
}
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
static void ompi_osc_rdma_put_complete_flush (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_module_t *module = (ompi_osc_rdma_module_t *) context;
assert (OPAL_SUCCESS == status);
/* the lowest bit is used as a flag indicating this put operation has a request */
if ((intptr_t) context & 0x1) {
ompi_osc_rdma_request_t *request = request = (ompi_osc_rdma_request_t *) ((intptr_t) context & ~1);
module = request->module;
/* NTH -- TODO: better error handling */
ompi_osc_rdma_request_complete (request, status);
}
OSC_RDMA_VERBOSE(status ? MCA_BASE_VERBOSE_ERROR : MCA_BASE_VERBOSE_TRACE, "btl put complete on module %p. local "
"address %p. opal status %d", (void *) module, local_address, status);
if (data) {
ompi_osc_rdma_frag_complete ((ompi_osc_rdma_frag_t *) data);
} else if (local_handle) {
ompi_osc_rdma_deregister (module, local_handle);
}
}
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,
mca_btl_base_rdma_completion_fn_t cb, void *context, void *cbdata) {
ompi_osc_rdma_module_t *module = sync->module;
int ret;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "initiating btl put of %lu bytes to remote address %" PRIx64 ", sync "
"object %p...", (unsigned long) size, target_address, (void *) sync);
/* flag outstanding rma requests */
ompi_osc_rdma_sync_rdma_inc (sync);
do {
ret = module->selected_btl->btl_put (module->selected_btl, peer->data_endpoint, ptr, target_address,
local_handle, target_handle, size, 0, MCA_BTL_NO_ORDER,
cb, context, cbdata);
if (OPAL_UNLIKELY(OMPI_SUCCESS == ret)) {
return OMPI_SUCCESS;
}
++module->put_retry_count;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (!ompi_osc_rdma_oor (ret)) {
break;
}
/* spin a bit on progress */
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ompi_osc_rdma_progress (module);
} while (1);
OSC_RDMA_VERBOSE(10, "btl put failed with opal error code %d", ret);
return ret;
}
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
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;
mca_btl_base_registration_handle_t *local_handle = NULL;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
mca_btl_base_rdma_completion_fn_t cbfunc = NULL;
ompi_osc_rdma_frag_t *frag = NULL;
char *ptr = source_buffer;
void *cbcontext;
int ret;
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)) {
ret = ompi_osc_rdma_register (module, peer->data_endpoint, source_buffer, size, 0, &local_handle);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
} else {
memcpy (ptr, source_buffer, size);
local_handle = frag->handle;
}
}
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (ompi_osc_rdma_use_btl_flush (module)) {
/* NTH: when using the btl_flush function there is no guarantee that the callback will happen
* before the flush is complete. because of this there is a chance that the sync object will be
* released before there is a callback. to handle this case we call different callback that doesn't
* use the sync object. its possible the btl sematics will change in the future and the callback
* will happen *before* flush is considered complete. if that is the case this workaround can be
* removed */
cbcontext = (void *) module;
if (request || local_handle || frag) {
cbfunc = ompi_osc_rdma_put_complete_flush;
}
/* else the callback function is a no-op so do not bother specifying one */
} else {
cbcontext = (void *) sync;
cbfunc = ompi_osc_rdma_put_complete;
}
/* increment the outstanding request counter in the request object */
if (request) {
(void) OPAL_THREAD_ADD_FETCH32 (&request->outstanding_requests, 1);
cbcontext = (void *) ((intptr_t) request | 1);
request->sync = sync;
}
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ret = ompi_osc_rdma_put_real (sync, peer, target_address, target_handle, ptr, local_handle, size, cbfunc,
cbcontext, frag);
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
ompi_osc_rdma_cleanup_rdma (sync, false, frag, local_handle, request);
}
return ret;
}
static void ompi_osc_rdma_get_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_request_t *request = (ompi_osc_rdma_request_t *) context;
intptr_t source = (intptr_t) local_address + request->offset;
ompi_osc_rdma_frag_t *frag = (ompi_osc_rdma_frag_t *) data;
ompi_osc_rdma_sync_t *sync = request->sync;
void *origin_addr = request->origin_addr;
OSC_RDMA_VERBOSE(status ? MCA_BASE_VERBOSE_ERROR : MCA_BASE_VERBOSE_TRACE, "btl get complete on sync %p. local "
"address %p. origin %p. opal status %d", (void *) sync, local_address, origin_addr, status);
assert (OPAL_SUCCESS == status);
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (request->buffer || frag) {
if (OPAL_LIKELY(OMPI_SUCCESS == status)) {
memcpy (origin_addr, (void *) source, request->len);
}
}
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (NULL == request->buffer) {
/* completion detection can handle this case without the counter when using btl_flush */
ompi_osc_rdma_sync_rdma_dec (sync);
} else {
/* the counter was needed to keep track of the number of outstanding operations */
ompi_osc_rdma_sync_rdma_dec_always (sync);
}
if (NULL != frag) {
ompi_osc_rdma_frag_complete (frag);
} else {
ompi_osc_rdma_deregister (sync->module, local_handle);
}
ompi_osc_rdma_request_complete (request, status);
}
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) {
ompi_osc_rdma_module_t *module = sync->module;
ompi_osc_rdma_request_t *subreq;
int ret;
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, subreq);
subreq->internal = true;
subreq->type = OMPI_OSC_RDMA_TYPE_RDMA;
subreq->parent_request = request;
(void) OPAL_THREAD_ADD_FETCH32 (&request->outstanding_requests, 1);
ret = ompi_osc_rdma_get_contig (sync, peer, source_address, source_handle, target_buffer, size, subreq);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OMPI_OSC_RDMA_REQUEST_RETURN(subreq);
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ompi_osc_rdma_request_deref (request);
}
return ret;
}
static int ompi_osc_rdma_get_contig (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)
{
ompi_osc_rdma_module_t *module = sync->module;
const size_t btl_alignment_mask = ALIGNMENT_MASK(module->selected_btl->btl_get_alignment);
mca_btl_base_registration_handle_t *local_handle = NULL;
ompi_osc_rdma_frag_t *frag = NULL;
osc_rdma_size_t aligned_len;
osc_rdma_base_t aligned_source_base, aligned_source_bound;
char *ptr = target_buffer;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
bool counter_needs_inc = false;
int ret;
aligned_source_base = source_address & ~btl_alignment_mask;
aligned_source_bound = (source_address + size + btl_alignment_mask) & ~btl_alignment_mask;
aligned_len = aligned_source_bound - aligned_source_base;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "initiating get of %lu bytes from remote ptr %" PRIx64 " to local ptr %p",
size, source_address, target_buffer);
if ((module->selected_btl->btl_register_mem && size > module->selected_btl->btl_get_local_registration_threshold) ||
(((uint64_t) target_buffer | size | source_address) & btl_alignment_mask)) {
ret = ompi_osc_rdma_frag_alloc (module, aligned_len, &frag, &ptr);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
if (OMPI_ERR_VALUE_OUT_OF_BOUNDS == ret) {
/* region is too large for a buffered read */
size_t subsize;
if ((source_address & btl_alignment_mask) && (source_address & btl_alignment_mask) == ((intptr_t) target_buffer & btl_alignment_mask)) {
/* remote region has the same alignment but the base is not aligned. perform a small
* buffered get of the beginning of the remote region */
aligned_source_base = OPAL_ALIGN(source_address, module->selected_btl->btl_get_alignment, osc_rdma_base_t);
subsize = (size_t) (aligned_source_base - source_address);
ret = ompi_osc_rdma_get_partial (sync, peer, source_address, source_handle, target_buffer, subsize, request);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
source_address += subsize;
target_buffer = (void *) ((intptr_t) target_buffer + subsize);
size -= subsize;
aligned_len = aligned_source_bound - aligned_source_base;
}
if (!(((uint64_t) target_buffer | source_address) & btl_alignment_mask) &&
(size & btl_alignment_mask)) {
/* remote region bases are aligned but the bounds are not. perform a
* small buffered get of the end of the remote region */
aligned_len = size & ~btl_alignment_mask;
subsize = size - aligned_len;
size = aligned_len;
ret = ompi_osc_rdma_get_partial (sync, peer, source_address + aligned_len, source_handle,
(void *) ((intptr_t) target_buffer + aligned_len), subsize, request);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
}
/* (remaining) user request is now correctly aligned */
}
if ((((uint64_t) target_buffer | size | source_address) & btl_alignment_mask)) {
/* local and remote alignments differ */
request->buffer = ptr = malloc (aligned_len);
} else {
ptr = target_buffer;
}
if (NULL != ptr) {
(void) ompi_osc_rdma_register (module, peer->data_endpoint, ptr, aligned_len, MCA_BTL_REG_FLAG_LOCAL_WRITE,
&local_handle);
}
if (OPAL_UNLIKELY(NULL == local_handle)) {
free (request->buffer);
request->buffer = NULL;
return ret;
}
} else {
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "using internal buffer %p in fragment %p for get of size %lu bytes, source address 0x%lx",
(void*)ptr, (void *) frag, (unsigned long) aligned_len, (unsigned long) aligned_source_base);
local_handle = frag->handle;
}
}
request->offset = source_address - aligned_source_base;
request->len = size;
request->origin_addr = target_buffer;
request->sync = sync;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (request->buffer) {
/* always increment the outstanding RDMA counter as the btl_flush function does not guarantee callback completion,
* just operation completion. */
counter_needs_inc = true;
ompi_osc_rdma_sync_rdma_inc_always (sync);
} else {
/* if this operation is being buffered with a frag then ompi_osc_rdma_sync_rdma_complete() can use the number
* of pending operations on the rdma_frag as an indicator as to whether the operation is complete. this can
* only be done since there is only on rdma frag per module. if that changes this logic will need to be changed
* as well. this path also covers the case where the get operation is not buffered. */
ompi_osc_rdma_sync_rdma_inc (sync);
}
do {
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ret = module->selected_btl->btl_get (module->selected_btl, peer->data_endpoint, ptr,
aligned_source_base, local_handle, source_handle,
aligned_len, 0, MCA_BTL_NO_ORDER, ompi_osc_rdma_get_complete,
request, frag);
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (OPAL_LIKELY(OMPI_SUCCESS == ret)) {
return OMPI_SUCCESS;
}
++module->get_retry_count;
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
if (!ompi_osc_rdma_oor (ret)) {
break;
}
/* spin a bit on progress */
for (int i = 0 ; i < 10 ; ++i) {
ompi_osc_rdma_progress (module);
}
} while (1);
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_ERROR, "btl get failed with opal error code %d", ret);
osc/rdma: performance improvments and bug fixes This commit is a large update to the osc/rdma component. Included in this commit: - Add support for using hardware atomics for fetch-and-op and single count accumulate when using the accumulate lock. This will improve the performance of these operations even when not setting the single intrinsic info key. - Rework how large accumulates are done. They now block on the get operation to fix some bugs discovered by an IBM one-sided test. I may roll back some of the changes if the underlying bug in the original design is discovered. There appear to be no real difference (on the hardware this was tested with) in performance so its probably a non-issue. References #2530. - Add support for an additional lock-all algorithm: on-demand. The on-demand algorithm will attempt to acquire the peer lock when starting an RMA operation. The lock algorithm default has not changed. The algorithm can be selected by setting the osc_rdma_locking_mode MCA variable. The valid values are two_level and on_demand. - Make use of the btl_flush function if available. This can improve performance with some btls. - When using btl_flush do not keep track of the number of put operations. This reduces the number of atomic operations in the critical path. - Make the window buffers more friendly to multi-threaded applications. This was done by dropping support for multiple buffers per MPI window. I intend to re-add that support once the underlying performance bug under the old buffering scheme is fixed. - Fix a bug in request completion in the accumulate, get, and put paths. This also helps with #2530. - General code cleanup and fixes. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-03-15 12:20:41 -06:00
ompi_osc_rdma_cleanup_rdma (sync, counter_needs_inc, frag, local_handle, request);
return ret;
}
static inline int ompi_osc_rdma_put_w_req (ompi_osc_rdma_sync_t *sync, const void *origin_addr, int origin_count,
ompi_datatype_t *origin_datatype, ompi_osc_rdma_peer_t *peer,
ptrdiff_t target_disp, int target_count,
ompi_datatype_t *target_datatype, ompi_osc_rdma_request_t *request)
{
ompi_osc_rdma_module_t *module = sync->module;
mca_btl_base_registration_handle_t *target_handle;
uint64_t target_address;
int ret;
/* short-circuit case */
if (0 == origin_count || 0 == target_count) {
if (request) {
ompi_osc_rdma_request_complete (request, MPI_SUCCESS);
}
return OMPI_SUCCESS;
}
ptrdiff_t len, offset;
// a buffer defined by (buf, count, dt)
// will have data starting at buf+offset and ending len bytes later:
len = opal_datatype_span(&target_datatype->super, target_count, &offset);
// the below function wants arg4 to be the number of bytes after
// source_disp that the data ends, which is offset+len
ret = osc_rdma_get_remote_segment (module, peer, target_disp, offset+len,
&target_address, &target_handle);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
/* optimize communication with peers that we can do direct load and store operations on */
if (ompi_osc_rdma_peer_local_base (peer)) {
return ompi_osc_rdma_copy_local (origin_addr, origin_count, origin_datatype, (void *) (intptr_t) target_address,
target_count, target_datatype, request);
}
return ompi_osc_rdma_master (sync, (void *) origin_addr, origin_count, origin_datatype, peer, target_address, target_handle,
target_count, target_datatype, request, module->selected_btl->btl_put_limit,
ompi_osc_rdma_put_contig, false);
}
static inline int ompi_osc_rdma_get_w_req (ompi_osc_rdma_sync_t *sync, void *origin_addr, int origin_count, ompi_datatype_t *origin_datatype,
ompi_osc_rdma_peer_t *peer, ptrdiff_t source_disp, int source_count,
ompi_datatype_t *source_datatype, ompi_osc_rdma_request_t *request)
{
ompi_osc_rdma_module_t *module = sync->module;
mca_btl_base_registration_handle_t *source_handle;
uint64_t source_address;
ptrdiff_t source_span, source_lb;
int ret;
/* short-circuit case */
if (0 == origin_count || 0 == source_count) {
if (request) {
ompi_osc_rdma_request_complete (request, MPI_SUCCESS);
}
return OMPI_SUCCESS;
}
// a buffer defined by (buf, count, dt)
// will have data starting at buf+offset and ending len bytes later:
source_span = opal_datatype_span(&source_datatype->super, source_count, &source_lb);
ret = osc_rdma_get_remote_segment (module, peer, source_disp, source_span+source_lb,
&source_address, &source_handle);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
/* optimize self/local communication */
if (ompi_osc_rdma_peer_local_base (peer)) {
return ompi_osc_rdma_copy_local ((void *) (intptr_t) source_address, source_count, source_datatype,
origin_addr, origin_count, origin_datatype, request);
}
return ompi_osc_rdma_master (sync, origin_addr, origin_count, origin_datatype, peer, source_address,
source_handle, source_count, source_datatype, request,
module->selected_btl->btl_get_limit, ompi_osc_rdma_get_contig, true);
}
int ompi_osc_rdma_put (const void *origin_addr, int origin_count, ompi_datatype_t *origin_datatype,
int target_rank, ptrdiff_t target_disp, int target_count,
ompi_datatype_t *target_datatype, ompi_win_t *win)
{
ompi_osc_rdma_module_t *module = GET_MODULE(win);
ompi_osc_rdma_peer_t *peer;
ompi_osc_rdma_sync_t *sync;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "put: 0x%lx, %d, %s, %d, %d, %d, %s, %s", (unsigned long) origin_addr,
origin_count, origin_datatype->name, target_rank, (int) target_disp, target_count,
target_datatype->name, win->w_name);
sync = ompi_osc_rdma_module_sync_lookup (module, target_rank, &peer);
if (OPAL_UNLIKELY(NULL == sync)) {
return OMPI_ERR_RMA_SYNC;
}
return ompi_osc_rdma_put_w_req (sync, origin_addr, origin_count, origin_datatype, peer, target_disp,
target_count, target_datatype, NULL);
}
int ompi_osc_rdma_rput (const void *origin_addr, int origin_count, ompi_datatype_t *origin_datatype,
int target_rank, ptrdiff_t target_disp, int target_count,
ompi_datatype_t *target_datatype, ompi_win_t *win,
ompi_request_t **request)
{
ompi_osc_rdma_module_t *module = GET_MODULE(win);
ompi_osc_rdma_peer_t *peer;
ompi_osc_rdma_request_t *rdma_request;
ompi_osc_rdma_sync_t *sync;
int ret;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "rput: 0x%lx, %d, %s, %d, %d, %d, %s, %s", (unsigned long) origin_addr, origin_count,
origin_datatype->name, target_rank, (int) target_disp, target_count, target_datatype->name, win->w_name);
sync = ompi_osc_rdma_module_sync_lookup (module, target_rank, &peer);
if (OPAL_UNLIKELY(NULL == sync)) {
return OMPI_ERR_RMA_SYNC;
}
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, rdma_request);
rdma_request->type = OMPI_OSC_RDMA_TYPE_PUT;
ret = ompi_osc_rdma_put_w_req (sync, origin_addr, origin_count, origin_datatype, peer, target_disp,
target_count, target_datatype, rdma_request);
if (OMPI_SUCCESS != ret) {
OMPI_OSC_RDMA_REQUEST_RETURN(rdma_request);
return ret;
}
*request = (ompi_request_t *) rdma_request;
return OMPI_SUCCESS;
}
int ompi_osc_rdma_get (void *origin_addr, int origin_count, ompi_datatype_t *origin_datatype,
int source_rank, ptrdiff_t source_disp, int source_count,
ompi_datatype_t *source_datatype, ompi_win_t *win)
{
ompi_osc_rdma_module_t *module = GET_MODULE(win);
ompi_osc_rdma_peer_t *peer;
ompi_osc_rdma_sync_t *sync;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "get: 0x%lx, %d, %s, %d, %d, %d, %s, %s", (unsigned long) origin_addr,
origin_count, origin_datatype->name, source_rank, (int) source_disp, source_count,
source_datatype->name, win->w_name);
sync = ompi_osc_rdma_module_sync_lookup (module, source_rank, &peer);
if (OPAL_UNLIKELY(NULL == sync)) {
return OMPI_ERR_RMA_SYNC;
}
return ompi_osc_rdma_get_w_req (sync, origin_addr, origin_count, origin_datatype, peer,
source_disp, source_count, source_datatype, NULL);
}
int ompi_osc_rdma_rget (void *origin_addr, int origin_count, ompi_datatype_t *origin_datatype,
int source_rank, ptrdiff_t source_disp, int source_count,
ompi_datatype_t *source_datatype, ompi_win_t *win,
ompi_request_t **request)
{
ompi_osc_rdma_module_t *module = GET_MODULE(win);
ompi_osc_rdma_peer_t *peer;
ompi_osc_rdma_request_t *rdma_request;
ompi_osc_rdma_sync_t *sync;
int ret;
OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_TRACE, "rget: 0x%lx, %d, %s, %d, %d, %d, %s, %s", (unsigned long) origin_addr,
origin_count, origin_datatype->name, source_rank, (int) source_disp, source_count,
source_datatype->name, win->w_name);
sync = ompi_osc_rdma_module_sync_lookup (module, source_rank, &peer);
if (OPAL_UNLIKELY(NULL == sync)) {
return OMPI_ERR_RMA_SYNC;
}
OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, rdma_request);
rdma_request->type = OMPI_OSC_RDMA_TYPE_GET;
ret = ompi_osc_rdma_get_w_req (sync, origin_addr, origin_count, origin_datatype, peer,
source_disp, source_count, source_datatype, rdma_request);
if (OMPI_SUCCESS != ret) {
OMPI_OSC_RDMA_REQUEST_RETURN(rdma_request);
return ret;
}
*request = (ompi_request_t *) rdma_request;
return OMPI_SUCCESS;
}