1
1
openmpi/opal/mca/btl/ugni/btl_ugni.h

618 строки
20 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 22:57:21 +03:00
* Copyright (c) 2011-2018 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2011 UT-Battelle, LLC. All rights reserved.
* Copyright (c) 2014 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/*
* The ugni btl is implemented with native Cray Gemini.
*
* Known issues with ugni:
* -
*/
#ifndef MCA_BTL_UGNI_H
#define MCA_BTL_UGNI_H
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 04:47:28 +04:00
#include "opal_config.h"
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 04:47:28 +04:00
#include "opal/mca/mpool/mpool.h"
#include "opal/mca/mpool/base/base.h"
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 22:07:08 +03:00
#include "opal/mca/rcache/base/base.h"
#include "opal/mca/rcache/udreg/rcache_udreg.h"
#include "opal/util/output.h"
#include "opal_stdint.h"
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 04:47:28 +04:00
#include "opal/mca/btl/btl.h"
#include "opal/mca/btl/base/base.h"
#include "opal/mca/btl/base/btl_base_error.h"
#include "opal/class/opal_hash_table.h"
#include "opal/class/opal_free_list.h"
#include <errno.h>
#include <stdint.h>
#include <sys/types.h>
#include <assert.h>
#include <sys/time.h>
#include <gni_pub.h>
/* datagram message ids */
#define MCA_BTL_UGNI_CONNECT_WILDCARD_ID 0x0000000000000000ull
#define MCA_BTL_UGNI_CONNECT_DIRECTED_ID 0x8000000000000000ull
#define MCA_BTL_UGNI_DATAGRAM_MASK 0x8000000000000000ull
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>
2017-03-13 07:37:35 +03:00
/** maximum number of supported virtual devices */
#define MCA_BTL_UGNI_MAX_DEV_HANDLES 128
/** number of rdma completion queue items to remove per progress loop */
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 22:57:21 +03:00
#define MCA_BTL_UGNI_COMPLETIONS_PER_LOOP 32
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>
2017-03-13 07:37:35 +03:00
/** how often to check for connection requests */
#define MCA_BTL_UGNI_CONNECT_USEC 10
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>
2017-03-13 07:37:35 +03:00
/**
* Modex data
*/
struct mca_btl_ugni_modex_t {
/** GNI NIC address */
uint32_t addr;
/** CDM identifier (base) */
int id;
};
typedef struct mca_btl_ugni_modex_t mca_btl_ugni_modex_t;
/* ompi and smsg endpoint attributes */
typedef struct mca_btl_ugni_endpoint_attr_t {
opal_process_name_t proc_name;
uint32_t index;
gni_smsg_attr_t smsg_attr;
gni_mem_handle_t rmt_irq_mem_hndl;
} mca_btl_ugni_endpoint_attr_t;
enum {
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 22:07:08 +03:00
MCA_BTL_UGNI_RCACHE_UDREG,
MCA_BTL_UGNI_RCACHE_GRDMA
};
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>
2017-03-13 07:37:35 +03:00
enum mca_btl_ugni_free_list_id_t {
/* eager fragment list (registered) */
MCA_BTL_UGNI_LIST_EAGER_SEND,
MCA_BTL_UGNI_LIST_EAGER_RECV,
/* SMSG fragment list (unregistered) */
MCA_BTL_UGNI_LIST_SMSG,
/* RDMA fragment list */
MCA_BTL_UGNI_LIST_RDMA,
MCA_BTL_UGNI_LIST_RDMA_INT,
MCA_BTL_UGNI_LIST_MAX,
};
struct mca_btl_ugni_cq_t {
/** ugni CQ handle */
gni_cq_handle_t gni_handle;
/** number of completions expected on the CQ */
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 22:57:21 +03:00
volatile int32_t active_operations;
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>
2017-03-13 07:37:35 +03:00
};
typedef struct mca_btl_ugni_cq_t mca_btl_ugni_cq_t;
/**
* GNI virtual device
*/
struct mca_btl_ugni_device_t {
/** Communication domain handle */
gni_cdm_handle_t dev_cd_handle;
/** protection for ugni access */
volatile int32_t lock;
/** Index of device in module devices array */
int dev_index;
/** number of SMSG connections */
opal_atomic_int32_t smsg_connections;
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>
2017-03-13 07:37:35 +03:00
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 22:57:21 +03:00
/** boolean indicating that the device was recently flushed */
volatile bool flushed;
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>
2017-03-13 07:37:35 +03:00
/** uGNI device handle */
gni_nic_handle_t dev_handle;
/** uGNI rdma completion queue */
mca_btl_ugni_cq_t dev_rdma_local_cq;
/** local rdma completion queue (async) */
mca_btl_ugni_cq_t dev_rdma_local_irq_cq;
/** local SMSG completion queue */
mca_btl_ugni_cq_t dev_smsg_local_cq;
/** IRQ memory handle for this device */
gni_mem_handle_t smsg_irq_mhndl;
/** RDMA endpoint free list */
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 22:57:21 +03:00
opal_free_list_t rdma_descs;
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>
2017-03-13 07:37:35 +03:00
};
typedef struct mca_btl_ugni_device_t mca_btl_ugni_device_t;
typedef intptr_t (*mca_btl_ugni_device_serialize_fn_t) (mca_btl_ugni_device_t *device, void *arg);
typedef struct mca_btl_ugni_module_t {
mca_btl_base_module_t super;
bool 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>
2017-03-13 07:37:35 +03:00
mca_btl_ugni_device_t devices[MCA_BTL_UGNI_MAX_DEV_HANDLES];
opal_mutex_t endpoint_lock;
size_t endpoint_count;
opal_pointer_array_t endpoints;
opal_hash_table_t id_to_endpoint;
/* lock for this list */
opal_mutex_t failed_frags_lock;
/** rdma frags waiting to be reposted */
opal_list_t failed_frags;
/** lock for the eager_get_pending list */
opal_mutex_t eager_get_pending_lock;
opal_list_t eager_get_pending;
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 22:07:08 +03:00
mca_mpool_base_module_t *mpool;
opal_free_list_t smsg_mboxes;
gni_ep_handle_t wildcard_ep;
struct mca_btl_base_endpoint_t *local_ep;
opal_atomic_int32_t active_datagrams;
opal_event_t connection_event;
struct mca_btl_ugni_endpoint_attr_t wc_remote_attr, wc_local_attr;
gni_cq_handle_t smsg_remote_cq;
gni_cq_handle_t smsg_remote_irq_cq;
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>
2017-03-13 07:37:35 +03:00
/** fragment free lists (see enum mca_btl_ugni_free_list_id_t) */
opal_free_list_t frags_lists[MCA_BTL_UGNI_LIST_MAX];
/* lock for this list */
opal_mutex_t ep_wait_list_lock;
/* endpoints waiting on credits */
opal_list_t ep_wait_list;
/* fragment id bounce buffer (smsg msg ids are only 32 bits) */
opal_pointer_array_t pending_smsg_frags_bb;
int32_t reg_max;
opal_atomic_int32_t reg_count;
/* used to calculate the fraction of registered memory resources
* this rank should be limited too */
int nlocal_procs;
opal_atomic_int32_t active_rdma_count;
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 22:07:08 +03:00
mca_rcache_base_module_t *rcache;
} mca_btl_ugni_module_t;
typedef struct mca_btl_ugni_component_t {
/* base BTL component */
mca_btl_base_component_3_0_0_t super;
/* maximum supported btls. hardcoded to 1 for now */
uint32_t ugni_max_btls;
/* Maximum number of entries a completion queue can hold */
uint32_t remote_cq_size;
uint32_t local_cq_size;
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 22:57:21 +03:00
uint32_t local_rdma_cq_size;
/* There is a hardware limitation that hurts BTE performance
* if we submit too many BTE requests. This acts as a throttle. */
int32_t active_rdma_threshold;
/* number of ugni modules */
uint32_t ugni_num_btls;
/* ugni modules */
mca_btl_ugni_module_t *modules;
size_t smsg_max_data;
/* After this message size switch to BTE protocols */
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 22:57:21 +03:00
long int ugni_fma_limit;
/** FMA switchover for get */
long int ugni_fma_get_limit;
/** FMA switchover for put */
long int ugni_fma_put_limit;
#if OPAL_C_HAVE__THREAD_LOCAL
bool bind_threads_to_devices;
#endif
/* Switch to get when sending above this size */
size_t ugni_smsg_limit;
/* RDMA/SMSG free list settings */
int ugni_free_list_num;
int ugni_free_list_max;
int ugni_free_list_inc;
/* eager free list settings */
int ugni_eager_num;
int ugni_eager_max;
int ugni_eager_inc;
int smsg_max_retries;
/* number of times to retry a post */
int rdma_max_retries;
/* Maximum number of outstanding eager messages */
int smsg_max_credits;
/* mailbox size (computed) */
int smsg_mbox_size;
/* Maximum number of memory registrations per process */
int max_mem_reg;
/* Page size to use for SMSG allocations (udreg mpool) */
unsigned int smsg_page_size;
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 22:07:08 +03:00
/* rcache type (grdma or udreg) */
int rcache_type;
/* memory pool hints */
char *mpool_hints;
/* Number of mailboxes to allocate in each block */
unsigned int mbox_increment;
/* Indicate whether progress thread requested */
bool progress_thread_requested;
/* Indicate whether progress thread allowed */
bool progress_thread_enabled;
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>
2017-03-13 07:37:35 +03:00
/** Number of ugni device contexts to create per GNI device */
int virtual_device_count;
/** Protection tag */
uint8_t ptag;
/** Unique id for this process assigned by the system */
uint32_t cookie;
/** Starting value of communication identifier */
uint32_t cdm_id_base;
/** GNI CDM flags */
uint32_t cdm_flags;
/** NIC address */
uint32_t dev_addr;
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 22:57:21 +03:00
/** MCA variable identifier for the cdm_flags variable */
int cdm_flags_id;
} mca_btl_ugni_component_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>
2017-03-13 07:37:35 +03:00
/* Global structures */
OPAL_MODULE_DECLSPEC extern mca_btl_ugni_component_t mca_btl_ugni_component;
OPAL_MODULE_DECLSPEC extern mca_btl_ugni_module_t mca_btl_ugni_module;
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 22:57:21 +03:00
static inline uint32_t mca_btl_ugni_ep_get_device_index (mca_btl_ugni_module_t *ugni_module)
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>
2017-03-13 07:37:35 +03:00
{
static volatile uint32_t device_index = (uint32_t) 0;
/* don't really care if the device index is atomically updated */
return opal_atomic_fetch_add_32 ((volatile int32_t *) &device_index, 1) % mca_btl_ugni_component.virtual_device_count;
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 22:57:21 +03:00
}
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>
2017-03-13 07:37:35 +03:00
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 22:57:21 +03:00
/**
* Get a virtual device for communication
*/
static inline mca_btl_ugni_device_t *mca_btl_ugni_ep_get_device (mca_btl_ugni_module_t *ugni_module)
{
return ugni_module->devices + mca_btl_ugni_ep_get_device_index (ugni_module);
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>
2017-03-13 07:37:35 +03:00
}
static inline int mca_btl_rc_ugni_to_opal (gni_return_t rc)
{
static int codes[] = {OPAL_SUCCESS,
OPAL_ERR_RESOURCE_BUSY,
OPAL_ERR_BAD_PARAM,
OPAL_ERR_OUT_OF_RESOURCE,
OPAL_ERR_TIMEOUT,
OPAL_ERR_PERM,
OPAL_ERROR,
OPAL_ERR_BAD_PARAM,
OPAL_ERR_BAD_PARAM,
OPAL_ERR_NOT_FOUND,
OPAL_ERR_VALUE_OUT_OF_BOUNDS,
OPAL_ERROR,
OPAL_ERR_NOT_SUPPORTED,
OPAL_ERR_OUT_OF_RESOURCE};
return codes[rc];
}
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 22:57:21 +03:00
int mca_btl_ugni_flush (mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint);
/**
2015-06-24 06:59:57 +03:00
* BML->BTL notification of change in the process list.
*
* location: btl_ugni_add_procs.c
*
* @param btl (IN) BTL module
* @param nprocs (IN) Number of processes
* @param procs (IN) Array of processes
* @param endpoint (OUT) Array of mca_btl_base_endpoint_t structures by BTL.
* @param reachable (OUT) Bitmask indicating set of peer processes that are reachable by this BTL.
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 04:47:28 +04:00
* @return OPAL_SUCCESS or error status on failure.
*/
int
mca_btl_ugni_add_procs (struct mca_btl_base_module_t* btl,
size_t nprocs,
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 04:47:28 +04:00
struct opal_proc_t **procs,
struct mca_btl_base_endpoint_t **peers,
opal_bitmap_t *reachable);
/**
* Notification of change to the process list.
*
* location: btl_ugni_add_procs.c
*
* @param btl (IN) BTL module
* @param nprocs (IN) Number of processes
* @param proc (IN) Set of processes
* @param peer (IN) Set of peer addressing information.
* @return Status indicating if cleanup was successful
*/
int
mca_btl_ugni_del_procs (struct mca_btl_base_module_t *btl,
size_t nprocs,
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 04:47:28 +04:00
struct opal_proc_t **procs,
struct mca_btl_base_endpoint_t **peers);
struct mca_btl_base_endpoint_t *mca_btl_ugni_get_ep (struct mca_btl_base_module_t *module, opal_proc_t *proc);
/**
* Initiate an asynchronous send.
*
* location: btl_ugni_send.c
*
* @param btl (IN) BTL module
* @param endpoint (IN) BTL addressing information
* @param descriptor (IN) Description of the data to be transfered
* @param tag (IN) The tag value used to notify the peer.
*/
int
mca_btl_ugni_send (struct mca_btl_base_module_t *btl,
struct mca_btl_base_endpoint_t *btl_peer,
struct mca_btl_base_descriptor_t *descriptor,
mca_btl_base_tag_t tag);
/**
2015-06-24 06:59:57 +03:00
* Initiate an immediate blocking send.
*
* location: btl_ugni_sendi.c
*
* @param btl (IN) BTL module
* @param endpoint (IN) BTL addressing information
* @param convertor (IN) Data type convertor
* @param header (IN) Pointer to header.
* @param header_size (IN) Size of header.
* @param payload_size (IN) Size of payload (from convertor).
* @param order (IN) The ordering tag (may be MCA_BTL_NO_ORDER)
* @param flags (IN) Flags.
* @param tag (IN) The tag value used to notify the peer.
* @param descriptor (OUT) The descriptor to be returned unable to be sent immediately
*/
int
mca_btl_ugni_sendi (struct mca_btl_base_module_t *btl,
struct mca_btl_base_endpoint_t *endpoint,
struct opal_convertor_t *convertor,
void *header, size_t header_size,
size_t payload_size, uint8_t order,
uint32_t flags, mca_btl_base_tag_t tag,
mca_btl_base_descriptor_t **descriptor);
int mca_btl_ugni_get (mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint, void *local_address,
uint64_t remote_address, mca_btl_base_registration_handle_t *local_handle,
mca_btl_base_registration_handle_t *remote_handle, size_t size, int flags,
int order, mca_btl_base_rdma_completion_fn_t cbfunc, void *cbcontext, void *cbdata);
int mca_btl_ugni_put (mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint, void *local_address,
uint64_t remote_address, mca_btl_base_registration_handle_t *local_handle,
mca_btl_base_registration_handle_t *remote_handle, size_t size, int flags,
int order, mca_btl_base_rdma_completion_fn_t cbfunc, void *cbcontext, void *cbdata);
int mca_btl_ugni_aop (struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint,
uint64_t remote_address, mca_btl_base_registration_handle_t *remote_handle,
mca_btl_base_atomic_op_t op, uint64_t operand, int flags, int order,
mca_btl_base_rdma_completion_fn_t cbfunc, void *cbcontext, void *cbdata);
int mca_btl_ugni_afop (struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint,
void *local_address, uint64_t remote_address, mca_btl_base_registration_handle_t *local_handle,
mca_btl_base_registration_handle_t *remote_handle, mca_btl_base_atomic_op_t op,
uint64_t operand, int flags, int order, mca_btl_base_rdma_completion_fn_t cbfunc,
void *cbcontext, void *cbdata);
int mca_btl_ugni_acswap (struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint,
void *local_address, uint64_t remote_address, mca_btl_base_registration_handle_t *local_handle,
mca_btl_base_registration_handle_t *remote_handle, uint64_t compare, uint64_t value,
int flags, int order, mca_btl_base_rdma_completion_fn_t cbfunc, void *cbcontext, void *cbdata);
int mca_btl_ugni_progress_send_wait_list (struct mca_btl_base_endpoint_t *endpoint);
int mca_btl_ugni_progress_datagram (mca_btl_ugni_device_t *device);
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);
struct mca_btl_base_registration_handle_t {
/** uGNI memory handle */
gni_mem_handle_t gni_handle;
};
typedef struct mca_btl_ugni_reg_t {
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 22:07:08 +03:00
mca_rcache_base_registration_t base;
mca_btl_base_registration_handle_t handle;
} mca_btl_ugni_reg_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>
2017-03-13 07:37:35 +03:00
/**
* Initialize uGNI support.
*/
int mca_btl_ugni_init (void);
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>
2017-03-13 07:37:35 +03:00
/**
* Finalize uGNI support.
*/
int mca_btl_ugni_fini (void);
int mca_btl_ugni_module_init (mca_btl_ugni_module_t *ugni_module);
/**
* Intialize a virtual device for device index 0.
*
* @param[inout] device Device to initialize
* @param[in] virtual_device_id Virtual device identified (up to max handles)
*/
int mca_btl_ugni_device_init (mca_btl_ugni_device_t *device, int virtual_device_id);
/**
* Finalize a virtual device.
*
* @param[in] device Device to finalize
*/
int mca_btl_ugni_device_fini (mca_btl_ugni_device_t *dev);
/* Get a unique 64-bit id for the process name */
static inline uint64_t mca_btl_ugni_proc_name_to_id (opal_process_name_t name) {
/* Throw away the top bit of the jobid for the datagram type */
return ((uint64_t) (name.jobid & 0x7fffffff) << 32 | name.vpid);
}
2014-12-05 02:18:16 +03:00
int mca_btl_ugni_spawn_progress_thread(struct mca_btl_base_module_t* btl);
int mca_btl_ugni_kill_progress_thread(void);
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 22:57:21 +03:00
struct mca_btl_ugni_post_descriptor_t;
void btl_ugni_dump_post_desc (struct mca_btl_ugni_post_descriptor_t *desc);
struct mca_btl_ugni_post_descriptor_t;
void mca_btl_ugni_handle_rdma_completions (mca_btl_ugni_module_t *ugni_module, mca_btl_ugni_device_t *device,
struct mca_btl_ugni_post_descriptor_t *post_desc, const int count);
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>
2017-03-13 07:37:35 +03:00
/**
* Try to lock a uGNI device for exclusive access
*/
static inline int mca_btl_ugni_device_trylock (mca_btl_ugni_device_t *device)
{
/* checking the lock non-atomically first can reduce the number of
* unnecessary atomic operations. */
return (device->lock || opal_atomic_swap_32 (&device->lock, 1));
}
/**
* Lock a uGNI device for exclusive access
*/
static inline void mca_btl_ugni_device_lock (mca_btl_ugni_device_t *device)
{
while (mca_btl_ugni_device_trylock (device));
}
/**
* Release exclusive access to the device
*/
static inline void mca_btl_ugni_device_unlock (mca_btl_ugni_device_t *device)
{
opal_atomic_wmb ();
device->lock = 0;
}
/**
* Serialize an operation on a uGNI device
*
* @params[in] device ugni device
* @params[in] fn function to serialize
* @params[in] arg function argument
*/
static inline intptr_t mca_btl_ugni_device_serialize (mca_btl_ugni_device_t *device,
mca_btl_ugni_device_serialize_fn_t fn, void *arg)
{
intptr_t rc;
if (!opal_using_threads ()) {
return fn (device, arg);
}
/* NTH: for now the device is just protected by a spin lock but this will change in the future */
mca_btl_ugni_device_lock (device);
rc = fn (device, arg);
mca_btl_ugni_device_unlock (device);
return rc;
}
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 22:57:21 +03:00
static inline intptr_t mca_btl_ugni_device_serialize_any (mca_btl_ugni_module_t *ugni_module,
mca_btl_ugni_device_serialize_fn_t fn, void *arg)
{
mca_btl_ugni_device_t *device;
intptr_t rc;
if (!opal_using_threads ()) {
return fn (ugni_module->devices, arg);
}
#if OPAL_C_HAVE__THREAD_LOCAL
if (mca_btl_ugni_component.bind_threads_to_devices) {
/* NTH: if we have C11 _Thread_local just go ahead and assign the devices round-robin to each
* thread. in testing this give much better performance than just picking any device */
static _Thread_local mca_btl_ugni_device_t *device_local = NULL;
device = device_local;
if (OPAL_UNLIKELY(NULL == device)) {
/* assign device contexts round-robin */
device_local = device = mca_btl_ugni_ep_get_device (ugni_module);
}
mca_btl_ugni_device_lock (device);
} else {
#endif
/* get the next starting index */
uint32_t device_index = mca_btl_ugni_ep_get_device_index (ugni_module);
const int device_count = mca_btl_ugni_component.virtual_device_count;
for (int i = 0 ; i < device_count ; ++i) {
device = ugni_module->devices + ((device_index + i) % device_count);
if (!mca_btl_ugni_device_trylock (device)) {
break;
}
device = NULL;
}
if (NULL == device) {
device = mca_btl_ugni_ep_get_device (ugni_module);
mca_btl_ugni_device_lock (device);
}
#if OPAL_C_HAVE__THREAD_LOCAL
}
#endif
rc = fn (device, arg);
mca_btl_ugni_device_unlock (device);
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>
2017-03-13 07:37:35 +03:00
/** Number of times the progress thread has woken up */
extern unsigned int mca_btl_ugni_progress_thread_wakeups;
#endif