Good practice of using chopin

Minimize errors

Consider using try() or tryCatch() to let an error will not halt all the work without results, especially when using higher-level functions.

Raster workflow: stacked vs file-based parallelization

!NOTE It’s up to the users’ system specification and the size of the data.

For faster computation, there are two strategies. One is parallelization which is implemented in chopin for example. The other strategy is to stack rasters and extract values at once. Adjustment of arguments in exactextractr::exact_extract will benefit many people who are dealing with sizable data that are able to be stacked. We compare the two strategies in terms of computation time.

Before proceeding, users need to consider the hardware specification. For example, memory and storage to leverage the maximal performance of exact_extract. Specifically speaking, memory capacity is crucial to store the stacked rasters in memory rather than to read the proxy of rasters from the disk as implemented in terra. max_cells_in_memory is a key argument to control the memory usage. The maximum possible value for this argument is 2³¹ − 1 = 2, 147, 483, 647, roughly 2.147e9, as applied in the example above. As memory bandwidth is much faster than disk input/output specification, the stacked rasters with high max_cells_in_memory applied will run faster than file-based parallelization or the extraction with lower value of max_cells_in_memory. The performance does not come without a cost. The memory-intensive setting is not suitable for the system with limited memory, for example, in consumer laptops with around 16 GB of RAM.

!NOTE Stacking rasters usually takes the large amount of memory. Users need to consider the memory capacity of the system before stacking rasters.

Raster-Vector overlay

Use chopin::par_pad_grid() to make exhaustive (fully filling the entire extent without any gaps) and padded grid objects in SpatVector class. chopin::par_grid() filters features that intersect with each grid. Post-processing is necessary as polygons may be taken more than twice in parallelized calculation. If users have a field containing divisible and hierarchical information, dataset can be processed with chopin::par_hierarchy().

Customization

chopin way to parallelize is to iterate calculation over a list then to use future.apply function. If one wants to customize parallelization workflow like chopin, consider making list objects in two steps. First, assign extent vectors to each element. In this case, users should be aware of the coordinate system of both the input data and the coordinates of the extent vectors. Second, use preprocessed vector objects with respect to each extent vector in the first list. We recommend to name each list element to debug easily. For par_multirasters() workflow, using file or database table paths is recommended instead of SpatRaster objects in future parallel processing workflow. This is because SpatRaster objects made from external files are Rcpp pointers, which cannot be exported to processes in parallelization.

More tips to save time and memory

1. Read vector data with sf package at first. It is a bit faster than terra and will save time for processing as exactextractr is designed to work with sf objects for vector inputs.
2. If your analysis does not require the high precision of vector data, simplification of geometries (e.g., using rmapshaper, i.e., ms_simplify) will result in considerable time savings.

Save computing costs

If HPC systems are used, you will have CPU-memory usage quota per user. Thus, it is always recommended to test your computation workflow with subsets in manageable sizes. It can give you an estimate for the total computational demand. profvis::profvis() helps to summarize runtime per function call and memory usage. For testing, use future::plan(future::sequential) to see single-thread. Keep chopin::par_pad_grid() arguments the same then only take one or two grids using lapply. Preferably the hardest case will be tested to estimate the maximum peak memory usage per thread. After prototyping, take the numbers as reference numbers to submit a job to HPC with a proper amount of computational assets: each HPC management system offers tools to see statistics and task summaries. Users are advised to use these records to plan asset allocation for the next time.