`ProteinStructureStore` speed

[AntonOresten]

ProteinStructureStore uses HDF5 for lazy IO of structures, but HDF5.jl is just a wrapper for some pre-built binary that doesn't support parallelization. This effectively bottlenecks the IO speed. A rough test showed that for a dataset with ~20 properties, ~300KB per structure, reading 100 structures takes ~1 second.

At the moment, this format and structure is viable for repositories of protein structures with chain and residue-wise information that is otherwise expensive to gather. It might not be optimal for direct use in workflows that require high throughput protein data look-ups.

Programs that require fast IO might have to serialize into some faster intermediate format.

[AntonOresten]

v0.6 uses JSON to serialize dictionary and nested vector properties, which should reduce calls to the HDF5 binary. Also cleaner file structure and less nesting which might help a little.

[AntonOresten]

Seems that cancelling a HDF5 reading loop in the REPL makes the next loop run waay slower, meaning long read/write runs should be done in clean sessions or as scripts. In ideal settings we can at least reach a stable >400 single-chain structures/second.

[AntonOresten]

Switching to JLD2 in v0.7. It's so much better. Similar writing speed, but doesn't randomly precompile/get slower, and ~5x faster reading.