[Bigdata-commit] SF.net SVN: bigdata:[7357] branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/ java/com/bigdata/rdf/graph/impl

[<tho...@us...>]

Revision: 7357
          http://bigdata.svn.sourceforge.net/bigdata/?rev=7357&view=rev
Author:   thompsonbry
Date:     2013-08-28 20:01:46 +0000 (Wed, 28 Aug 2013)
Log Message:
-----------
Modified StaticFrontier2 to clear out any non-null elements in the backing array.  This addresses a GC leak during the computation.

Modified TLScheduler to clear the map of per-thread schedulers rather than clearing each per-thread scheduler. This is a big hot spot.

See #629

Modified Paths:
--------------
    branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/StaticFrontier2.java
    branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/scheduler/TLScheduler.java

Modified: branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/StaticFrontier2.java
===================================================================
--- branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/StaticFrontier2.java	2013-08-28 19:47:36 UTC (rev 7356)
+++ branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/StaticFrontier2.java	2013-08-28 20:01:46 UTC (rev 7357)
@@ -119,6 +119,18 @@
         }
 
         /*
+         * Null fill until the end of the last frontier. That will help out GC.
+         * Otherwise those IV references are pinned and can hang around. We
+         * could track the high water mark on the backing array for this
+         * purpose.
+         */
+        for (int i = nvertices; i < a.length; i++) {
+            if (a[i] == null)
+                break;
+            a[i] = null;
+        }
+        
+        /*
          * Take a slice of the backing showing only the valid entries and use it
          * to replace the view of the backing array.
          */

Modified: branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/scheduler/TLScheduler.java
===================================================================
--- branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/scheduler/TLScheduler.java	2013-08-28 19:47:36 UTC (rev 7356)
+++ branches/BIGDATA_RELEASE_1_2_0/bigdata-rdf/src/java/com/bigdata/rdf/graph/impl/scheduler/TLScheduler.java	2013-08-28 20:01:46 UTC (rev 7357)
@@ -122,13 +122,16 @@
         /*
          * Clear the per-thread maps, but do not discard. They will be reused in
          * the next round.
+         * 
+         * FIXME This is a big cost. Try simply clearing [map] and see if that
+         * is less expensive.
          */
-        for (STScheduler s : map.values()) {
-
-            s.clear();
-
-        }
-
+//        for (STScheduler s : map.values()) {
+//
+//            s.clear();
+//
+//        }
+        map.clear();
     }
 
     @Override

This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.