rtree benchmark

A comprehensive understanding of the rtree's performance need a comprehensive benchmark.

Dave Moten provides us an awesome pure-java implementation of rtree, which supports multiple split strategies (e.g. quadratic and R*-tree huristics) and features (e.g. STR bulk loading and Flatbuffers backed structure). This project is created to present a comprehensive performance evaluation on the various rtree types.

Note: this is an on-going work, helps or suggests are very appreciated.

How to run

The project is based on jmh, and is supported to be used as the official guide recommends:

$ mvn clean install
$ java -jar target/microbenchmarks.jar

Important note: since currently this benchmark runs on a snapshot version of rtree (to test the feature of STR bulk loading), you may need to pull the latest code from master branch of the rtree project and build and deploy the snapshot maven library to your local environment.

Results

We conduct the experiments on a desktop server with:

4-core Intel(R) Core(TM) i5-4590 CPU @ 3.30GHz
16GB DDR3 Memory
CentOS Linux 7 with linux kernel 3.10.0-327.el7.x86_64
openjdk-1.8.0.71-2.b15.el7_2.x86_64

The experiments are conducted on two datasets:

Greek: a collection of some 38,377 entries corresponding to the epicentres of earthquakes in Greece between 1964 and 2000.
1k: uniformly generated 1000 rectangles.

The benchmarked rtree types are:

DefaultRTree: the original rtree implementation proposed by Guttman with quadratic split.
StarRTree: R*-tree huristics.
STRRTree: STR bulk loaded R*-tree, with a default full ratio (0.7) in each node.
STRFullRTree: STR bulk loaded R*-tree, with full nodes.
FBSRTree: R*-tree loaded from a Flatbuffers serialized byte array.

We test the performane of several operations on each type of rtree against various maxChildren parameter. TODO: we should test the performance against dataset size.

The results are plotted with plot.py and are presented as below.

Note: since the creation time of indexes span a large range, we use a logarithmic scale for the y axis.

Greek	1k

Full results are presented in rtreebm.txt.

Analysis

Creation

FBSRTree is much faster than others, because it just loads the context and the root node, and it lazily loaded the nodes when need.
STR-Rtree shows an order of magnitude of speedup comparing against the normal (R*-tree or DefaultRtree) creation process, and shows little difference about the node full ratio (but maybe differ in index size).
R-tree* is the slowest to construct.

Insertion

Default Rtree shows the best insertion performance, both in one insertion and batch insertion.
STR-Rtree is not bad in single insertion if not full. Full STR-Rtree works as bad as the FBSRtree for one insertion.
If there are a lot of insertions occurs (need node split anyway), the performance of STR-Rtree is similar to R*-tree.
FBS RTree is slow at insertion.
It seems that less children is better for insertion.

Deletion

It shows similar deletion performance for DefaultRTree and R*-tree (for deletion, STR-Rtree is also R*-tree).
FBS RTree is 10x slower at deletion.
It seems that 10 children per node is good for deletion.

Range Search

R*-tree shows the best search performance for most cases.
STR-Rtree is not bad, and non-full nodes shows better than full nodes.
FBS RTree is 10x slower at search.
Backpressure shows performance decline.
Small nodes (less children) is better at search for most cases.

Nearest Search

Currently (rtree-0.8-RC11-SNAPSHOT), the nearest search is based on a range search, therefore the results are similar.

ambling/rtree-benchmark