NDB wl-1533 tux optim 17 - allow slack in interior nodes

ndb/include/kernel/ndb_limits.h

@@ -110,7 +110,7 @@
  */
 #define MAX_TTREE_NODE_SIZE 64		// total words in node
 #define MAX_TTREE_PREF_SIZE 4		// words in min prefix
-#define MAX_TTREE_NODE_SLACK 3		// diff between max and min occupancy
+#define MAX_TTREE_NODE_SLACK 2		// diff between max and min occupancy
 
 /*
  * Blobs.

ndb/src/kernel/blocks/dbtux/Dbtux.hpp

@@ -591,7 +591,7 @@ private:
   void nodePopDown(Signal* signal, NodeHandle& node, unsigned pos, TreeEnt& ent);
   void nodePushDown(Signal* signal, NodeHandle& node, unsigned pos, TreeEnt& ent);
   void nodePopUp(Signal* signal, NodeHandle& node, unsigned pos, TreeEnt& ent);
-  void nodeSlide(Signal* signal, NodeHandle& dstNode, NodeHandle& srcNode, unsigned i);
+  void nodeSlide(Signal* signal, NodeHandle& dstNode, NodeHandle& srcNode, unsigned cnt, unsigned i);
   // scans linked to node
   void linkScan(NodeHandle& node, ScanOpPtr scanPtr);
   void unlinkScan(NodeHandle& node, ScanOpPtr scanPtr);
@@ -600,8 +600,19 @@ private:
   /*
    * DbtuxTree.cpp
    */
+  // add entry
   void treeAdd(Signal* signal, Frag& frag, TreePos treePos, TreeEnt ent);
+  void treeAddFull(Signal* signal, Frag& frag, NodeHandle lubNode, unsigned pos, TreeEnt ent);
+  void treeAddNode(Signal* signal, Frag& frag, NodeHandle lubNode, unsigned pos, TreeEnt ent, NodeHandle parentNode, unsigned i);
+  void treeAddRebalance(Signal* signal, Frag& frag, NodeHandle node, unsigned i);
+  // remove entry
   void treeRemove(Signal* signal, Frag& frag, TreePos treePos);
+  void treeRemoveInner(Signal* signal, Frag& frag, NodeHandle lubNode, unsigned pos);
+  void treeRemoveSemi(Signal* signal, Frag& frag, NodeHandle node, unsigned i);
+  void treeRemoveLeaf(Signal* signal, Frag& frag, NodeHandle node);
+  void treeRemoveNode(Signal* signal, Frag& frag, NodeHandle node);
+  void treeRemoveRebalance(Signal* signal, Frag& frag, NodeHandle node, unsigned i);
+  // rotate
   void treeRotateSingle(Signal* signal, Frag& frag, NodeHandle& node, unsigned i);
   void treeRotateDouble(Signal* signal, Frag& frag, NodeHandle& node, unsigned i);
 

ndb/src/kernel/blocks/dbtux/DbtuxDebug.cpp

@@ -178,16 +178,19 @@ Dbtux::printNode(Signal* signal, Frag& frag, NdbOut& out, TupLoc loc, PrintPar&
     out << "occupancy " << node.getOccup() << " of interior node";
     out << " less than min " << tree.m_minOccup << endl;
   }
-  // check missed half-leaf/leaf merge
+#ifdef dbtux_totally_groks_t_trees
+  // check missed semi-leaf/leaf merge
   for (unsigned i = 0; i <= 1; i++) {
     if (node.getLink(i) != NullTupLoc &&
         node.getLink(1 - i) == NullTupLoc &&
-        node.getOccup() + cpar[i].m_occup <= tree.m_maxOccup) {
+        // our semi-leaf seems to satify interior minOccup condition
+        node.getOccup() < tree.m_minOccup) {
       par.m_ok = false;
       out << par.m_path << sep;
       out << "missed merge with child " << i << endl;
     }
   }
+#endif
   // check inline prefix
   { ConstData data1 = node.getPref();
     Uint32 data2[MaxPrefSize];

ndb/src/kernel/blocks/dbtux/DbtuxMeta.cpp

@@ -211,11 +211,7 @@ Dbtux::execTUX_ADD_ATTRREQ(Signal* signal)
       // make these configurable later
       tree.m_nodeSize = MAX_TTREE_NODE_SIZE;
       tree.m_prefSize = MAX_TTREE_PREF_SIZE;
-#ifdef dbtux_min_occup_less_max_occup
       const unsigned maxSlack = MAX_TTREE_NODE_SLACK;
-#else
-      const unsigned maxSlack = 0;
-#endif
       // size up to and including first 2 entries
       const unsigned pref = tree.getSize(AccPref);
       if (! (pref <= tree.m_nodeSize)) {

ndb/src/kernel/blocks/dbtux/DbtuxNode.cpp

@@ -386,19 +386,20 @@ Dbtux::nodePopUp(Signal* signal, NodeHandle& node, unsigned pos, TreeEnt& ent)
 }
 
 /*
- * Move all possible entries from another node before the min (i=0) or
- * after the max (i=1).  XXX can be optimized
+ * Move number of entries from another node to this node before the min
+ * (i=0) or after the max (i=1).  Expensive but not often used.
  */
 void
-Dbtux::nodeSlide(Signal* signal, NodeHandle& dstNode, NodeHandle& srcNode, unsigned i)
+Dbtux::nodeSlide(Signal* signal, NodeHandle& dstNode, NodeHandle& srcNode, unsigned cnt, unsigned i)
 {
   Frag& frag = dstNode.m_frag;
   TreeHead& tree = frag.m_tree;
   ndbrequire(i <= 1);
-  while (dstNode.getOccup() < tree.m_maxOccup && srcNode.getOccup() != 0) {
+  while (cnt != 0) {
     TreeEnt ent;
     nodePopDown(signal, srcNode, i == 0 ? srcNode.getOccup() - 1 : 0, ent);
     nodePushUp(signal, dstNode, i == 0 ? 0 : dstNode.getOccup(), ent);
+    cnt--;
   }
 }
 

ndb/src/kernel/blocks/dbtux/DbtuxSearch.cpp

@@ -159,7 +159,7 @@ Dbtux::searchToAdd(Signal* signal, Frag& frag, ConstData searchKey, TreeEnt sear
  *
  * Compares search key to each node min.  A move to right subtree can
  * overshoot target node.  The last such node is saved.  The final node
- * is a half-leaf or leaf.  If search key is less than final node min
+ * is a semi-leaf or leaf.  If search key is less than final node min
  * then the saved node is the g.l.b of the final node and we move back
  * to it.
  */

ndb/src/kernel/blocks/dbtux/Times.txt

@@ -29,6 +29,7 @@ shows ms / 1000 rows for each and index time overhead
 samples 10% of all PKs (100,000 pk reads, 100,000 scans)
 
 the "pct" values are from more accurate total times (not shown)
+comments [ ... ] are after the case
 
 040616          mc02/a  40 ms   87 ms   114 pct
                 mc02/b  51 ms  128 ms   148 pct
@@ -76,13 +77,12 @@ optim 13        mc02/a  40 ms   57 ms    42 pct
                 mc02/c   9 ms   13 ms    50 pct
                 mc02/d 170 ms  256 ms    50 pct
 
-after wl-1884 store all-NULL keys (the tests have pctnull=10 per column)
-
 optim 13        mc02/a  39 ms   59 ms    50 pct
                 mc02/b  47 ms   77 ms    61 pct
                 mc02/c   9 ms   12 ms    44 pct
                 mc02/d 246 ms  289 ms    17 pct
 
+[ after wl-1884 store all-NULL keys (the tests have pctnull=10 per column) ]
 [ case d: bug in testOIBasic killed PK read performance ]
 
 optim 14        mc02/a  41 ms   60 ms    44 pct
@@ -98,8 +98,7 @@ none            mc02/a  35 ms   60 ms    71 pct
                 mc02/c   5 ms   12 ms   106 pct
                 mc02/d 165 ms  238 ms    44 pct
 
-[ johan re-installed mc02 as fedora gcc-3.3.2 ]
-[ case c: table scan has improved... ]
+[ johan re-installed mc02 as fedora gcc-3.3.2, tux uses more C++ stuff than tup]
 
 charsets        mc02/a  35 ms   60 ms    71 pct
                 mc02/b  42 ms   84 ms    97 pct
@@ -118,6 +117,16 @@ optim 15        mc02/a  34 ms   60 ms    72 pct
 optim 16        mc02/a  34 ms   53 ms    53 pct
                 mc02/b  42 ms   75 ms    75 pct
 
-[ case a, b: binary search of bounding node when adding entry ]
+[ binary search of bounding node when adding entry ]
+
+none            mc02/a  35 ms   53 ms    51 pct
+                mc02/b  42 ms   75 ms    76 pct
+
+[ rewrote treeAdd / treeRemove ]
+
+optim 17        mc02/a  35 ms   52 ms    49 pct
+                mc02/b  43 ms   75 ms    75 pct
+
+[ allow slack (2) in interior nodes - almost no effect?? ]
 
 vim: set et:

ndb/test/ndbapi/testOIBasic.cpp

@@ -2963,9 +2963,10 @@ tbuild(Par par)
       RUNSTEP(par, createindex, ST);
       RUNSTEP(par, invalidateindex, MT);
     }
-    RUNSTEP(par, readverify, MT);
+    RUNSTEP(par, pkupdate, MT);
+    RUNSTEP(par, readverify, ST);
     RUNSTEP(par, pkdelete, MT);
-    RUNSTEP(par, readverify, MT);
+    RUNSTEP(par, readverify, ST);
     RUNSTEP(par, dropindex, ST);
   }
   return 0;