HIDEFIX

This library provides an alternative reader for the HDF5 file or NetCDF4 file (which uses HDF5) which supports concurrent access to data. This is achieved by building an index of the chunks, allowing a thread to use many file handles to read the file. The original (native) HDF5 library is used to build the index, but once it has been created it is no longer needed. The index can be serialized to disk so that performing the indexing is not necessary.

use hidefix::prelude::*;

let idx = Index::index("tests/data/coads_climatology.nc4").unwrap();
let mut r = idx.reader("SST").unwrap();

let values = r.values::<f32>(None, None).unwrap();

println!("SST: {:?}", values);

Motivation

The HDF5 library requires internal locks to be thread-safe since it relies on internal buffers which cannot be safely accessed/written to from multiple threads. This effectively causes multi-threaded applications to use sequential reads, while competing for the locks. And also apparently cause each other trouble, perhaps through dropping cached chunks which other threads still need. It can be safely used from different processes, but that requires potentially much more overhead than multi-threaded or asynchronous code.

Some basic benchmarks

hidefix is intended to perform better when concurrent reads are made either to the same dataset, same file or to different files from a single process. For basic benchmarks the performance is on-par or slightly better compared to doing standard sequential reads than the native HDF5 library (through its rust-bindings). Where hidefix shines is once the multiple threads in the same process tries to read in any way from a HDF5 file simultaneously.

This simple benchmark tries to read a small dataset sequentially or concurrently using the cached reader from hidefix and the native reader from HDF5. The dataset is chunked, shuffled and compressed (using gzip):

$ cargo bench --bench concurrency -- --ignored

test shuffled_compressed::cache_concurrent_reads  ... bench:  15,903,406 ns/iter (+/- 220,824)
test shuffled_compressed::cache_sequential        ... bench:  59,778,761 ns/iter (+/- 602,316)
test shuffled_compressed::native_concurrent_reads ... bench: 411,605,868 ns/iter (+/- 35,346,233)
test shuffled_compressed::native_sequential       ... bench: 103,457,237 ns/iter (+/- 7,703,936)

Inspiration and other projects

This work is based in part on the DMR++ module of the OPeNDAP Hyrax server. The zarr format does something similar, and the same approach has been tested out on HDF5 as swell.