The first paper reading——Guttman, R-TREES A DYNAMIC INDEX STRUCTURE FOR SPATIAL SEARCHING

Guttman, R-TREES A DYNAMIC INDEX STRUCTURE FOR SPATIAL SEARCHING

When reading this paper, I didn't follow the standard three-phase process. Because I know I am new to this area, I can not step into the third phase, and I don't have the judgement ability to decide whether the paper is useful yet. So I mainly put my effort into the second phase.

Summary of the Paper

1. Introduction to background of spatial search, and state that there is no good method to deal with this problem yet. Then R-tree index structure is born.
2. Compare R-tree with several traditional index mechanism. Traditional index methods have drawbacks such as incapable of dealing with multi-dimensional data, or ignore the cost on disk-visiting, or not supporting well for dynamic structure.
3. The basic structure of R-tree.
   R-tree is an extension of B-tree which is nearly perfect in indexing one-dimensional data. But multi-dimensional data doesn't has a good total order, so linear alignment like B-tree is impossible. Thus the most different part of R-tree is the existence of containment and overlapping.
   
   Terms about R-tree.
   root, leaf, internal node, tuple-identifier, bounding box, child-pointer, node, entry
4. Basic operation on R-tree
   
   • Searching
     All variant of searching including searching existing entry, searching overlap or containment etc.
   • Insertion
     Insert is divided into three steps. Each steps follow the optimization principle of minimizing total area.
     • ChooseLeaf
     • Split Node
     • Adjust Tree
   • Deletion
     Deletion also involves three steps, say
     • Find Leaf
     • CondenseTree
     • Shorten Tree
5. More on splitting
   An efficient implementation of R-tree relies on a powerful split method. This method should be fast itself and provide fast access in searching process. And it should also fully utilize the space.
   A key to a good splitting is to minimize overlaps so that search on a certain area won't need to recursive on subtrees more than once.
   
   Guttman introduces three ways to implement splitting:
   
   1. Exhaustive
   2. Quadratic
   3. Linear
6. Performance of R-tree
   Several tests have been performed on the R-tree based on different parameter. Parameters are the size of m, the  page size, the splitting method adopted.
   Test result shows that R-tree is really efficient dealing with spatial search.
7. My Question
   
   1. The first time I read this paper I feel frustrated because there are so many unfamiliar terms and I can't get the meaning. Then I think about it, perhaps it is because I don't have enough background information. So I read about second storage system of DBMS, and review the B-tree, and think about some application of the spatial search. After that, I gradually grasp the meaning, and getting started.
   2. About the deletion. In the end, Guttman mentioned that high-level nodes must be placed higher in the tree, so that leaves of their dependent subtrees will be on the same level as leaves of the main tree. I don't get the meaning. What to be re-inserted is the entry in the leaf node but not the internal node, so what makes the difference?
   3. In splitting method, Guttman says overlapping slows down the search, but in his method, he choose to minimize the total area. Sometimes this works, but there are occasions this will contradict to minimizing overlapping principle. I think he is not convincing in the part. 
   4. About the test. The test is only performed on R-trees. I would not be convinced if only the author provide more comparisons with other index method.