This repository contains a set of benchmarks to compare the performance of different independent HDF5 implementations against the original HDF5 C library.
Every time there is a commit to the master branch of this repository, the benchmarks are run in the GitHub Actions CI pipeline and the graphs are updated with the new results.
The original C library is not included directly, but through the thin C# wrapper HDF5.PInvoke.1.10, which should not affect the benchmark performance.
Please create a PR if you want to add more benchmarks or improve existing ones.
You can view the graphs here: https://apollo3zehn.github.io/hdf5-benchmark/
Add --recurse-submodules to your git clone command to clone the submodules as well. See .github/workflows/benchmark.yml to see how to run the individual benchmarks.
Chunk caches are very important to improve performance. However, they also impact benchmarks negatively in that when the benchmark repeats the read operation on the very same dataset multiple times, no real read operation happens anymore and the benchmark result becomes useless. Thus the following benchmark principles need to be applied, especially to the baseline project:
-
1D array: no cache needed as all data are read only cache
-
2D array: make cache size at least the size of the dataset and reopen the file and the dataset for each loop iteration (but exclude this from the actual benchmark!).
Closing and reopening is currently not required for PureHDF because the lifetime of the cache is only during the read operation itself. However, this might change in future.