The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.

Using Parquet Storage for Large Datasets

Introduction

When working with large brain connectome datasets, storing all matrices in memory as R lists can be inefficient. The riemtan package now supports Apache Arrow Parquet format for efficient storage and lazy loading of SPD matrices.

This vignette demonstrates how to:

Benefits of Parquet Storage

Installation

Make sure you have the arrow package installed:

install.packages("arrow")

Basic Workflow

1. Writing Connectomes to Parquet

First, let’s create some example SPD matrices and save them to Parquet format:

library(riemtan)
library(Matrix)

# Create example connectomes (4x4 SPD matrices)
set.seed(42)
connectomes <- lapply(1:50, function(i) {
  mat <- diag(4) + matrix(rnorm(16, 0, 0.1), 4, 4)
  mat <- (mat + t(mat)) / 2  # Make symmetric
  mat <- mat + diag(4) * 0.5  # Ensure positive definite
  Matrix::pack(Matrix::Matrix(mat, sparse = FALSE))
})

# Write to Parquet format
write_connectomes_to_parquet(
  connectomes,
  output_dir = "my_connectomes",
  subject_ids = paste0("subject_", 1:50),
  provenance = list(
    study = "Example Study",
    acquisition_date = "2024-01-01",
    preprocessing = "Standard pipeline v1.0"
  )
)

This creates a directory structure:

my_connectomes/
├── metadata.json
├── matrix_0001.parquet
├── matrix_0002.parquet
├── ...
└── matrix_0050.parquet

2. Validating Parquet Directory

Before using a Parquet directory, you can validate its structure:

# Detailed validation with verbose output
validate_parquet_directory("my_connectomes", verbose = TRUE)

3. Creating a CSample with Parquet Backend

Now use the Parquet backend to create a CSample:

# Load AIRM metric
data(airm)

# Create Parquet backend (default cache size: 10 matrices)
backend <- create_parquet_backend(
  "my_connectomes",
  cache_size = 10
)

# Create CSample with the backend
sample <- CSample$new(
  backend = backend,
  metric_obj = airm
)

# Sample info
print(paste("Sample size:", sample$sample_size))
print(paste("Matrix dimension:", sample$matrix_size))

4. Computing with Parquet-backed CSample

All standard CSample operations work transparently with the Parquet backend:

# Compute tangent images
sample$compute_tangents()

# Compute vectorized images
sample$compute_vecs()

# Compute Frechet mean
sample$compute_fmean(tol = 0.01, max_iter = 50)

# Center the sample
sample$center()

# Compute variation
sample$compute_variation()
print(paste("Variation:", sample$variation))

5. Lazy Loading Behavior

The Parquet backend loads matrices on-demand:

# Access specific connectome (loads from disk if not cached)
conn_1 <- sample$connectomes[[1]]

# Access all connectomes (loads all from disk)
all_conns <- sample$connectomes

# Cache management
backend$get_cache_size()  # Check current cache usage
backend$clear_cache()     # Clear cache to free memory

Working with CSuperSample

CSuperSample works seamlessly with Parquet-backed samples:

# Create two samples with different backends
backend1 <- create_parquet_backend("study1_connectomes")
backend2 <- create_parquet_backend("study2_connectomes")

sample1 <- CSample$new(backend = backend1, metric_obj = airm)
sample2 <- CSample$new(backend = backend2, metric_obj = airm)

# Create super sample
super_sample <- CSuperSample$new(list(sample1, sample2))

# Gather all connectomes
super_sample$gather()

# Compute statistics
super_sample$compute_fmean()
super_sample$compute_variation()

Backwards Compatibility

The Parquet backend is fully compatible with the traditional list-based approach:

# Traditional approach (all in memory)
sample_memory <- CSample$new(
  conns = connectomes,
  metric_obj = airm
)

# Parquet approach (lazy loading)
backend <- create_parquet_backend("my_connectomes")
sample_parquet <- CSample$new(
  backend = backend,
  metric_obj = airm
)

# Both work identically
sample_memory$compute_fmean()
sample_parquet$compute_fmean()

Performance Tips

Cache Size Tuning

Adjust cache size based on available memory:

# Small cache (memory-constrained environments)
backend_small <- ParquetBackend$new("my_connectomes", cache_size = 5)

# Large cache (memory-rich environments)
backend_large <- ParquetBackend$new("my_connectomes", cache_size = 50)

Batch Operations

For operations that need all matrices, consider loading in batches:

# Process in chunks
n <- sample$sample_size
batch_size <- 10

for (start in seq(1, n, by = batch_size)) {
  end <- min(start + batch_size - 1, n)

  # Load batch
  batch <- lapply(start:end, function(i) backend$get_matrix(i))

  # Process batch...

  # Clear cache to free memory
  backend$clear_cache()
}

Parallel Processing with Parquet

New in v0.2.5: riemtan now supports cross-platform parallel processing that works seamlessly with the Parquet backend.

Enabling Parallel Processing

Configure parallel processing using the futureverse framework:

library(riemtan)

# Enable parallel processing (works on all platforms including Windows!)
set_parallel_plan("multisession", workers = 4)

# Check status
is_parallel_enabled()  # TRUE
get_n_workers()        # 4

# Create Parquet-backed sample
backend <- create_parquet_backend("large_dataset", cache_size = 20)
sample <- CSample$new(backend = backend, metric_obj = airm)

Parallel I/O and Computation

All major operations automatically use parallel processing when beneficial:

# Parallel tangent computations with progress bar
sample$compute_tangents(progress = TRUE)   # 3-8x faster

# Parallel vectorization
sample$compute_vecs(progress = TRUE)       # 2-4x speedup

# Parallel Frechet mean computation
sample$compute_fmean(progress = TRUE)      # 2-5x faster for large samples

Batch Loading with Parallel I/O

For very large Parquet datasets, use batch loading with parallel I/O and memory management:

# Load specific subset in batches
subset_conns <- sample$load_connectomes_batched(
  indices = 1:500,        # Load first 500 matrices
  batch_size = 50,        # 50 matrices per batch
  progress = TRUE         # Show progress
)

# This loads 500 matrices in 10 batches, clearing cache between batches
# Each batch is loaded in parallel for 5-10x speedup

Performance Comparison

# Sequential (default if not configured)
set_parallel_plan("sequential")
system.time(sample$compute_tangents())  # Baseline

# Parallel with 4 workers
set_parallel_plan("multisession", workers = 4)
system.time(sample$compute_tangents())  # 3-4x faster

# Parallel with 8 workers
set_parallel_plan("multisession", workers = 8)
system.time(sample$compute_tangents())  # 6-8x faster

Progress Reporting

Enable progress bars to monitor long-running operations:

# Install progressr for progress bars (optional)
install.packages("progressr")

# Enable parallel processing
set_parallel_plan("multisession", workers = 4)

# All operations support progress parameter
sample$compute_tangents(progress = TRUE)
sample$compute_vecs(progress = TRUE)
sample$compute_fmean(progress = TRUE)

# Batch loading with progress
conns <- sample$load_connectomes_batched(
  indices = 1:1000,
  batch_size = 100,
  progress = TRUE  # Shows "Batch 1/10: loading matrices 1-100"
)

Best Practices

1. Choose appropriate worker count:

# Conservative (leave cores for system)
set_parallel_plan("multisession", workers = parallel::detectCores() - 1)

# Maximum performance (use all cores)
set_parallel_plan("multisession", workers = parallel::detectCores())

2. Memory management with parallelization:

# Each worker loads its own data copy
# For 4 workers with 100 matrices of 200x200, expect:
# Memory = 4 workers × cache_size × matrix_size ≈ 4 × 20 × 320 KB ≈ 25 MB

# Use smaller cache with more workers
backend <- create_parquet_backend("dataset", cache_size = 10)
set_parallel_plan("multisession", workers = 8)

3. Reset when done:

# Compute with parallelization
set_parallel_plan("multisession", workers = 4)
sample$compute_tangents(progress = TRUE)
sample$compute_fmean(progress = TRUE)

# Reset to free worker resources
reset_parallel_plan()

Auto-Detection

Parallel processing is automatically disabled for small datasets to avoid overhead:

# Small dataset (n < 10): Uses sequential processing automatically
small_sample <- CSample$new(conns = small_connectomes[1:5], metric_obj = airm)
small_sample$compute_tangents()  # Sequential (no overhead)

# Large dataset (n >= 10): Uses parallel processing automatically
large_sample <- CSample$new(backend = backend, metric_obj = airm)
large_sample$compute_tangents()  # Parallel (if plan is active)

Parallel Strategies

Different strategies for different environments:

# multisession: Works on all platforms (Windows, Mac, Linux)
set_parallel_plan("multisession", workers = 4)

# multicore: Unix only, lower overhead
set_parallel_plan("multicore", workers = 4)  # Auto-fallback to multisession on Windows

# cluster: For remote/distributed computing
set_parallel_plan("cluster", workers = c("node1", "node2"))

Metadata Access

Access metadata stored with your Parquet data:

# Get metadata
metadata <- backend$get_metadata()

# Access subject IDs
subject_ids <- metadata$subject_ids

# Access provenance information
provenance <- metadata$provenance
print(provenance$study)
print(provenance$preprocessing)

Advanced: Custom File Patterns

Customize the Parquet file naming pattern:

write_connectomes_to_parquet(
  connectomes,
  output_dir = "custom_naming",
  file_pattern = "conn_%03d.parquet"
)

# Files will be named: conn_001.parquet, conn_002.parquet, ...

Troubleshooting

Large Datasets

For very large datasets (1000+ matrices):

# Use minimal cache
backend <- ParquetBackend$new("large_dataset", cache_size = 3)

# Compute statistics without loading all matrices at once
sample <- CSample$new(backend = backend, metric_obj = airm)
sample$compute_tangents()
sample$compute_vecs()

# Operations that don't need all matrices in memory
sample$compute_fmean(batch_size = 32)  # Uses batching

Memory Monitoring

# Check cache usage
cache_size <- backend$get_cache_size()
print(paste("Cached matrices:", cache_size))

# Free memory when needed
backend$clear_cache()

Summary

The Parquet backend provides:

For small to medium datasets that fit in memory, the traditional list approach remains perfectly valid. For large brain connectome studies, the Parquet backend enables analysis at scale.

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.