Add timers to measure cost of intra-node aggregation

Author: wkliao

src/drivers/ncmpio/ncmpio_NC.h

@@ -446,6 +446,9 @@ struct NC {
     int  my_aggr;            /* rank ID of my aggregator */
     int  num_nonaggrs;       /* number of non-aggregators assigned */
     int *nonaggr_ranks;      /* ranks of assigned non-aggregators */
+#ifdef PNETCDF_PROFILING
+    double aggr_time;
+#endif
 };
 
 #define NC_readonly(ncp)   fIsSet((ncp)->flags, NC_MODE_RDONLY)
@@ -647,11 +650,11 @@ extern int
 ncmpio_intra_node_aggr_init(NC *ncp);
 
 extern int
-ncmpio_intra_node_aggregation_nreqs(NC *ncp, int num_reqs, NC_req *put_list,
-                                    MPI_Offset newnumrecs);
+ncmpio_intra_node_aggregation_nreqs(NC *ncp, int mode, int num_reqs,
+                                    NC_req *put_list, MPI_Offset newnumrecs);
 extern int
-ncmpio_intra_node_aggregation(NC *ncp, NC_var *varp, const MPI_Offset *start,
-                              const MPI_Offset *count,
+ncmpio_intra_node_aggregation(NC *ncp, int mode, NC_var *varp,
+                              const MPI_Offset *start, const MPI_Offset *count,
                               const MPI_Offset *stride, MPI_Offset bufCount,
                               MPI_Datatype bufType, void *buf);
 

src/drivers/ncmpio/ncmpio_getput.m4

@@ -270,7 +270,8 @@ err_check:
     if (fIsSet(reqMode, NC_REQ_COLL) && ncp->my_aggr >= 0 && ncp->nprocs > 1) {
         /* intra-node write aggregation must be in collective mode */
         void *wbuf = (nbytes == 0) ?  NULL : xbuf;
-        err = ncmpio_intra_node_aggregation(ncp, varp, start, count, stride, nelems, xtype, wbuf);
+        err = ncmpio_intra_node_aggregation(ncp, NC_REQ_WR, varp, start, count,
+                                            stride, nelems, xtype, wbuf);
         if (status == NC_NOERR) status = err;
     }
     else {

src/drivers/ncmpio/ncmpio_intra_node.c

@@ -1,6 +1,29 @@
 /*
  *  Copyright (C) 2024, Northwestern University and Argonne National Laboratory
  *  See COPYRIGHT notice in top-level directory.
+ *
+ * This file contains the implementation of intra-node aggregation feature,
+ * which is designed for the I/O patterns that contain many noncontiguous
+ * requests interleaved among processes, and spreading across a wide range of
+ * file space. It is particularly useful when the number of MPI processes
+ * allocated to a compute node is large.
+ *
+ * This feature is enabled by setting the PnetCDF hint 'nc_num_aggrs_per_node'
+ * to a positive integral value indicating the desired number of processes per
+ * compute node to be selected as the intra-node I/O aggregators. Each process
+ * is assigned a unique aggregator. The non-aggregators send their requests to
+ * the assigned aggregators, and then the aggregators make MPI-IO requests to
+ * the file.
+ *
+ * Such strategy can effectively reduce communication congestion due to many
+ * pending asynchronous messages produced in the collective write inside of
+ * MPI-IO.
+ *
+ * The concept of intra-node request aggregation is based on the paper:
+ * Q. Kang, S. Lee, K. Hou, R. Ross, A. Agrawal, A. Choudhary, and W. Liao.
+ * Improving MPI Collective I/O for High Volume Non-Contiguous Requests With
+ * Intra-Node Aggregation. IEEE Transactions on Parallel and Distributed
+ * Systems (TPDS), 31(11):2682-2695, November 2020.
  */
 
 #ifdef HAVE_CONFIG_H
@@ -48,6 +71,11 @@
                       ((*(b) < *(c)) ? (b) : ((*(a) < *(c)) ? (c) : (a))) : \
                       ((*(b) > *(c)) ? (b) : ((*(a) < *(c)) ? (a) : (c))))
 
+/*----< qsort_off_len_buf() >------------------------------------------------*/
+/* Sort three arrays of offsets, lengths, and buffer addresses based on the
+ * increasing order of offsets. This code is based on the qsort routine from
+ * Bentley & McIlroy's "Engineering a Sort Function".
+ */
 static void
 qsort_off_len_buf(MPI_Aint num,
 #ifdef HAVE_MPI_LARGE_COUNT
@@ -125,7 +153,7 @@ qsort_off_len_buf(MPI_Aint num,
         if ((r = pb - pa) > 1)
             qsort_off_len_buf(r, offsets, lengths, bufAddr);
         if ((r = pd - pc) > 1) {
-            /* Iterate rather than recurse to save stack space */
+            /* Iterate rather than recursively call self to save stack space */
             lengths = lengths + (num - r);
             bufAddr = bufAddr + (num - r);
             offsets = pn - r;
@@ -161,10 +189,18 @@ ncmpio_intra_node_aggr_init(NC *ncp)
     ncp->num_nonaggrs = 0;     /* number of non-aggregators assigned */
     ncp->nonaggr_ranks = NULL; /* ranks of assigned non-aggregators */
 
+#ifdef PNETCDF_PROFILING
+    ncp->aggr_time = 0.0;
+#endif
+
     if (ncp->num_aggrs_per_node == 0 || ncp->num_aggrs_per_node == ncp->nprocs)
         /* disable intra-node aggregation */
         return NC_NOERR;
 
+#ifdef PNETCDF_PROFILING
+    double timing = MPI_Wtime();
+#endif
+
     /* allocate space for storing the rank IDs of non-aggregators assigned to
      * this rank. Note ncp->nonaggr_ranks[] will be freed when closing the
      * file, if allocated.
@@ -368,6 +404,10 @@ ncmpio_intra_node_aggr_init(NC *ncp)
      * of processes are allocated on the same node.
      */
 
+#ifdef PNETCDF_PROFILING
+    ncp->aggr_time = MPI_Wtime() - timing;
+#endif
+
     return NC_NOERR;
 }
 
@@ -791,7 +831,9 @@ intra_node_aggregation(NC           *ncp,
     MPI_Datatype recvTypes, fileType=MPI_BYTE;
     MPI_File fh;
     MPI_Request *req=NULL;
-
+#ifdef PNETCDF_PROFILING
+    double timing = MPI_Wtime();
+#endif
 #ifdef HAVE_MPI_LARGE_COUNT
     MPI_Count bufLen;
     MPI_Type_size_c(bufType, &bufLen);
@@ -1190,6 +1232,10 @@ intra_node_aggregation(NC           *ncp,
         NCI_Free(lengths);
     }
 
+#ifdef PNETCDF_PROFILING
+    ncp->aggr_time += MPI_Wtime() - timing;
+#endif
+
     if (ncp->rank != ncp->my_aggr) /* non-aggregator writes nothing */
         buf_count = 0;
 
@@ -1221,6 +1267,7 @@ intra_node_aggregation(NC           *ncp,
 /* This is a collective call */
 int
 ncmpio_intra_node_aggregation_nreqs(NC         *ncp,
+                                    int         reqMode,
                                     int         num_reqs,
                                     NC_req     *put_list,
                                     MPI_Offset  newnumrecs)
@@ -1234,6 +1281,12 @@ ncmpio_intra_node_aggregation_nreqs(NC         *ncp,
     int *lengths=NULL;
 #endif
     MPI_Datatype bufType=MPI_BYTE;
+#ifdef PNETCDF_PROFILING
+    double timing = MPI_Wtime();
+#endif
+
+    /* currently supports write requests only */
+    if (fIsSet(reqMode, NC_REQ_RD)) return NC_NOERR;
 
     assert(ncp->my_aggr >= 0);
 
@@ -1260,6 +1313,10 @@ ncmpio_intra_node_aggregation_nreqs(NC         *ncp,
     if (put_list != NULL)
         NCI_Free(put_list);
 
+#ifdef PNETCDF_PROFILING
+    ncp->aggr_time += MPI_Wtime() - timing;
+#endif
+
     err = intra_node_aggregation(ncp, num_pairs, offsets, lengths, bufLen,
                                  bufType, NULL);
     if (status == NC_NOERR) status = err;
@@ -1291,6 +1348,7 @@ ncmpio_intra_node_aggregation_nreqs(NC         *ncp,
 /* This is a collective call */
 int
 ncmpio_intra_node_aggregation(NC               *ncp,
+                              int               reqMode,
                               NC_var           *varp,
                               const MPI_Offset *start,
                               const MPI_Offset *count,
@@ -1307,6 +1365,12 @@ ncmpio_intra_node_aggregation(NC               *ncp,
     MPI_Aint *offsets=NULL;
     int *lengths=NULL;
 #endif
+#ifdef PNETCDF_PROFILING
+    double timing = MPI_Wtime();
+#endif
+
+    /* currently supports write requests only */
+    if (fIsSet(reqMode, NC_REQ_RD)) return NC_NOERR;
 
     if (buf == NULL) /* zero-length request */
         return intra_node_aggregation(ncp, 0, NULL, NULL, 0, MPI_BYTE, NULL);
@@ -1326,6 +1390,10 @@ ncmpio_intra_node_aggregation(NC               *ncp,
     }
     status = err;
 
+#ifdef PNETCDF_PROFILING
+    ncp->aggr_time += MPI_Wtime() - timing;
+#endif
+
     err = intra_node_aggregation(ncp, num_pairs, offsets, lengths, bufCount,
                                  bufType, buf);
     if (status == NC_NOERR) status = err;

src/drivers/ncmpio/ncmpio_wait.c

@@ -989,7 +989,9 @@ req_commit(NC  *ncp,
 
         if (ncp->my_aggr >= 0 && coll_indep == NC_REQ_COLL && ncp->nprocs > 1)
             /* intra-node write aggregation must be in collective mode */
-            err = ncmpio_intra_node_aggregation_nreqs(ncp, num_w_reqs, put_list, newnumrecs);
+            err = ncmpio_intra_node_aggregation_nreqs(ncp, NC_REQ_WR,
+                                                      num_w_reqs, put_list,
+                                                      newnumrecs);
         else
             err = wait_getput(ncp, num_w_reqs, put_list, NC_REQ_WR, coll_indep,
                               newnumrecs);