Merge pull request #1 from cheng-zhao/develop

MPI optimization
cheng-zhao · Jun 26, 2022 · 16e18b3 · 16e18b3
2 parents c0c0d5d + 634d272
commit 16e18b3
Show file tree

Hide file tree

Showing 33 changed files with 572 additions and 483 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -0,0 +1,7 @@
+# FCFC Change Log
+
+## [1.0.1] &ndash; 2022-06-26
+
+-   Refactorized the implementation of MPI and OpenMP parallelisms.
+-   Optimized MPI communications with custom structs, and fixed the deadlock due to multiple broadcasts.
+
diff --git a/Makefile b/Makefile
@@ -60,8 +60,8 @@ TARGETS = FCFC_2PT_BOX FCFC_2PT
 
 # Set mandatory options (do not edit below this line).
 LIBS += -lm
-INCL += -Isrc/fcfc/common -Isrc/io -Isrc/lib -Isrc/math -Isrc/tree
-SRCS = $(wildcard src/fcfc/common/*.c src/*/*.c)
+INCL += -Isrc/util -Isrc/io -Isrc/lib -Isrc/math -Isrc/tree
+SRCS = $(wildcard src/*/*.c)
 
 
 all: $(TARGETS)

diff --git a/options.mk b/options.mk
@@ -26,14 +26,14 @@ WITH_MU_ONE :=
 # Set `WITH_CFITSIO := T` to enable FITS file format.
 # The paths for cfitsio header (fitsio.h) and library (libcfitsio.{a,so,dylib})
 # files can be set via `CFITSIO_INC_DIR` and `CFITSIO_LIB_DIR`, respectively.
-WITH_CFITSIO := T
+WITH_CFITSIO := 
 CFITSIO_INC_DIR := 
 CFITSIO_LIB_DIR := 
 
 # Set `WITH_HDF5 := T` to enable HDF5 file format.
 # The paths for libhdf5 header (hdf5.h) and library (libhdf5.{a,so,dylib})
 # files can be set via `HDF5_INC_DIR` and `HDF5_LIB_DIR`, respectively.
-WITH_HDF5 := T
+WITH_HDF5 := 
 HDF5_INC_DIR := 
 HDF5_LIB_DIR := 
 
diff --git a/src/fcfc/2pt/build_tree.c b/src/fcfc/2pt/build_tree.c
@@ -18,9 +18,6 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-#ifdef MPI
-#include <mpi.h>
-#endif
 
 /*============================================================================*\
                  Function for pre-processing the input catalog
@@ -147,13 +144,13 @@ Function `tree_create`:
   * `conf`:     structure for storing configurations;
   * `cf`:       structure for correlation function evaluations;
   * `idx`:      index of the catalogue to be processed;
-  * `rank`:     ID of MPI task.
+  * `para`:     structure for parallelisms.
 Return:
   Address of the tree on success; NULL on error.
 ******************************************************************************/
 void *tree_create(const CONF *conf, CF *cf, const int idx
 #ifdef MPI
-    , const int rank
+    , const PARA *para
 #endif
     ) {
   void *tree = NULL;
@@ -165,7 +162,7 @@ void *tree_create(const CONF *conf, CF *cf, const int idx
   DATA *data = cf->data + idx;
 
 #ifdef MPI
-  if (rank == FCFC_MPI_ROOT) {
+  if (para->rank == para->root) {
 #endif
     if (!conf) {
       P_ERR("configuration parameters are not loaded\n");
@@ -317,9 +314,8 @@ void *tree_create(const CONF *conf, CF *cf, const int idx
   /* Broadcast the data catalog and tree. */
   switch (cf->treetype) {
     case FCFC_STRUCT_KDTREE:
-      kdtree_broadcast((KDT **) (&tree), &nnode, cf->wt[idx], FCFC_MPI_ROOT,
-          rank);
-      if (rank != FCFC_MPI_ROOT) {
+      kdtree_broadcast((KDT **) (&tree), &nnode, cf->wt[idx], para);
+      if (para->rank != para->root) {
         KDT *root = (KDT *) tree;
         cf->data[idx].n = root->n;
         cf->data[idx].x[0] = root->x[0];
@@ -330,9 +326,8 @@ void *tree_create(const CONF *conf, CF *cf, const int idx
       }
       break;
     case FCFC_STRUCT_BALLTREE:
-      balltree_broadcast((BLT **) (&tree), &nnode, cf->wt[idx], FCFC_MPI_ROOT,
-          rank);
-      if (rank != FCFC_MPI_ROOT) {
+      balltree_broadcast((BLT **) (&tree), &nnode, cf->wt[idx], para);
+      if (para->rank != para->root) {
         BLT *root = (BLT *) tree;
         cf->data[idx].n = root->n;
         cf->data[idx].x[0] = root->x[0];
@@ -347,7 +342,7 @@ void *tree_create(const CONF *conf, CF *cf, const int idx
       return NULL;
   }
 
-  if (rank == FCFC_MPI_ROOT) {
+  if (para->rank == para->root) {
 #endif
     printf(FMT_DONE);
 #ifdef MPI

diff --git a/src/fcfc/2pt/build_tree.h b/src/fcfc/2pt/build_tree.h
@@ -41,13 +41,13 @@ Function `tree_create`:
   * `conf`:     structure for storing configurations;
   * `cf`:       structure for correlation function evaluations;
   * `idx`:      index of the catalogue to be processed;
-  * `rank`:     ID of MPI task.
+  * `para`:     structure for parallelisms.
 Return:
   Address of the tree on success; NULL on error.
 ******************************************************************************/
 void *tree_create(const CONF *conf, CF *cf, const int idx
 #ifdef MPI
-    , const int rank
+    , const PARA *para
 #endif
     );
 

diff --git a/src/fcfc/2pt/count_func.c b/src/fcfc/2pt/count_func.c
@@ -19,9 +19,6 @@
 #include <stdint.h>
 #include <stdlib.h>
 #include <math.h>
-#ifdef MPI
-#include <mpi.h>
-#endif
 #ifdef OMP
 #include <omp.h>
 #endif
@@ -4842,15 +4839,14 @@ Function `count_pairs`:
   * `cnt`:      array for storing pair counts;
   * `isauto`:   true for counting auto pairs;
   * `withwt`:   true for enabling weights;
-  * `ntask`:    number of MPI tasks;
-  * `rank`:     ID of MPI task.
+  * `para`:     structure for parallelisms.
 Return:
   Zero on success; non-zero on error.
 ******************************************************************************/
 int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
     const bool isauto, const bool withwt
 #ifdef MPI
-    , const int ntask, const int rank
+    , const PARA *para
 #endif
     ) {
   void (*paircnt_func) (STACK_DUAL_NODE *, const void *, void *) = NULL;
@@ -7114,7 +7110,7 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
   /* Prepare for dynamic MPI scheduler. */
   uint32_t *addr = NULL;
   MPI_Aint win_size = 0;
-  if (rank == FCFC_MPI_ROOT) {
+  if (para->rank == para->root) {
     win_size = sizeof(uint32_t);
     if (MPI_Alloc_mem(win_size, MPI_INFO_NULL, &addr)) {
       P_ERR("failed to allocate memory for MPI window\n");
@@ -7124,14 +7120,14 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
   }
   MPI_Win win;
   if (MPI_Win_create(addr, win_size, sizeof(uint32_t), MPI_INFO_NULL,
-      MPI_COMM_WORLD, &win)) {
+      para->comm, &win)) {
     P_ERR("failed to create the MPI window\n");
     free(conf); FCFC_QUIT(FCFC_ERR_MPI);
   }
   uint32_t start, step = cf->nthread * FCFC_STACK_SIZE_PER_THREAD;
 
   /* Compute the tree level for dual-node distribution. */
-  size_t nnodes = ntask * FCFC_STACK_SIZE_PER_TASK;
+  size_t nnodes = para->ntask * FCFC_STACK_SIZE_PER_TASK;
   #ifdef OMP    /* both MPI and OMP */
   nnodes *= cf->nthread * FCFC_STACK_SIZE_PER_THREAD;
   #endif
@@ -7147,9 +7143,9 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
 
       /* Distribute dual nodes to tasks dynamically. */
       do {
-        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, FCFC_MPI_ROOT, 0, win) ||
-            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, FCFC_MPI_ROOT, 0,
-            MPI_SUM, win) || MPI_Win_unlock(FCFC_MPI_ROOT, win)) {
+        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, para->root, 0, win) ||
+            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, para->root, 0,
+            MPI_SUM, win) || MPI_Win_unlock(para->root, win)) {
           P_ERR("failed to fetch dual nodes with MPI tasks\n");
           free(conf); free(nodes);
           FCFC_QUIT(FCFC_ERR_MPI);
@@ -7210,9 +7206,9 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
 
       /* Distribute dual nodes to tasks dynamically. */
       do {
-        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, FCFC_MPI_ROOT, 0, win) ||
-            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, FCFC_MPI_ROOT, 0,
-            MPI_SUM, win) || MPI_Win_unlock(FCFC_MPI_ROOT, win)) {
+        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, para->root, 0, win) ||
+            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, para->root, 0,
+            MPI_SUM, win) || MPI_Win_unlock(para->root, win)) {
           P_ERR("failed to fetch dual nodes with MPI tasks\n");
           free(conf); free(nodes1); free(nodes2);
           FCFC_QUIT(FCFC_ERR_MPI);
@@ -7270,9 +7266,9 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
 
       /* Distribute dual nodes to tasks dynamically. */
       do {
-        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, FCFC_MPI_ROOT, 0, win) ||
-            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, FCFC_MPI_ROOT, 0,
-            MPI_SUM, win) || MPI_Win_unlock(FCFC_MPI_ROOT, win)) {
+        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, para->root, 0, win) ||
+            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, para->root, 0,
+            MPI_SUM, win) || MPI_Win_unlock(para->root, win)) {
           P_ERR("failed to fetch dual nodes with MPI tasks\n");
           free(conf); free(nodes);
           FCFC_QUIT(FCFC_ERR_MPI);
@@ -7333,9 +7329,9 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
 
       /* Distribute dual nodes to tasks dynamically. */
       do {
-        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, FCFC_MPI_ROOT, 0, win) ||
-            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, FCFC_MPI_ROOT, 0,
-            MPI_SUM, win) || MPI_Win_unlock(FCFC_MPI_ROOT, win)) {
+        if (MPI_Win_lock(MPI_LOCK_EXCLUSIVE, para->root, 0, win) ||
+            MPI_Fetch_and_op(&step, &start, MPI_UINT32_T, para->root, 0,
+            MPI_SUM, win) || MPI_Win_unlock(para->root, win)) {
           P_ERR("failed to fetch dual nodes with MPI tasks\n");
           free(conf); free(nodes1); free(nodes2);
           FCFC_QUIT(FCFC_ERR_MPI);
@@ -7386,7 +7382,7 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
 
   /* Cleanup MPI scheduler. */
   if (MPI_Win_free(&win) ||
-      (rank == FCFC_MPI_ROOT && MPI_Free_mem(addr))) {
+      (para->rank == para->root && MPI_Free_mem(addr))) {
     P_ERR("failed to release memory for MPI scheduler\n");
     free(conf); stack_destroy(&stack);
     FCFC_QUIT(FCFC_ERR_MPI);
@@ -7630,16 +7626,16 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
       for (size_t i = 0; i < cf->ntot; i++) pcnt[i] = cnt[i].d;
     }
 
-    if (rank == FCFC_MPI_ROOT) {
+    if (para->rank == para->root) {
       if (MPI_Ireduce(MPI_IN_PLACE, pcnt, cf->ntot, MPI_DOUBLE, MPI_SUM,
-          FCFC_MPI_ROOT, MPI_COMM_WORLD, &req)) {
+          para->root, para->comm, &req)) {
         P_ERR("failed to gather pair counts from MPI tasks\n");
         FCFC_QUIT(FCFC_ERR_MPI);
       }
     }
     else {
       if (MPI_Ireduce(pcnt, NULL, cf->ntot, MPI_DOUBLE, MPI_SUM,
-                      FCFC_MPI_ROOT, MPI_COMM_WORLD, &req)) {
+                      para->root, para->comm, &req)) {
         P_ERR("failed to gather pair counts from MPI tasks\n");
         FCFC_QUIT(FCFC_ERR_MPI);
       }
@@ -7662,16 +7658,16 @@ int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
       for (size_t i = 0; i < cf->ntot; i++) pcnt[i] = cnt[i].i;
     }
 
-    if (rank == FCFC_MPI_ROOT) {
+    if (para->rank == para->root) {
       if (MPI_Ireduce(MPI_IN_PLACE, pcnt, cf->ntot, MPI_INT64_T, MPI_SUM,
-          FCFC_MPI_ROOT, MPI_COMM_WORLD, &req)) {
+          para->root, para->comm, &req)) {
         P_ERR("failed to gather pair counts from MPI tasks\n");
         FCFC_QUIT(FCFC_ERR_MPI);
       }
     }
     else {
       if (MPI_Ireduce(pcnt, NULL, cf->ntot, MPI_INT64_T, MPI_SUM,
-          FCFC_MPI_ROOT, MPI_COMM_WORLD, &req)) {
+          para->root, para->comm, &req)) {
         P_ERR("failed to gather pair counts from MPI tasks\n");
         FCFC_QUIT(FCFC_ERR_MPI);
       }

diff --git a/src/fcfc/2pt/count_func.h b/src/fcfc/2pt/count_func.h
@@ -47,15 +47,14 @@ Function `kdtree_count_cross_pairs`:
   * `cnt`:      array for storing pair counts;
   * `isauto`:   true for counting auto pairs;
   * `withwt`:   true for enabling weights;
-  * `ntask`:    number of MPI tasks;
-  * `rank`:     ID of MPI task.
+  * `para`:     structure for parallelisms.
 Return:
   Zero on success; non-zero on error.
 ******************************************************************************/
 int count_pairs(const void *tree1, const void *tree2, CF *cf, COUNT *cnt,
     const bool isauto, const bool withwt
 #ifdef MPI
-    , const int ntask, const int rank
+    , const PARA *para
 #endif
     );