Uploaded image for project: 'Lucene - Core'
  1. Lucene - Core
  2. LUCENE-9087

Should the BKD tree use a fixed maxPointsInLeafNode?

    XMLWordPrintableJSON

    Details

    • Type: Improvement
    • Status: Closed
    • Priority: Major
    • Resolution: Fixed
    • Affects Version/s: None
    • Fix Version/s: 8.6
    • Component/s: None
    • Labels:
      None
    • Lucene Fields:
      New

      Description

      Currently the BKD tree uses a fixed maxPointsInLeafNode provided in the constructor. For the current default codec the value is set to 1024. This is a good compromise between memory usage and performance of the BKD tree.

      Lowering this value can increase search performance but it has a penalty in memory usage. Now that the BKD tree can be load off-heap, this can be less of a concern. Note that lowering too much that value can hurt performance as well as the tree becomes too deep and benefits are gone.

      For data types that use the tree as an effective R-tree (ranges and shapes datatypes) the benefits are larger as it can minimise the overlap between leaf nodes. 

      Finally, creating too many leaf nodes can be dangerous at write time as memory usage depends on the number of leaf nodes created. The writer creates a long array of length = numberOfLeafNodes.

      What I am wondering here is if we can improve this situation in order to create the most efficient tree? My current ideas are:
       

      • We can adapt the points per leaf depending on that number so we create a tree with the best depth and best points per leaf. Note that for the for 1D case we have an upper estimation of the number of points that the tree will be indexing. 
      • Add a mechanism so field types can easily define their best points per leaf. In the case, field types like ranges or shapes can define its own value to minimise overlap.
      • Maybe the default is just too high now that we can load the tree off-heap.

      Any thoughts?

        Attachments

          Issue Links

            Activity

              People

              • Assignee:
                ivera Ignacio Vera
                Reporter:
                ivera Ignacio Vera
              • Votes:
                0 Vote for this issue
                Watchers:
                5 Start watching this issue

                Dates

                • Created:
                  Updated:
                  Resolved:

                  Time Tracking

                  Estimated:
                  Original Estimate - Not Specified
                  Not Specified
                  Remaining:
                  Remaining Estimate - 0h
                  0h
                  Logged:
                  Time Spent - 1.5h
                  1.5h