1
1
openmpi/opal/mca/btl/ugni/btl_ugni_module.c

363 строки
13 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
* Copyright (c) 2011-2018 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2011 UT-Battelle, LLC. All rights reserved.
* Copyright (c) 2014-2016 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2017 Intel, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 00:47:28 +00:00
#include "opal_config.h"
#include "btl_ugni.h"
#include "btl_ugni_frag.h"
#include "btl_ugni_endpoint.h"
#include "btl_ugni_prepare.h"
#include "btl_ugni_smsg.h"
static int
mca_btl_ugni_free (struct mca_btl_base_module_t *btl,
mca_btl_base_descriptor_t *des);
static int
mca_btl_ugni_module_finalize (struct mca_btl_base_module_t* btl);
static struct mca_btl_base_descriptor_t *
mca_btl_ugni_prepare_src (struct mca_btl_base_module_t *btl,
struct mca_btl_base_endpoint_t *endpoint,
struct opal_convertor_t *convertor,
uint8_t order, size_t reserve, size_t *size,
uint32_t flags);
static mca_btl_base_registration_handle_t *
mca_btl_ugni_register_mem (mca_btl_base_module_t *btl, mca_btl_base_endpoint_t *endpoint, void *base,
size_t size, uint32_t flags);
static int mca_btl_ugni_deregister_mem (mca_btl_base_module_t *btl, mca_btl_base_registration_handle_t *handle);
mca_btl_ugni_module_t mca_btl_ugni_module = {
.super = {
.btl_component = &mca_btl_ugni_component.super,
.btl_add_procs = mca_btl_ugni_add_procs,
.btl_del_procs = mca_btl_ugni_del_procs,
.btl_finalize = mca_btl_ugni_module_finalize,
.btl_alloc = mca_btl_ugni_alloc,
.btl_free = mca_btl_ugni_free,
.btl_prepare_src = mca_btl_ugni_prepare_src,
.btl_send = mca_btl_ugni_send,
.btl_sendi = mca_btl_ugni_sendi,
.btl_put = mca_btl_ugni_put,
.btl_get = mca_btl_ugni_get,
.btl_register_mem = mca_btl_ugni_register_mem,
.btl_deregister_mem = mca_btl_ugni_deregister_mem,
.btl_atomic_op = mca_btl_ugni_aop,
.btl_atomic_fop = mca_btl_ugni_afop,
.btl_atomic_cswap = mca_btl_ugni_acswap,
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
.btl_flush = mca_btl_ugni_flush,
}
};
static void mca_btl_ugni_datagram_event (int foo, short bar, void *arg)
{
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *) arg;
mca_btl_ugni_device_t *device = ugni_module->devices;
struct timeval tv = {.tv_sec = 0, .tv_usec = MCA_BTL_UGNI_CONNECT_USEC};
mca_btl_ugni_progress_datagram (device);
opal_event_evtimer_add (&ugni_module->connection_event, &tv);
}
int
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
mca_btl_ugni_module_init (mca_btl_ugni_module_t *ugni_module)
{
int rc;
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
BTL_VERBOSE(("binding module %p to device 0", (void *) ugni_module));
/* copy module defaults (and function pointers) */
memmove (ugni_module, &mca_btl_ugni_module, sizeof (mca_btl_ugni_module));
ugni_module->initialized = false;
ugni_module->nlocal_procs = 0;
ugni_module->active_datagrams = 0;
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
ugni_module->active_rdma_count = 0;
opal_event_evtimer_set (opal_sync_event_base, &ugni_module->connection_event,
mca_btl_ugni_datagram_event, ugni_module);
OBJ_CONSTRUCT(&ugni_module->failed_frags, opal_list_t);
OBJ_CONSTRUCT(&ugni_module->failed_frags_lock, opal_mutex_t);
OBJ_CONSTRUCT(&ugni_module->eager_get_pending, opal_list_t);
OBJ_CONSTRUCT(&ugni_module->eager_get_pending_lock,opal_mutex_t);
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
for (int i = 0 ; i < MCA_BTL_UGNI_LIST_MAX ; ++i) {
OBJ_CONSTRUCT(ugni_module->frags_lists + i, opal_free_list_t);
}
OBJ_CONSTRUCT(&ugni_module->pending_smsg_frags_bb, opal_pointer_array_t);
OBJ_CONSTRUCT(&ugni_module->ep_wait_list_lock,opal_mutex_t);
OBJ_CONSTRUCT(&ugni_module->ep_wait_list, opal_list_t);
OBJ_CONSTRUCT(&ugni_module->endpoint_lock, opal_mutex_t);
OBJ_CONSTRUCT(&ugni_module->endpoints, opal_pointer_array_t);
OBJ_CONSTRUCT(&ugni_module->id_to_endpoint, opal_hash_table_t);
OBJ_CONSTRUCT(&ugni_module->smsg_mboxes, opal_free_list_t);
OBJ_CONSTRUCT(&ugni_module->eager_get_pending, opal_list_t);
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
/* set up virtual device handles */
for (int i = 0 ; i < mca_btl_ugni_component.virtual_device_count ; ++i) {
rc = mca_btl_ugni_device_init (ugni_module->devices + i, i);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
BTL_VERBOSE(("error initializing uGNI device handle"));
return rc;
}
}
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
/* create wildcard endpoint on first device to listen for connections.
* there is no need to bind this endpoint. We are single threaded
* here so there is no need for a device lock. */
rc = GNI_EpCreate (ugni_module->devices[0].dev_handle, NULL,
&ugni_module->wildcard_ep);
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 00:47:28 +00:00
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
BTL_ERROR(("error creating wildcard ugni endpoint"));
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
return mca_btl_rc_ugni_to_opal (rc);
}
/* post wildcard datagram */
rc = mca_btl_ugni_wildcard_ep_post (ugni_module);
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 00:47:28 +00:00
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
BTL_ERROR(("error posting wildcard datagram"));
return rc;
}
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 00:47:28 +00:00
return OPAL_SUCCESS;
}
static int
mca_btl_ugni_module_finalize (struct mca_btl_base_module_t *btl)
{
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *)btl;
mca_btl_base_endpoint_t *ep;
uint64_t key;
2012-04-24 21:16:51 +00:00
int rc;
if (ugni_module->initialized) {
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
/* close all open connections and release endpoints */
OPAL_HASH_TABLE_FOREACH(key, uint64, ep, &ugni_module->id_to_endpoint) {
if (NULL != ep) {
mca_btl_ugni_release_ep (ep);
}
}
if (mca_btl_ugni_component.progress_thread_enabled) {
2014-12-04 16:18:16 -07:00
mca_btl_ugni_kill_progress_thread();
}
/* destroy all cqs */
rc = GNI_CqDestroy (ugni_module->smsg_remote_cq);
if (GNI_RC_SUCCESS != rc) {
2014-12-04 16:18:16 -07:00
BTL_ERROR(("error tearing down RX SMSG CQ - %s",gni_err_str[rc]));
}
if (mca_btl_ugni_component.progress_thread_enabled) {
rc = GNI_CqDestroy (ugni_module->smsg_remote_irq_cq);
if (GNI_RC_SUCCESS != rc) {
BTL_ERROR(("error tearing down remote SMSG CQ - %s",gni_err_str[rc]));
}
}
/* cancel wildcard post */
rc = GNI_EpPostDataCancelById (ugni_module->wildcard_ep,
MCA_BTL_UGNI_CONNECT_WILDCARD_ID |
OPAL_PROC_MY_NAME.vpid);
if (GNI_RC_SUCCESS != rc) {
BTL_VERBOSE(("btl/ugni error cancelling wildcard post"));
}
/* tear down wildcard endpoint */
rc = GNI_EpDestroy (ugni_module->wildcard_ep);
if (GNI_RC_SUCCESS != rc) {
2014-12-04 16:18:16 -07:00
BTL_VERBOSE(("btl/ugni error destroying endpoint - %s",gni_err_str[rc]));
}
opal_event_del (&ugni_module->connection_event);
}
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
for (int i = 0 ; i < MCA_BTL_UGNI_LIST_MAX ; ++i) {
OBJ_DESTRUCT(ugni_module->frags_lists + i);
}
OBJ_DESTRUCT(&ugni_module->ep_wait_list);
OBJ_DESTRUCT(&ugni_module->smsg_mboxes);
OBJ_DESTRUCT(&ugni_module->pending_smsg_frags_bb);
OBJ_DESTRUCT(&ugni_module->id_to_endpoint);
OBJ_DESTRUCT(&ugni_module->endpoint_lock);
OBJ_DESTRUCT(&ugni_module->endpoints);
OBJ_DESTRUCT(&ugni_module->eager_get_pending);
OBJ_DESTRUCT(&ugni_module->eager_get_pending_lock);
opal: rework mpool and rcache frameworks This commit rewrites both the mpool and rcache frameworks. Summary of changes: - Before this change a significant portion of the rcache functionality lived in mpool components. This meant that it was impossible to add a new memory pool to use with rdma networks (ugni, openib, etc) without duplicating the functionality of an existing mpool component. All the registration functionality has been removed from the mpool and placed in the rcache framework. - All registration cache mpools components (udreg, grdma, gpusm, rgpusm) have been changed to rcache components. rcaches are allocated and released in the same way mpool components were. - It is now valid to pass NULL as the resources argument when creating an rcache. At this time the gpusm and rgpusm components support this. All other rcache components require non-NULL resources. - A new mpool component has been added: hugepage. This component supports huge page allocations on linux. - Memory pools are now allocated using "hints". Each mpool component is queried with the hints and returns a priority. The current hints supported are NULL (uses posix_memalign/malloc), page_size=x (huge page mpool), and mpool=x. - The sm mpool has been moved to common/sm. This reflects that the sm mpool is specialized and not meant for any general allocations. This mpool may be moved back into the mpool framework if there is any objection. - The opal_free_list_init arguments have been updated. The unused0 argument is not used to pass in the registration cache module. The mpool registration flags are now rcache registration flags. - All components have been updated to make use of the new framework interfaces. As this commit makes significant changes to both the mpool and rcache frameworks both versions have been bumped to 3.0.0. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2015-11-02 12:07:08 -07:00
if (ugni_module->rcache) {
mca_rcache_base_module_destroy (ugni_module->rcache);
}
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
for (int i = 0 ; i < mca_btl_ugni_component.virtual_device_count ; ++i) {
mca_btl_ugni_device_fini (ugni_module->devices + i);
}
ugni_module->initialized = false;
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 00:47:28 +00:00
return OPAL_SUCCESS;
}
mca_btl_base_descriptor_t *
mca_btl_ugni_alloc(struct mca_btl_base_module_t *btl,
struct mca_btl_base_endpoint_t *endpoint,
uint8_t order, size_t size, uint32_t flags)
{
mca_btl_ugni_base_frag_t *frag = NULL;
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
/* do not allocate a fragment unless the wait list is relatively small. this
* reduces the potential for resource exhaustion. note the wait list only exists
* because we have no way to notify the sender that credits are available. */
if (OPAL_UNLIKELY(opal_list_get_size (&endpoint->frag_wait_list) > 32)) {
return NULL;
}
if (size <= mca_btl_ugni_component.smsg_max_data) {
frag = mca_btl_ugni_frag_alloc_smsg (endpoint);
} else if (size <= btl->btl_eager_limit) {
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
frag = mca_btl_ugni_frag_alloc_eager_send (endpoint);
}
if (OPAL_UNLIKELY(NULL == frag)) {
return NULL;
}
BTL_VERBOSE(("btl/ugni_module allocated frag of size: %u, flags: %x. frag = %p",
(unsigned int)size, flags, (void *) frag));
frag->base.des_flags = flags;
frag->base.order = order;
frag->base.des_segments = &frag->segments[1];
frag->base.des_segment_count = 1;
frag->segments[0].seg_addr.pval = NULL;
frag->segments[0].seg_len = 0;
frag->segments[1].seg_addr.pval = frag->base.super.ptr;
frag->segments[1].seg_len = size;
frag->flags = MCA_BTL_UGNI_FRAG_BUFFERED;
if (size > mca_btl_ugni_component.smsg_max_data) {
mca_btl_ugni_reg_t *registration;
frag->hdr_size = sizeof (frag->hdr.eager);
frag->flags |= MCA_BTL_UGNI_FRAG_EAGER | MCA_BTL_UGNI_FRAG_IGNORE;
registration = (mca_btl_ugni_reg_t *) frag->base.super.registration;
frag->hdr.eager.memory_handle = registration->handle;
} else {
frag->hdr_size = sizeof (frag->hdr.send);
}
return &frag->base;
}
static int
mca_btl_ugni_free (struct mca_btl_base_module_t *btl,
mca_btl_base_descriptor_t *des)
{
return mca_btl_ugni_frag_return ((mca_btl_ugni_base_frag_t *) des);
}
static struct mca_btl_base_descriptor_t *
mca_btl_ugni_prepare_src (struct mca_btl_base_module_t *btl,
mca_btl_base_endpoint_t *endpoint,
struct opal_convertor_t *convertor,
uint8_t order, size_t reserve, size_t *size,
uint32_t flags)
{
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
/* do not allocate a fragment unless the wait list is relatively small. this
* reduces the potential for resource exhaustion. note the wait list only exists
* because we have no way to notify the sender that credits are available. */
if (OPAL_UNLIKELY(opal_list_get_size (&endpoint->frag_wait_list) > 32)) {
return NULL;
}
return mca_btl_ugni_prepare_src_send (btl, endpoint, convertor,
order, reserve, size, flags);
}
static mca_btl_base_registration_handle_t *
mca_btl_ugni_register_mem (mca_btl_base_module_t *btl, mca_btl_base_endpoint_t *endpoint, void *base,
size_t size, uint32_t flags)
{
opal: rework mpool and rcache frameworks This commit rewrites both the mpool and rcache frameworks. Summary of changes: - Before this change a significant portion of the rcache functionality lived in mpool components. This meant that it was impossible to add a new memory pool to use with rdma networks (ugni, openib, etc) without duplicating the functionality of an existing mpool component. All the registration functionality has been removed from the mpool and placed in the rcache framework. - All registration cache mpools components (udreg, grdma, gpusm, rgpusm) have been changed to rcache components. rcaches are allocated and released in the same way mpool components were. - It is now valid to pass NULL as the resources argument when creating an rcache. At this time the gpusm and rgpusm components support this. All other rcache components require non-NULL resources. - A new mpool component has been added: hugepage. This component supports huge page allocations on linux. - Memory pools are now allocated using "hints". Each mpool component is queried with the hints and returns a priority. The current hints supported are NULL (uses posix_memalign/malloc), page_size=x (huge page mpool), and mpool=x. - The sm mpool has been moved to common/sm. This reflects that the sm mpool is specialized and not meant for any general allocations. This mpool may be moved back into the mpool framework if there is any objection. - The opal_free_list_init arguments have been updated. The unused0 argument is not used to pass in the registration cache module. The mpool registration flags are now rcache registration flags. - All components have been updated to make use of the new framework interfaces. As this commit makes significant changes to both the mpool and rcache frameworks both versions have been bumped to 3.0.0. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2015-11-02 12:07:08 -07:00
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *) btl;
mca_btl_ugni_reg_t *reg;
int access_flags = flags & MCA_BTL_REG_FLAG_ACCESS_ANY;
int rc;
opal: rework mpool and rcache frameworks This commit rewrites both the mpool and rcache frameworks. Summary of changes: - Before this change a significant portion of the rcache functionality lived in mpool components. This meant that it was impossible to add a new memory pool to use with rdma networks (ugni, openib, etc) without duplicating the functionality of an existing mpool component. All the registration functionality has been removed from the mpool and placed in the rcache framework. - All registration cache mpools components (udreg, grdma, gpusm, rgpusm) have been changed to rcache components. rcaches are allocated and released in the same way mpool components were. - It is now valid to pass NULL as the resources argument when creating an rcache. At this time the gpusm and rgpusm components support this. All other rcache components require non-NULL resources. - A new mpool component has been added: hugepage. This component supports huge page allocations on linux. - Memory pools are now allocated using "hints". Each mpool component is queried with the hints and returns a priority. The current hints supported are NULL (uses posix_memalign/malloc), page_size=x (huge page mpool), and mpool=x. - The sm mpool has been moved to common/sm. This reflects that the sm mpool is specialized and not meant for any general allocations. This mpool may be moved back into the mpool framework if there is any objection. - The opal_free_list_init arguments have been updated. The unused0 argument is not used to pass in the registration cache module. The mpool registration flags are now rcache registration flags. - All components have been updated to make use of the new framework interfaces. As this commit makes significant changes to both the mpool and rcache frameworks both versions have been bumped to 3.0.0. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2015-11-02 12:07:08 -07:00
rc = ugni_module->rcache->rcache_register (ugni_module->rcache, base, size, 0, access_flags,
(mca_rcache_base_registration_t **) &reg);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
return NULL;
}
return &reg->handle;
}
static int mca_btl_ugni_deregister_mem (mca_btl_base_module_t *btl, mca_btl_base_registration_handle_t *handle)
{
opal: rework mpool and rcache frameworks This commit rewrites both the mpool and rcache frameworks. Summary of changes: - Before this change a significant portion of the rcache functionality lived in mpool components. This meant that it was impossible to add a new memory pool to use with rdma networks (ugni, openib, etc) without duplicating the functionality of an existing mpool component. All the registration functionality has been removed from the mpool and placed in the rcache framework. - All registration cache mpools components (udreg, grdma, gpusm, rgpusm) have been changed to rcache components. rcaches are allocated and released in the same way mpool components were. - It is now valid to pass NULL as the resources argument when creating an rcache. At this time the gpusm and rgpusm components support this. All other rcache components require non-NULL resources. - A new mpool component has been added: hugepage. This component supports huge page allocations on linux. - Memory pools are now allocated using "hints". Each mpool component is queried with the hints and returns a priority. The current hints supported are NULL (uses posix_memalign/malloc), page_size=x (huge page mpool), and mpool=x. - The sm mpool has been moved to common/sm. This reflects that the sm mpool is specialized and not meant for any general allocations. This mpool may be moved back into the mpool framework if there is any objection. - The opal_free_list_init arguments have been updated. The unused0 argument is not used to pass in the registration cache module. The mpool registration flags are now rcache registration flags. - All components have been updated to make use of the new framework interfaces. As this commit makes significant changes to both the mpool and rcache frameworks both versions have been bumped to 3.0.0. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2015-11-02 12:07:08 -07:00
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *) btl;
mca_btl_ugni_reg_t *reg =
(mca_btl_ugni_reg_t *)((intptr_t) handle - offsetof (mca_btl_ugni_reg_t, handle));
opal: rework mpool and rcache frameworks This commit rewrites both the mpool and rcache frameworks. Summary of changes: - Before this change a significant portion of the rcache functionality lived in mpool components. This meant that it was impossible to add a new memory pool to use with rdma networks (ugni, openib, etc) without duplicating the functionality of an existing mpool component. All the registration functionality has been removed from the mpool and placed in the rcache framework. - All registration cache mpools components (udreg, grdma, gpusm, rgpusm) have been changed to rcache components. rcaches are allocated and released in the same way mpool components were. - It is now valid to pass NULL as the resources argument when creating an rcache. At this time the gpusm and rgpusm components support this. All other rcache components require non-NULL resources. - A new mpool component has been added: hugepage. This component supports huge page allocations on linux. - Memory pools are now allocated using "hints". Each mpool component is queried with the hints and returns a priority. The current hints supported are NULL (uses posix_memalign/malloc), page_size=x (huge page mpool), and mpool=x. - The sm mpool has been moved to common/sm. This reflects that the sm mpool is specialized and not meant for any general allocations. This mpool may be moved back into the mpool framework if there is any objection. - The opal_free_list_init arguments have been updated. The unused0 argument is not used to pass in the registration cache module. The mpool registration flags are now rcache registration flags. - All components have been updated to make use of the new framework interfaces. As this commit makes significant changes to both the mpool and rcache frameworks both versions have been bumped to 3.0.0. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2015-11-02 12:07:08 -07:00
(void) ugni_module->rcache->rcache_deregister (ugni_module->rcache, &reg->base);
return OPAL_SUCCESS;
}
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
int mca_btl_ugni_event_fatal_error (gni_return_t grc, gni_cq_entry_t event_data)
{
/* combined error check for get event and get completed. we might miss exactly
* what happened but it is unrecoverable anyway. fwiw, this error path has
* never been seen in production. */
if (GNI_CQ_OVERRUN(event_data)) {
/* TODO -- need to handle overrun -- how do we do this without an event?
will the event eventually come back? Ask Cray */
BTL_ERROR(("CQ overrun detected in RDMA event data. can not recover"));
} else {
BTL_ERROR(("Error in GNI_GetComplete %s", gni_err_str[grc]));
}
return mca_btl_rc_ugni_to_opal (grc);
}
int mca_btl_ugni_device_handle_event_error (mca_btl_ugni_rdma_desc_t *rdma_desc, gni_cq_entry_t event_data)
{
mca_btl_ugni_device_t *device = rdma_desc->device;
uint32_t recoverable = 1;
(void) GNI_CqErrorRecoverable (event_data, &recoverable);
if (OPAL_UNLIKELY(++rdma_desc->tries >= mca_btl_ugni_component.rdma_max_retries || !recoverable)) {
char char_buffer[1024];
GNI_CqErrorStr (event_data, char_buffer, sizeof (char_buffer));
BTL_ERROR(("giving up on desciptor %p, recoverable %d: %s", (void *) rdma_desc, recoverable, char_buffer));
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
return OPAL_ERROR;
}
return _mca_btl_ugni_repost_rdma_desc_device (device, rdma_desc);
}