Benchmarks

Benchmarks answer two practical questions:

• How fast can TSEdge write and read local time-series data?
• How much space do compressed segment files save compared with simple formats?

The numbers are meant to be reproducible and useful for comparison. They are not a claim that TSEdge is faster than every existing time-series database.

Setup

Common benchmark runs use 1,000,000 points.

Datasets:

• smooth: slowly changing values with small noise.
• noisy: random values.
• step: values that stay constant for a range of points, then change.
• constant: identical values.
• irregular_timestamps: non-uniform timestamp intervals.

Systems:

• TSEdge.
• Raw binary files.
• CSV files.
• SQLite, when SQLite support is available during the build.

Metrics:

• storage size
• compression ratio
• write throughput
• read throughput
• AVG aggregate time

Running

Build the benchmark programs first:

mkdir build
cd build
cmake ..
cmake --build .

Then run the main comparison programs:

./tsedge_bench 1000000
./file_bench 1000000
./sqlite_bench 1000000

If SQLite was not found by CMake, sqlite_bench is not built. The TSEdge and file benchmarks can still run.

The benchmark programs can also write CSV result files for notebook or spreadsheet analysis.

Block-size tuning results

Block size controls how many points are compressed into one block. Small blocks can reduce extra decoding for tiny ranges. Larger blocks reduce metadata overhead and often improve full scans.

The following table reflects the benchmark workload used when the default block size was selected.

Block size, points	Batch fast, points/sec	Append fast, points/sec	Tiny read, sec	Small read, sec	Full read, points/sec	AVG full, sec	MIN/MAX full, sec	Compression ratio	Bytes/point	Block count
1024	5.14M	3.06M	0.000021	0.000043	47.35M	0.013701	0.013135	1.5868	10.083	977
4096	12.81M	6.82M	0.000039	0.000077	83.12M	0.003304	0.003385	1.5967	10.021	245
8192	16.12M	7.90M	0.000057	0.000111	103.26M	0.001582	0.001677	1.5983	10.010	123
16384	16.45M	8.25M	0.000091	0.000092	115.31M	0.000830	0.000795	1.5992	10.005	62
32768	17.70M	6.12M	0.000170	0.000175	121.70M	0.000439	0.000412	1.5996	10.003	31

The default block size is 16384 points.

For the current workload, it gives a good balance between write throughput, full-range reads and small-range query overhead.

Additional benchmark details are available in the repository's benchmark notes and scripts.