Tools for defining and running large computations on DiskArrays.
The package DiskArrays.jl implements Julia's AbstractArray interface for chunked (and possibly compressed) n-dimensional arrays that are stored on disk and operated on lazily.
Although DiskArrays.jl provides basic implementations for e.g. broadcasting or reductions over dimensions it has clear limitations when it comes to parallel computations or when broadcasting over arrays from different sources with non-aligning chunks. With DiskArrayEngine intend to provide a general-purpose computing backend that scales to very large n-dimensional arrays (GBs, TBs or larger) typically stored in a DiskArrays.jl-supported format like NetCDF, Zarr, ArchGDAL, HDF5Utils etc with parallelism supported by Dagger.jl.
Before starting to jump into this package it is worth checking if it is actually the right tool for your problem. Here is a quick check-list of things to consider and possible alternatives:
- Your data is too large to fit into one machine's memory (otherwise just use normal Julia Arrays)
Mmapis not an option (e.g. because your data comes in a compressed format, or data is stored in the cloud or your queueing system sees unrealistic memory usage by mmap)- Your data is too large to fit into the memory of all your workers when distributed among them (otherwise try DistributedArrays.jl)
- You want to process all or almost all of your data and not just a small subset. Otherwise just read the subset of interest into memory and do your processing based on this one
If you are still here you should also note that this package is not intended to be used by end-users directly, but the plan is to wrap functionality from this package in other packages, in particular YAXArrays.jl, DimensionalData.jl or PyramidSchemes.jl that provide more user-friendly interfaces for the end users.
This package is still under active development and should be considered experimental. Expect things to break and to already be broken. In particular, extensive documentation and tests are still missing. However, some core functionality of the package is already used by e.g. PyramidScheme.jl which is why we decided to already register this package while still under active development.
To be done, describe the generalized moving window concept, how to define user functions, lazy interface and which runner options exist
The simplest way to use some the machinery in DiskArrayEngine is to wrap any existing DiskArray into an EngineArray by calling engine(mydiskarray). Afterwards, many operations like mapslices, mapreduce, broadcast and simple statistics like mean, median, max/min etc will be dispatch using DiskArrayEngine instead of the simple
DiskArrays.jl implementation and might give significant speedups. However, currently we still default to using the LocalRunner, which will only use a single process, we will experiment with defaulting to DaggerRunner as soon as multiple processes are available.