Parallel-Split MPI_Alltoall algorithm as part of acoll collective.

-A new parallel-split algorithm for MPI_Alltoall is introduced as part
of acoll collective component, primarily targeting smaller message sizes
(<= 4KB). The algorithm, at a high level, operates by diving the ranks
into n groups, performing alltoall (using a base alltoall routine)
within the n groups in parallel, following which data is exchanged
between groups of n adjacent ranks (starting from rank 0). For example
if n=2, this algorithm splits the ranks into 2 groups, one containing
all even ranked processes and another containing all odd ranked
processes. Alltoall is performed within these 2 groups in parallel,
followed by which each adjacent even-odd pairs (pairs being [0,1],
[2,3],..) exchanges data to complete Alltoall operation. If n =4 or n=8,
alltoall is performed within 4 or 8 groups in parallel. Following this
step, groups of adjacent 4 or 8 ranks(starting from 0) exchanges data
among themselves to complete the alltoall operation.
-Additionally for intra node cases, an xpmem based linear algorithm for
MPI_Alltoall is added as part of acoll. When sbuf and rbuf can be
exposed via xpmem, the alltoall algorithm can be implemented as linear
direct copy from sbuf of all the other ranks to rbuf of a given rank.

Signed-off-by: Mithun Mohan <[email protected]>

Author: MithunMohanKadavil

ompi/mca/coll/acoll/Makefile.am

@@ -15,6 +15,7 @@ sources = \
         coll_acoll_allgather.c \
         coll_acoll_bcast.c \
         coll_acoll_gather.c \
+        coll_acoll_alltoall.c \
         coll_acoll_reduce.c \
         coll_acoll_allreduce.c \
         coll_acoll_barrier.c \

ompi/mca/coll/acoll/coll_acoll.h

@@ -66,6 +66,13 @@ int mca_coll_acoll_gather_intra(const void *sbuf, size_t scount, struct ompi_dat
                                 void *rbuf, size_t rcount, struct ompi_datatype_t *rdtype, int root,
                                 struct ompi_communicator_t *comm, mca_coll_base_module_t *module);
 
+int mca_coll_acoll_alltoall(const void *sbuf, size_t scount,
+                            struct ompi_datatype_t *sdtype,
+                            void* rbuf, size_t rcount,
+                            struct ompi_datatype_t *rdtype,
+                            struct ompi_communicator_t *comm,
+                            mca_coll_base_module_t *module);
+
 int mca_coll_acoll_reduce_intra(const void *sbuf, void *rbuf, size_t count,
                                 struct ompi_datatype_t *dtype, struct ompi_op_t *op, int root,
                                 struct ompi_communicator_t *comm, mca_coll_base_module_t *module);
@@ -80,6 +87,7 @@ int mca_coll_acoll_barrier_intra(struct ompi_communicator_t *comm, mca_coll_base
 END_C_DECLS
 
 #define MCA_COLL_ACOLL_ROOT_CHANGE_THRESH 10
+#define MCA_COLL_ACOLL_SPLIT_FACTOR_LIST_LEN 3
 
 typedef enum MCA_COLL_ACOLL_SG_SIZES {
     MCA_COLL_ACOLL_SG_SIZE_1 = 8,
@@ -142,6 +150,18 @@ typedef struct coll_acoll_data {
     int sync[2];
 } coll_acoll_data_t;
 
+/* The enum literals are used as indices into arrays and values are
+ * assigned to the enum literals so as to ensure it is valid irrespective
+ * of what the compiler assigns. */
+typedef enum MCA_COLL_ACOLL_R2R_DIST {
+    DIST_CORE = 0,
+    DIST_L3CACHE,
+    DIST_NUMA,
+    DIST_SOCKET,
+    DIST_NODE,
+    DIST_END
+} MCA_COLL_ACOLL_R2R_DIST_T;
+
 typedef struct coll_acoll_subcomms {
     ompi_communicator_t *local_comm;
     ompi_communicator_t *local_r_comm;
@@ -152,6 +172,7 @@ typedef struct coll_acoll_subcomms {
     ompi_communicator_t *orig_comm;
     ompi_communicator_t *socket_comm;
     ompi_communicator_t *socket_ldr_comm;
+    ompi_communicator_t *split_comm[MCA_COLL_ACOLL_SPLIT_FACTOR_LIST_LEN]; // AllToAll odd even split comm
     int num_nodes;
     int derived_node_size;
     int is_root_node;
@@ -170,6 +191,7 @@ typedef struct coll_acoll_subcomms {
     int initialized;
     int prev_init_root;
     int num_root_change;
+    MCA_COLL_ACOLL_R2R_DIST_T r2r_dist;
 
     ompi_communicator_t *numa_comm_ldrs;
     ompi_communicator_t *node_comm;