Uses the shard package to place the full data matrix in POSIX shared
memory, enabling zero-copy parallel ROI extraction across furrr workers.
Activated per model spec via use_shard.
Details
When enabled, the default serial ROI-extraction loop in
mvpa_iterate is skipped.
Instead, each worker receives lightweight ALTREP handles that point into the
shared-memory segment and extracts its own ROI columns on demand.
This eliminates the two main memory bottlenecks of the default pipeline:
Serial
extract_roi()in the main process.Serialisation of large ROI matrices to workers via furrr.
Searchlight batches carry only neighborhood indices, so automatic shard
batches are larger than default-backend batches to amortize scheduling
overhead. Full garbage collection is not forced between shard batches by
default; set options(rMVPA.shard_gc_each_batch = TRUE) to restore
that conservative behavior when diagnosing memory pressure.
Supported dataset types
mvpa_image_dataset(volumetric)mvpa_surface_dataset(cortical surface)mvpa_clustered_dataset(parcellated)
Cleanup
Shared-memory segments are released automatically at the end of
mvpa_iterate().
You can also call shard_cleanup(mod_spec$shard_data) explicitly.