This is the repository of the TinSpin benchmark framework. Click here for the TinSpin spatial index collection.
TinSpin is a framework for benchmarking in-memory spatial indexes.
TinSpin provides several dataset generators for point data and rectangle data. The datasets can be scaled with size and dimensionality. Each index can be tested with various loads, such as insertion, window queries, exact match queries, nearest neighbor queries, updates (moving objects) and removal.
The framework was originally developed at ETH Zurich in the GlobIS group. It is now maintained by Tilmann Zäschke.
If you want to reference this project, please consider referencing the PH-Tree instead, as TinSpin was originally developed to benchmark the PH-Tree and because there is currently no TinSpin publication: The PH-tree: A Space-Efficient Storage Structure and Multi-Dimensional Index, T. Zäschke, C. Zimmerli and M.C. Norrie, Proc. of the 2014 ACM SIGMOD Intl. Conf. on Management of Data.
2023-07:
- Updated to new tinspin-indexes 2.0.0 with new API
- Indexes now all hold a unique identifier (int) as value instead of a singleton Object.
- Added support for multimap tests. Note: this allows testing of multimap, however, the generated test data does currently not contain duplicates.
2018-12:
- Major refactoring with modularization, new Maven module 'tinspin-common'
- Added new Example for custom testing
- Added (very basic!) import for HDF5 files.
- Testing: Updated Window Query generator again; Use result count only from first run, for comparability; Updated timings to use nanoSecs internally
- CHANGE: WQ timings now return time/query instead of time/result
2018-07:
- New logging output with ops/sec instead of time/op
- Operations without side effects (window query, point query, kNN search) are repeatedly executed until a minimum time has passed (in order to avoid problems with warm-up)
The TestManager
allows running many tests in one session. Output files are written to subfolders,
see below.
The TestRunner
can run individual tests, all output is written to console. This is mostly useful for debugging. The
meaning of the output is described below, the second row uses the new TinSpin 1.0 format
(the first row uses the old TinSpin 0.0.1 format).
The Example.java class demonstrates how to test you own index.
The data
folder contains plugins for importing data from various formats.
The db
folder contains a "cache" that can be used for file formats that are slow to import. It allows importing to a
DB file (binary format) which is faster to read than text/xml files such as OSM.
The wrappers
folder contains wrappers for various index implementation and different data types: point data, rectangle
data. *MM*
wrappers support multimap indexes. Note that generated data does NOT contain duplicates, i.e. the
generators generate at most one entry for each coordinate (point) or pair of coordinates (rectangle).
Some results can be found in the doc folder.
Visualized results from a complete run with 2D and 3D data can be found here (PDF).
The file benchmark-high-dim-2018-12.ods contains benchmarks with high-dimensional datasets, such as GloVe-25, GloVe-50, SIFT-128, NYTimes-256 and MNIST-784 ( see ann-benchmark).
The test data is written to tab-separated value files in target/logs.
Log files from tests are written to several possible folders. The default folders are:
dimsP
: Point data scaled with dimensionalitydimsR
: Rectangle data scaled with dimensionalitysizeP
: Point data scaled with dataset sizesizeR
: Rectangle data scaled with dataset sizesizePWQS
: Point data scaled with size of query windowsizeRWQS
: Rectangle data scaled with size of query window
- column names:
Index data ...
- comment:
% Averages
- comment:
% ========
- averaged results:
RSZ-R-AVG-3/3 ...
(Index=RSZ, datatype=Rectangle, average of 3 successful runs of 3 initiated runs - comment:
% Measurements
- comment:
% ============
- results:
RSZ-R-0 CLUSTER(5.0,0.0,null) ...
(Index=RSZ, datatype=Rectangle, random seed=0 (equals test run ID) - results new:
RSZ-R-M-0 CLUSTER(5.0,0.0,null) ...
(Index=RSZ, datatype=Rectangle, map/multimap=Map, random seed=0 (equals test run ID)
By default, TinSpin averages three consecutive test runs into one average.
Columns with general information. An example spreadsheet file for interpreting and visualizing results can be found here, e.g. 2nd sheet.
Index
: Index and test descriptor, such asRSZ-R-M
for rectangle index orPHT-P-MM
for a point multimap index.data
: Test data descriptor, Such asCUBE(4, 1.0, 0.0, null)
, see abovedim
: number of data dimensionsbits
: number of bits (deprecated), always 64N
: dataset size (number of points or rectangles)memory
: Total measured JVM memory [bytes]memory/n
: Total measured JVM memory [bytes per entry]
Columns with timing. Most parts of the test are executed in two runs or more, each run consisting of a predefined
number of
execution. For example, each exact match run consists of 100,000 exact match queries as defined in the TestStats
class.
Except for load/unload, runs are repeated until at least 2 seconds (default) have passed, this is in order to give more
precise timings for very short runs.
gen
: time [ms] for dataset generationload/s
: adding entries throughput [entries/s]wq1/s
: window query throughput run #1 [queries/s]wq2/s
: window query throughput run #2 [queries/s]pq1/s
: exact match query throughput run #1 [queries/s] (was called point query)pq2/s
: exact match query throughput run #2 [queries/s]1-NN1/s
: 1 nearest neighbor query throughput run #1 [queries/s]1-NN2/s
: 1 nearest neighbor query throughput run #2 [queries/s]10-NN1/s
: 10 nearest neighbor query throughput run #1 [queries/s]10-NN2/s
: 10 nearest neighbor query throughput run #2 [queries/s]up1/s
: position update throughput run #1 [updates/s]up2/s
: position update throughput run #2 [updates/s]unload/s
: removing entries throughput [entries/s]
Columns with Tree statistics. The following columns contain tree statistics, such as number of nodes or depth. The meaning may differ between trees.
nodes
: Number of nodespostLen
: PH-tree: Average length of postfixes; R-Tree & Quadtrees: depthAHC
: PH-tree:Number of AHC nodesNT
: PH-tree:Number of NT-NodesNTinternal
: PH-tree: Number of NT-subnodes in all NT-Nodes
Columns with result statistics. The following columns give an indicator of the results returned by the first test run, even if runs are repeated if they are faster than 2 seconds (default), see above. The counts should not vary much between runs. This is a basic form of correctness testing. For a given test scenarion, results should be the same for all indexes.
wq1-n
: Number of returned window query objectswq2-n
: Number of returned window query objectsq1p-n
: Number of found objects in point queryq2p-n
: Number of found objects in point query1-NN1-d
: Average distance over all 1-nearest neighbors1-NN2-d
: Average distance over all 1-nearest neighbors10-NN1-d
: Average distance over all 10-nearest neighbors10-NN2-d
: Average distance over all 10-nearest neighborsup1-n
: Number of updated objectsup2-n
: Number of updated objects
Columns with JVM statistics For each test part, the following column contain garbage collection statistics based on
Java instrumentation. They are a
good indicator, but not precise! -s
is the estimated memory [MB] freed up by GC. -t
is the estimated time
used by the GC in [ms].
load-s
: Estimated size of garbage collected memory [MB]load-t
: Estimated CPU time of garbage collector [ms]w-query-s
w-query-t
p-query-s
p-query-t
update-s
update-t
1-NN-s
1-NN-t
10-NN-s
10-NN-t
unload-s
unload-t
General messages column
msg
: General messages by index wrapper and test runner
In general:
- Index: Index and test descriptor, such as
RSZ-R
for rectangle index, see above - data: Test data descriptor, Such as
CUBE(1.0,0.0,null)
, see above - dim: number of data dimensions
- bits: number of bits (deprecated), always 64
- N: dataset size (number of points or rectangles)
- calcMem: Estimated memory requirement [bytes] per entry, only PH-Tree variants
- memory: Total measured JVM memory [bytes]
- memory/n: Total measured JVM memory [bytes per entry]
Timing. Most parts of the test are executed in two runs or more, each run consisting of a predefined number of
execution. For example, each exact match run consists of 100,000 exact match queries as defined in the TestStats
class.
Except for load/unload, runs are repeated until at least 2 seconds (default) have passed, this is in order to give more
precise timings for very short runs.
- gen: time [ms] for dataset generation
- load: total index loading time [ms]
- load/n: average loading time [micro-s/entry]
- q1/n: window query time run #1 [micro-s/returned entry]
- q2/n: window query time run #2 [micro-s/returned entry]
- pq1/n: exact match query time run #1 [micro-s/query] (was called point query)
- pq2/n: exact match query time run #2 [micro-s/query]
- up1/n: update time run #1 [micro-s/update]
- up2/n: update time run #2 [micro-s/update]
- 1-NN1: query time run #1 [micro-s/update]
- 1-NN2: query time run #2 [micro-s/update]
- 10-NN1: query time run #1 [micro-s/update]
- 10-NN2: query time run #2 [micro-s/update]
- unload: total index unloading time [ms]
- unload/n: average removal time [micro-s/entry]
Tree statistics. The following columns contain tree statistics, such as number of nodes or depth. The meaning may differ between trees.
- nodes: Number of nodes
- postLen: PH-tree: Average length of postfixes; R-Tree & Quadtrees: depth
- AHC: PH-tree:Number of AHC nodes
- NT: PH-tree:Number of NT-Nodes
- NTinternal: PH-tree: Number of NT-subnodes in all NT-Nodes
Result statistics. The following columns give an indicator of the result returned by the first test run, even if runs are repeated if they are faster than 2 seconds (default), see above. The counts should not vary much between runs, but using the first runs allows comparing the result counts of different tree as a basic form of correctness testing.
- q1-n: Number of returned window query objects
- q2-n: Number of returned window query objects
- q1p-n: Number of found objects in point query
- q2p-n: Number of found objects in point query
- d1-1NN: Average distance of nearest neighbors
- d2-1NN: Average distance of nearest neighbors
- d1-kNN: Average of sum of distance of 10 nearest neighbors
- d2-kNN: Average of sum of distance of 10 nearest neighbors
- up1-n: Number of updated objects
- up2-n: Number of updated objects
- distCalc-n : Number of distance calculation for insert, query and deletion
- distCalc1NN-n : Number of distance calculations of 1NN queries
- distCalcKNN-n : Number of distance calculations of kNN queries
For each test part, the following column contain garbage collection statistics based on Java instrumentation. They are a
good indicator, but not precise! -s
is the estimated memory [MB] freed up by GC. -t
is the estimated time in [ms]
used by the GC.
- load-s: Estimated size of garbage collected memory [MB]
- load-t: Estimated runtime of garbage collector [ms]
- w-query-s
- w-query-t
- p-query-s
- p-query-t
- update-s
- update-t
- 1-NN-s
- 1-NN-t
- 10-NN-s
- 10-NN-t
- unload-s
- unload-t
General messages column:
- msg: General messages by index wrapper and test runner