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

299 строки
11 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
* Copyright (c) 2011-2017 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2011 UT-Battelle, LLC. All rights reserved.
* Copyright (c) 2013 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#if !defined(MCA_BTL_UGNI_FRAG_H)
#define MCA_BTL_UGNI_FRAG_H
#include "btl_ugni.h"
#include "btl_ugni_endpoint.h"
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
#include <string.h>
typedef struct mca_btl_ugni_send_frag_hdr_t {
uint32_t lag;
} mca_btl_ugni_send_frag_hdr_t;
typedef struct mca_btl_ugni_send_ex_frag_hdr_t {
mca_btl_ugni_send_frag_hdr_t send;
uint8_t pml_header[128];
} mca_btl_ugni_send_ex_frag_hdr_t;
typedef struct mca_btl_ugni_rdma_frag_hdr_t {
void *ctx;
} mca_btl_ugni_rdma_frag_hdr_t;
typedef struct mca_btl_ugni_eager_frag_hdr_t {
mca_btl_ugni_send_frag_hdr_t send;
uint32_t size;
uint64_t address;
mca_btl_base_registration_handle_t memory_handle;
void *ctx;
} mca_btl_ugni_eager_frag_hdr_t;
typedef struct mca_btl_ugni_eager_ex_frag_hdr_t {
mca_btl_ugni_eager_frag_hdr_t eager;
uint8_t pml_header[128];
} mca_btl_ugni_eager_ex_frag_hdr_t;
typedef union mca_btl_ugni_frag_hdr_t {
mca_btl_ugni_send_frag_hdr_t send;
mca_btl_ugni_send_ex_frag_hdr_t send_ex;
mca_btl_ugni_rdma_frag_hdr_t rdma;
mca_btl_ugni_eager_frag_hdr_t eager;
mca_btl_ugni_eager_ex_frag_hdr_t eager_ex;
} mca_btl_ugni_frag_hdr_t;
enum {
MCA_BTL_UGNI_FRAG_BUFFERED = 1, /* frag data is buffered */
MCA_BTL_UGNI_FRAG_COMPLETE = 2, /* smsg complete for frag */
MCA_BTL_UGNI_FRAG_EAGER = 4, /* eager get frag */
MCA_BTL_UGNI_FRAG_IGNORE = 8, /* ignore local smsg completion */
MCA_BTL_UGNI_FRAG_SMSG_COMPLETE = 16, /* SMSG has completed for this message */
MCA_BTL_UGNI_FRAG_RESPONSE = 32,
};
struct mca_btl_ugni_base_frag_t;
typedef struct mca_btl_ugni_base_frag_t {
mca_btl_base_descriptor_t base;
opal_atomic_int32_t ref_cnt;
uint32_t msg_id;
uint16_t hdr_size;
uint16_t flags;
mca_btl_ugni_frag_hdr_t hdr;
mca_btl_base_segment_t segments[2];
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
gni_post_descriptor_t post_desc;
mca_btl_base_endpoint_t *endpoint;
mca_btl_ugni_reg_t *registration;
opal_free_list_t *my_list;
mca_btl_base_registration_handle_t memory_handle;
} mca_btl_ugni_base_frag_t;
typedef struct mca_btl_ugni_base_frag_t mca_btl_ugni_smsg_frag_t;
typedef struct mca_btl_ugni_base_frag_t mca_btl_ugni_rdma_frag_t;
typedef struct mca_btl_ugni_base_frag_t mca_btl_ugni_eager_frag_t;
typedef struct mca_btl_ugni_post_descriptor_t {
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
/** endpoint currently associated with this desctiptor */
mca_btl_base_endpoint_t *endpoint;
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
/** local memory handle (for callback) */
mca_btl_base_registration_handle_t *local_handle;
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
/** currently associated completion queue */
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
mca_btl_ugni_cq_t *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 12:57:21 -07:00
/** user callback function */
mca_btl_base_rdma_completion_fn_t cbfunc;
/** user callback data */
void *cbdata;
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
/** user callback context */
void *ctx;
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
/** opal status of this descriptor. filled in by
* mca_btl_ugni_cq_get_completed_desc_device() */
int rc;
/** true if posted with the BTE. false if FMA. this is used as part
* of the BTE throttling code. */
bool use_bte;
/** uGNI library post descriptor. this is last in this structure
* to try to keep it hot in the cache after copying this descriptor
* into the allocated descritor. (post follows almost immediately
* after allocate. */
gni_post_descriptor_t gni_desc;
} mca_btl_ugni_post_descriptor_t;
OBJ_CLASS_DECLARATION(mca_btl_ugni_post_descriptor_t);
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
typedef struct mca_btl_ugni_rdma_desc_t {
opal_free_list_item_t super;
mca_btl_ugni_post_descriptor_t btl_ugni_desc;
mca_btl_ugni_device_t *device;
gni_ep_handle_t gni_handle;
int tries;
} mca_btl_ugni_rdma_desc_t;
OBJ_CLASS_DECLARATION(mca_btl_ugni_rdma_desc_t);
#define MCA_BTL_UGNI_GNI_DESC_TO_RDMA_DESC(desc) \
((mca_btl_ugni_rdma_desc_t *) ((uintptr_t)(desc) - offsetof (mca_btl_ugni_rdma_desc_t, btl_ugni_desc) - offsetof (mca_btl_ugni_post_descriptor_t, gni_desc)))
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
/**
* Initialize a RDMA descriptor
*
* @param[in] item free list item (must be of class mca_btl_ugni_rdma_desc_t)
* @param[in] ctx pointer to ugni device context
*
* This function initializes a mca_btl_ugni_rdma_desc_t for use. It allocates
* resources from the ugni library. This must be called before a RDMA
* descriptor can be used. Usually this is passed as an argument to
* opal_free_list_init().
*/
int mca_btl_ugni_rdma_desc_init (opal_free_list_item_t *item, void *ctx);
/**
* @brief get an endpoint handle from a device's free list
*
* @param[in] ep btl endpoint
* @param[in] device btl device to use
* @param[in] use_bte whether this descriptor will be used with the BTE
*
* This function MUST be called with the device lock held. This was done over using
* the atomic free list to avoid unnecessary atomics in the critical path.
*/
static inline mca_btl_ugni_rdma_desc_t *
mca_btl_ugni_alloc_rdma_desc (mca_btl_ugni_device_t *device, mca_btl_ugni_post_descriptor_t *ugni_desc, const bool use_bte)
{
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
mca_btl_ugni_rdma_desc_t *desc = (mca_btl_ugni_rdma_desc_t *) opal_free_list_get_st (&device->rdma_descs);
mca_btl_ugni_endpoint_t *ep = ugni_desc->endpoint;
gni_return_t grc;
if (OPAL_LIKELY(NULL != desc)) {
grc = GNI_EpBind (desc->gni_handle, ep->ep_rem_addr, ep->ep_rem_id | device->dev_index);
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc)) {
opal_free_list_return_st (&device->rdma_descs, &desc->super);
return NULL;
}
desc->device = device;
desc->tries = 0;
desc->btl_ugni_desc = *ugni_desc;
desc->btl_ugni_desc.use_bte = use_bte;
}
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
return desc;
}
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
static inline void mca_btl_ugni_return_rdma_desc (mca_btl_ugni_rdma_desc_t *desc)
{
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
(void) GNI_EpUnbind (desc->gni_handle);
opal_free_list_return_st (&desc->device->rdma_descs, &desc->super);
}
static inline void mca_btl_ugni_post_desc_complete (mca_btl_ugni_module_t *module, mca_btl_ugni_post_descriptor_t *desc, int rc)
{
BTL_VERBOSE(("RDMA/FMA/ATOMIC operation complete for post descriptor %p. rc = %d", (void *) desc, rc));
if (NULL != desc->cbfunc) {
/* call the user's callback function */
btl/ugni: improve multi-threaded RDMA performance This commit improves the injection rate and latency for RDMA operations. This is done by the following improvements: - If C11's _Thread_local keyword is available then always use the same virtual device index for the same thread when using RDMA. If the keyword is not available then attempt to use any device that isn't already in use. The binding support is enabled by default but can be disabled via the btl_ugni_bind_devices MCA variable. - When posting FMA and RDMA operations always attempt to reap completions after posting the operation. This allows us to better balance the work of reaping completions across all application threads. - Limit the total number of outstanding BTE transactions. This fixes a performance bug when using many threads. - Split out RDMA and local SMSG completion queue sizes. The RDMA queue size is better tuned for performance with RMA-MT. - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA variable is deprecated. The new variable names are: btl_ugni_fma_put_limit and btl_ugni_fma_get_limit. - Change how post descriptors are handled. They are no longer allocated seperately from the RDMA endpoints. - Some cleanup to move error code out of the critical path. - Disable the FMA sharing flag on the CDM when we detect that there should be enough FMA descriptors for the number of virtual devices we plan will create. If the user sets this flag we will not unset it. This change should improve the small-message RMA performance by ~ 10%. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-02-15 12:57:21 -07:00
desc->cbfunc (&module->super, desc->endpoint, (void *)(intptr_t) desc->gni_desc.local_addr,
desc->local_handle, desc->ctx, desc->cbdata, rc);
}
}
OBJ_CLASS_DECLARATION(mca_btl_ugni_smsg_frag_t);
OBJ_CLASS_DECLARATION(mca_btl_ugni_rdma_frag_t);
OBJ_CLASS_DECLARATION(mca_btl_ugni_eager_frag_t);
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
int mca_btl_ugni_frag_init (mca_btl_ugni_base_frag_t *frag, void *id);
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
static inline mca_btl_ugni_base_frag_t *mca_btl_ugni_frag_alloc (mca_btl_base_endpoint_t *ep,
opal_free_list_t *list)
{
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
mca_btl_ugni_base_frag_t *frag = (mca_btl_ugni_base_frag_t *) opal_free_list_get (list);
if (OPAL_LIKELY(NULL != frag)) {
frag->endpoint = ep;
frag->ref_cnt = 1;
}
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
return frag;
}
static inline int mca_btl_ugni_frag_return (mca_btl_ugni_base_frag_t *frag)
{
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (frag->endpoint);
if (frag->registration) {
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
ugni_module->rcache->rcache_deregister (ugni_module->rcache,
(mca_rcache_base_registration_t *) frag->registration);
frag->registration = NULL;
}
frag->flags = 0;
opal_free_list_return (frag->my_list, (opal_free_list_item_t *) frag);
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-) WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic. This commit was SVN r32317.
2014-07-26 00:47:28 +00:00
return OPAL_SUCCESS;
}
static inline bool mca_btl_ugni_frag_del_ref (mca_btl_ugni_base_frag_t *frag, int rc) {
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (frag->endpoint);
/* save the descriptor flags since the callback is allowed to free the frag */
int des_flags = frag->base.des_flags;
int32_t ref_cnt;
opal_atomic_mb ();
ref_cnt = OPAL_THREAD_ADD_FETCH32(&frag->ref_cnt, -1);
if (ref_cnt) {
assert (ref_cnt > 0);
return false;
}
/* call callback if specified */
if (des_flags & MCA_BTL_DES_SEND_ALWAYS_CALLBACK) {
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
frag->base.des_cbfunc(&ugni_module->super, frag->endpoint, &frag->base, rc);
}
if (des_flags & MCA_BTL_DES_FLAGS_BTL_OWNERSHIP) {
mca_btl_ugni_frag_return (frag);
}
return true;
}
static inline void mca_btl_ugni_frag_complete (mca_btl_ugni_base_frag_t *frag, int rc) {
BTL_VERBOSE(("frag complete. flags = %d", frag->base.des_flags));
frag->flags |= MCA_BTL_UGNI_FRAG_COMPLETE;
mca_btl_ugni_frag_del_ref (frag, rc);
}
static inline bool mca_btl_ugni_frag_check_complete (mca_btl_ugni_base_frag_t *frag) {
return !!(MCA_BTL_UGNI_FRAG_COMPLETE & frag->flags);
}
btl/ugni: improve multi-threaded performance This commit updates the ugni btl to make use of multiple device contexts to improve the multi-threaded RMA performance. This commit contains the following: - Cleanup the endpoint structure by removing unnecessary field. The structure now also contains all the fields originally handled by the common/ugni endpoint. - Clean up the fragment allocation code to remove the need to initialize the my_list member of the fragment structure. This member is not initialized by the free list initializer function. - Remove the (now unused) common/ugni component. btl/ugni no longer need the component. common/ugni was originally split out of btl/ugni to support bcol/ugni. As that component exists there is no reason to keep this component. - Create wrappers for the ugni functionality required by btl/ugni. This was done to ease supporting multiple device contexts. The wrappers are thread safe and currently use a spin lock instead of a mutex. This produces better performance when using multiple threads spread over multiple cores. In the future this lock may be replaced by another serialization mechanism. The wrappers are located in a new file: btl_ugni_device.h. - Remove unnecessary device locking from serial parts of the ugni btl. This includes the first add-procs and module finalize. - Clean up fragment wait list code by moving enqueue into common function. - Expose the communication domain flags as an MCA variable. The defaults have been updated to reflect the recommended setting for knl and haswell. - Avoid allocating fragments for communication with already overloaded peers. - Allocate RDMA endpoints dyncamically. This is needed to support spreading RMA operations accross multiple contexts. - Add support for spreading RMA communication over multiple ugni device contexts. This should greatly improve the threading performance when communicating with multiple peers. By default the number of virtual devices depends on 1) whether opal_using_threads() is set, 2) how many local processes are in the job, and 3) how many bits are available in the pid. The last is used to ensure that each CDM is created with a unique id. Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-12 22:37:35 -06:00
void mca_btl_ugni_wait_list_append (mca_btl_ugni_module_t *ugni_module, mca_btl_base_endpoint_t *endpoint,
mca_btl_ugni_base_frag_t *frag);
static inline mca_btl_ugni_base_frag_t *mca_btl_ugni_frag_alloc_smsg (mca_btl_base_endpoint_t *ep)
{
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (ep);
return mca_btl_ugni_frag_alloc (ep, ugni_module->frags_lists + MCA_BTL_UGNI_LIST_SMSG);
}
static inline mca_btl_ugni_base_frag_t *mca_btl_ugni_frag_alloc_rdma (mca_btl_base_endpoint_t *ep)
{
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (ep);
return mca_btl_ugni_frag_alloc (ep, ugni_module->frags_lists + MCA_BTL_UGNI_LIST_RDMA);
}
static inline mca_btl_ugni_base_frag_t *mca_btl_ugni_frag_alloc_rdma_int (mca_btl_base_endpoint_t *ep)
{
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (ep);
return mca_btl_ugni_frag_alloc (ep, ugni_module->frags_lists + MCA_BTL_UGNI_LIST_RDMA_INT);
}
static inline mca_btl_ugni_base_frag_t *mca_btl_ugni_frag_alloc_eager_send (mca_btl_base_endpoint_t *ep)
{
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (ep);
return mca_btl_ugni_frag_alloc (ep, ugni_module->frags_lists + MCA_BTL_UGNI_LIST_EAGER_SEND);
}
static inline mca_btl_ugni_base_frag_t *mca_btl_ugni_frag_alloc_eager_recv (mca_btl_base_endpoint_t *ep)
{
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (ep);
return mca_btl_ugni_frag_alloc (ep, ugni_module->frags_lists + MCA_BTL_UGNI_LIST_EAGER_RECV);
}
#endif /* MCA_BTL_UGNI_FRAG_H */