srsadmm-core

This is the core library containing the distributed serverless ADMM algorithm, as well as binaries to solve a LASSO regression problem. It is meant for usage with the tokio runtime and the srsadmm-lambda-mm AWS Lambda function.

Documentation

docs.rs

Binaries

• Run the generate_problem binary to generate a problem instance.
• Run the lasso binary to solve the problem instance, or lasso_prox to solve the problem instance with the proximal gradient method. The binaries will pull your AWS credentials from the default environment variables. Please deploy the srsadmm-lambda-mm Lambda function to your AWS account first and ensure that it has access to the S3 bucket you are using. To enable a specific backend, run the binary as cargo run --release --bin <binary> --no-default-features --features <accelerate/netlib/openblas> -- <args>.

Features

• accelerate - Use the accelerate backend for matrix operations
• netlib - Use the netlib backend for matrix operations
• openblas - Use the openblas backend for matrix operations
• linfa - Adds a utility function to compute the optimal objective value for Lasso regression using the linfa and linfa-elasticnet libraries. Useful for testing and validation.
• rayon - Adds support for parallelization using the rayon library. Really not necessary for the ADMM algorithm, but speeds up the problem instance generation.

Usage

Alternatively, you can use the srsadmm-core library in your own project. While you can technically install it with cargo add srsadmm-core, it might be better to directly copy this directory into your own project and use it as a dependency. Take a look at lasso.rs and lasso_prox.rs for examples of how to use the library. You can also install with features accelerate, netlib, or openblas to enable different backends.

Recommended install: cargo add srsadmm-core --features openblas,rayon

Core ADMM Framework

ADMMProblem<G, S> - The main trait defining the ADMM algorithm interface. Implementations must provide methods for:

• precompute() - One-time setup and matrix factorizations
• update_x() - Primal variable update step
• update_z() - Auxiliary variable update (often with proximal operators)
• update_y() - Dual variable update step
• update_residuals() - Compute convergence metrics
• check_stopping_criteria() - Determine if algorithm should terminate

ADMMSolver<G, S, P> - Orchestrates the iterative ADMM optimization process with timing tracking, iteration control, and result export capabilities.

ADMMContext<G, S> - Execution context containing shared global state (G) and local subproblem state (S) with thread-safe synchronization primitives.

Distributed Matrix Variables

MatrixVariable - A distributed matrix that can be stored and synchronized across multiple backends (local disk, S3, memory). Provides high-level matrix operations for ADMM algorithms while handling distributed storage complexity. Key features:

• Multi-location storage (Local, S3, Memory) with automatic synchronization
• Lazy loading and caching strategies
• Matrix operations: addition, subtraction, multiplication, inversion
• Subproblem-aware row-wise operations with parallel processing

DataMatrixVariable - Specialized for large read-only matrices (like training data) with memory-mapped file access and efficient chunking for subproblem processing.

ScalarVariable - Similar distributed storage for scalar values with the same multi-backend synchronization.

Storage and Resource Management

ResourceLocation - Enum defining storage backends:

• Local - Compressed local filesystem storage
• S3 - AWS S3 cloud storage for distributed access
• Memory - In-memory storage for fast access

StorageConfig - Configuration for all storage backends with settings for local paths, S3 buckets, and memory management.

ProblemResourceImpl<T> - Internal resource manager handling storage, synchronization, and caching across multiple backends with automatic consistency management.

Matrix Operations

ops module provides distributed matrix operations:

• mm() - Matrix multiplication (local or cloud-based via AWS Lambda)
• lasso_factor() - Computes (A^T A + ρI)^-1 for LASSO problems
• soft_threshold() - L1 regularization proximal operator
• scale() - In-place matrix scaling

Performance and Monitoring

TimingTracker - Performance tracking for ADMM iterations, outputting a CSV file with the timing data.

Subproblem Management

subproblem module provides utilities for decomposing problems:

• split_matrix_into_subproblems() - Partitions matrices into row-wise chunks for parallel processing
• combine_subproblems() - Reassembles subproblem results into final solution