Merge branch 'main' into topic/main/detect-host-numa-as-device-mem

Author: Akshay-Venkatesh

docs/installing-open-mpi/configure-cli-options/networking.rst

@@ -73,7 +73,7 @@ can be used with ``configure``:
   only necessary if the PSM headers and libraries are not in default
   compiler/linker search paths.
 
-  PSM is the support library for QLogic InfiniPath and Intel TrueScale
+  PSM is the support library for QLogic InfiniPath and Intel True Scale
   network adapters.
 
 * ``--with-psm-libdir=DIR``:

docs/tuning-apps/networking/cuda.rst

@@ -160,21 +160,21 @@ PSM2 support for CUDA
 ---------------------
 
 CUDA-aware support is present in PSM2 MTL.  When running CUDA-aware
-Open MPI on Intel Omni-path, the PSM2 MTL will automatically set
+Open MPI on Cornelis Networks Omni-Path, the PSM2 MTL will automatically set
 ``PSM2_CUDA`` environment variable which enables PSM2 to handle GPU
 buffers.  If the user wants to use host buffers with a CUDA-aware Open
 MPI, it is recommended to set ``PSM2_CUDA`` to ``0`` in the execution
 environment. PSM2 also has support for the NVIDIA GPUDirect support
 feature. To enable this, users will need to set ``PSM2_GPUDIRECT``
 to ``1`` in the execution environment.
 
-Note: The PSM2 library and ``hfi1`` driver with CUDA support are
-requirements to use GPUDirect support on Intel Omni-Path. The minimum
+Note: The PSM2 library and ``hfi1`` driver with CUDA support are requirements
+to use GPUDirect support on Cornelis Networks Omni-Path. The minimum
 PSM2 build version required is `PSM2 10.2.175
 <https://github.com/01org/opa-psm2/releases/tag/PSM2_10.2-175>`_.
 
-For more information refer to the `Intel Omni-Path documentation
-<https://www.intel.com/content/www/us/en/support/articles/000016242/network-and-i-o/fabric-products.html>`_.
+For more information refer to the `Cornelis Networks Customer Center
+<https://customercenter.cornelisnetworks.com/>`_.
 
 /////////////////////////////////////////////////////////////////////////
 
@@ -629,10 +629,9 @@ limit is reached:
 What are some guidelines for using CUDA and Open MPI with Omni-Path?
 --------------------------------------------------------------------
 
-When developing CUDA-aware Open MPI applications for OPA-based
-fabrics, the PSM2 transport is preferred and a CUDA-aware version of
-PSM2 is provided with all versions of the Intel Omni-Path IFS software
-suite.
+When developing CUDA-aware Open MPI applications for OPA-based fabrics, the
+PSM2 transport is preferred and a CUDA-aware version of PSM2 is provided with
+all versions of the Cornelis Networks Omni-Path OPXS software suite.
 
 .. error:: TODO Are Intel/OPA references still correct?
 
@@ -655,10 +654,10 @@ processes do not move between NUMA nodes. See the section on
 :ref:`NUMA Node Issues <faq-cuda-mpi-cuda-numa-issues-label>`, for
 more information.
 
-For more information see the *Intel Performance Scaled Messaging 2
-(PSM2) Programmer's Guide* and the *Intel Omni-Path Performance Tuning
-Guide*, which can be found on the `Intel Omni-Path web site
-<https://www.intel.com/omnipath/FabricSoftwarePublications>`_.
+For more information see the *Cornelis Networks Performance Scaled Messaging 2
+(PSM2) Programmer's Guide* and the *Cornelis Networks Omni-Path Performance
+Tuning Guide*, which can be found in the `Cornelis Networks Customer Center
+<https://customercenter.cornelisnetworks.com/>`_.
 
 .. error:: TODO Are Intel/OPA references still correct?
 

docs/tuning-apps/networking/ofi.rst

@@ -131,6 +131,5 @@ Users can restrict access to a single HFI using the environment variable:
 More details can be found on the PSM2 Programmer's Guide and the Omni-Path
 Fabric Performance Tuning Guide.
 
-Please also see `the full Omni-Path documentation
-<https://www.intel.com/content/www/us/en/support/articles/000016242/network-and-i-o/fabric-products.html>`_
+Please see the `Cornelis Networks Customer Center <https://customercenter.cornelisnetworks.com/>`_
 for more details.

ompi/mca/coll/han/Makefile.am

@@ -16,6 +16,7 @@ coll_han.h \
 coll_han_trigger.h \
 coll_han_algorithms.h \
 coll_han_alltoall.c \
+coll_han_alltoallv.c \
 coll_han_dynamic.h \
 coll_han_dynamic_file.h \
 coll_han_barrier.c \
@@ -34,8 +35,7 @@ coll_han_algorithms.c \
 coll_han_dynamic.c \
 coll_han_dynamic_file.c \
 coll_han_topo.c \
-coll_han_subcomms.c \
-coll_han_utils.c
+coll_han_subcomms.c
 
 # Make the output library in this directory, and name it either
 # mca_<type>_<name>.la (for DSO builds) or libmca_<type>_<name>.la

ompi/mca/coll/han/coll_han.h

@@ -199,6 +199,7 @@ typedef struct mca_coll_han_op_module_name_t {
     mca_coll_han_op_up_low_module_name_t scatter;
     mca_coll_han_op_up_low_module_name_t scatterv;
     mca_coll_han_op_up_low_module_name_t alltoall;
+    mca_coll_han_op_up_low_module_name_t alltoallv;
 } mca_coll_han_op_module_name_t;
 
 /**
@@ -260,6 +261,11 @@ typedef struct mca_coll_han_component_t {
     /* alltoall: parallel stages */
     int32_t han_alltoall_pstages;
 
+    /* low level module for alltoallv */
+    uint32_t han_alltoallv_low_module;
+    int64_t han_alltoallv_smsc_avg_send_limit;
+    double han_alltoallv_smsc_noncontig_activation_limit;
+
 
     /* name of the modules */
     mca_coll_han_op_module_name_t han_op_module_name;
@@ -286,6 +292,8 @@ typedef struct mca_coll_han_component_t {
 
     /* Define maximum dynamic errors printed by rank 0 with a 0 verbosity level */
     int max_dynamic_errors;
+
+    opal_free_list_t pack_buffers;
 } mca_coll_han_component_t;
 
 /*
@@ -297,6 +305,7 @@ typedef struct mca_coll_han_single_collective_fallback_s
     union
     {
         mca_coll_base_module_alltoall_fn_t alltoall;
+        mca_coll_base_module_alltoallv_fn_t alltoallv;
         mca_coll_base_module_allgather_fn_t allgather;
         mca_coll_base_module_allgatherv_fn_t allgatherv;
         mca_coll_base_module_allreduce_fn_t allreduce;
@@ -319,6 +328,7 @@ typedef struct mca_coll_han_single_collective_fallback_s
 typedef struct mca_coll_han_collectives_fallback_s
 {
     mca_coll_han_single_collective_fallback_t alltoall;
+    mca_coll_han_single_collective_fallback_t alltoallv;
     mca_coll_han_single_collective_fallback_t allgather;
     mca_coll_han_single_collective_fallback_t allgatherv;
     mca_coll_han_single_collective_fallback_t allreduce;
@@ -384,6 +394,9 @@ OBJ_CLASS_DECLARATION(mca_coll_han_module_t);
 #define previous_alltoall           fallback.alltoall.alltoall
 #define previous_alltoall_module    fallback.alltoall.module
 
+#define previous_alltoallv           fallback.alltoallv.alltoallv
+#define previous_alltoallv_module    fallback.alltoallv.module
+
 #define previous_allgather          fallback.allgather.allgather
 #define previous_allgather_module   fallback.allgather.module
 
@@ -440,6 +453,7 @@ OBJ_CLASS_DECLARATION(mca_coll_han_module_t);
         HAN_UNINSTALL_COLL_API(COMM, HANM, allgather);                 \
         HAN_UNINSTALL_COLL_API(COMM, HANM, allgatherv);                \
         HAN_UNINSTALL_COLL_API(COMM, HANM, alltoall);                  \
+        HAN_UNINSTALL_COLL_API(COMM, HANM, alltoallv);                 \
         han_module->enabled = false;  /* entire module set to pass-through from now on */ \
     } while(0)
 
@@ -503,6 +517,9 @@ int
 mca_coll_han_alltoall_intra_dynamic(ALLTOALL_BASE_ARGS,
                                     mca_coll_base_module_t *module);
 int
+mca_coll_han_alltoallv_intra_dynamic(ALLTOALLV_BASE_ARGS,
+                                    mca_coll_base_module_t *module);
+int
 mca_coll_han_allgather_intra_dynamic(ALLGATHER_BASE_ARGS,
                                      mca_coll_base_module_t *module);
 int
@@ -544,13 +561,6 @@ ompi_coll_han_reorder_gather(const void *sbuf,
                              struct ompi_communicator_t *comm,
                              int * topo);
 
-size_t
-coll_han_utils_gcd(const uint64_t *numerators, const size_t size);
-
-int
-coll_han_utils_create_contiguous_datatype(size_t count, const ompi_datatype_t *oldType,
-                                          ompi_datatype_t **newType);
-
 static inline struct mca_smsc_endpoint_t *mca_coll_han_get_smsc_endpoint (struct ompi_proc_t *proc) {
     extern opal_mutex_t mca_coll_han_lock;
     if (NULL == proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_SMSC]) {
@@ -566,4 +576,7 @@ static inline struct mca_smsc_endpoint_t *mca_coll_han_get_smsc_endpoint (struct
     return (struct mca_smsc_endpoint_t *) proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_SMSC];
 }
 
+#define COLL_HAN_PACKBUF_PAYLOAD_BYTES (128*1024)
+
+
 #endif                          /* MCA_COLL_HAN_EXPORT_H */

ompi/mca/coll/han/coll_han_algorithms.c

@@ -82,6 +82,10 @@ mca_coll_han_algorithm_value_t*  mca_coll_han_available_algorithms[COLLCOUNT] =
         {"smsc", (fnptr_t)&mca_coll_han_alltoall_using_smsc}, // 2-level
         { 0 }
     },
+    [ALLTOALLV] = (mca_coll_han_algorithm_value_t[]){
+        {"smsc", (fnptr_t)&mca_coll_han_alltoallv_using_smsc}, // 2-level
+        { 0 }
+    },
 };
 
 int

ompi/mca/coll/han/coll_han_algorithms.h

@@ -214,5 +214,10 @@ int
 mca_coll_han_alltoall_using_smsc(ALLTOALL_BASE_ARGS,
                                     mca_coll_base_module_t *module);
 
+/* Alltoallv */
+int
+mca_coll_han_alltoallv_using_smsc(ALLTOALLV_BASE_ARGS,
+                                    mca_coll_base_module_t *module);
+
 
 #endif