When flushing, the metadata cache attempts to flush entries in increasing

address order.  To facilitate this, the metadata cache needs a list of
of dirty entries in increasing address order.  This is implemented via a
skip list of all dirty entries in the cache.

To date this skip list has been maintained at all times.

However, profiling indicates that we can avoid significant overhead by
constructing the skip list of dirty entries just before a flush, taking
it down afterwareds, and not maintaining it during normal operation.

This commit implements this optimization for both serial and parallel.

Tested serial and parallel, debug and production on charis and jelly.

1 files changed, 31 insertions, 11 deletions

diff --git a/src/H5ACmpio.c b/src/H5ACmpio.c
index f097e83..1c914b0 100644
--- a/src/H5ACmpio.c
+++ b/src/H5ACmpio.c
@@ -791,7 +791,7 @@ H5AC__log_dirtied_entry(const H5AC_info_t *entry_ptr)
     else {
         aux_ptr->dirty_bytes += entry_ptr->size;
 #if H5AC_DEBUG_DIRTY_BYTES_CREATION
-        aux_ptr->unprotect_dirty_bytes += entry_size;
+        aux_ptr->unprotect_dirty_bytes += entry_ptr->size;
         aux_ptr->unprotect_dirty_bytes_updates += 1;
 #endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */
     } /* end else */
@@ -989,7 +989,7 @@ H5AC__log_inserted_entry(const H5AC_info_t *entry_ptr)
     aux_ptr->dirty_bytes += entry_ptr->size;
 
 #if H5AC_DEBUG_DIRTY_BYTES_CREATION
-    aux_ptr->insert_dirty_bytes += size;
+    aux_ptr->insert_dirty_bytes += entry_ptr->size;
     aux_ptr->insert_dirty_bytes_updates += 1;
 #endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */
 
@@ -1875,6 +1875,8 @@ done:
  * Programmer:  John Mainzer
  *              April 28, 2010
  *
+ * Changes:     None.
+ *
  *-------------------------------------------------------------------------
  */
 static herr_t
@@ -1894,7 +1896,8 @@ H5AC__rsp__p0_only__flush(H5F_t *f)
     aux_ptr = (H5AC_aux_t *)H5C_get_aux_ptr(cache_ptr);
     HDassert(aux_ptr != NULL);
     HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC);
-    HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__PROCESS_0_ONLY);
+    HDassert(aux_ptr->metadata_write_strategy == \
+             H5AC_METADATA_WRITE_STRATEGY__PROCESS_0_ONLY);
 
     /* To prevent "messages from the future" we must
      * synchronize all processes before we start the flush.
@@ -1903,9 +1906,12 @@ H5AC__rsp__p0_only__flush(H5F_t *f)
      * However, when flushing from within the close operation from a file,
      * it's possible to skip this barrier (on the second flush of the cache).
      */
-    if(!H5CX_get_mpi_file_flushing())
-        if(MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm)))
+    if ( ! H5CX_get_mpi_file_flushing() ) {
+
+        if ( MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm)) )
+
             HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_result)
+    }
 
     /* Flush data to disk, from rank 0 process */
     if(aux_ptr->mpi_rank == 0) {
@@ -1921,23 +1927,30 @@ H5AC__rsp__p0_only__flush(H5F_t *f)
         aux_ptr->write_permitted = FALSE;
 
         /* Check for error on the write operation */
-        if(result < 0)
+        if ( result < 0 )
+
             HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't flush.")
 
         /* this code exists primarily for the test bed -- it allows us to
          * enforce POSIX semantics on the server that pretends to be a
          * file system in our parallel tests.
          */
-        if(aux_ptr->write_done)
+        if ( aux_ptr->write_done ) {
+
             (aux_ptr->write_done)();
+        }
     } /* end if */
 
     /* Propagate cleaned entries to other ranks. */
-    if(H5AC__propagate_flushed_and_still_clean_entries_list(f) < 0)
-        HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't propagate clean entries list.")
+    if ( H5AC__propagate_flushed_and_still_clean_entries_list(f) < 0 )
+
+        HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \
+                    "Can't propagate clean entries list.")
 
 done:
+
     FUNC_LEAVE_NOAPI(ret_value)
+
 } /* H5AC__rsp__p0_only__flush() */
 
 
@@ -2104,16 +2117,22 @@ H5AC__run_sync_point(H5F_t *f, int sync_point_op)
 
     /* Sanity checks */
     HDassert(f != NULL);
+
     cache_ptr = f->shared->cache;
+
     HDassert(cache_ptr != NULL);
+
     aux_ptr = (H5AC_aux_t *)H5C_get_aux_ptr(cache_ptr);
+
     HDassert(aux_ptr != NULL);
     HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC);
     HDassert((sync_point_op == H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN) ||
-        (sync_point_op == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED));
+             (sync_point_op == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED));
+
 
 #if H5AC_DEBUG_DIRTY_BYTES_CREATION
-HDfprintf(stdout, "%d:H5AC_propagate...:%u: (u/uu/i/iu/r/ru) = %zu/%u/%zu/%u/%zu/%u\n",
+HDfprintf(stdout, 
+    "%d:H5AC_propagate...:%u: (u/uu/i/iu/r/ru) = %zu/%u/%zu/%u/%zu/%u\n",
     aux_ptr->mpi_rank,
     aux_ptr->dirty_bytes_propagations,
     aux_ptr->unprotect_dirty_bytes,
@@ -2188,6 +2207,7 @@ HDfprintf(stdout, "%d:H5AC_propagate...:%u: (u/uu/i/iu/r/ru) = %zu/%u/%zu/%u/%zu
 #endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */
 
 done:
+
     FUNC_LEAVE_NOAPI(ret_value)
 } /* H5AC__run_sync_point() */