mesh-sieve

mesh-sieve is a modular, high-performance Rust library for mesh and data management, designed for scientific computing and PDE codes. It provides abstractions for mesh topology, field data, parallel partitioning, and communication, supporting both serial and MPI-based distributed workflows.

Features

• Mesh Topology: Flexible, generic Sieve data structures for representing mesh connectivity.
• Field Data: Atlas and Section types for mapping mesh points to data arrays.
• Parallel Communication: Pluggable communication backends (serial, Rayon, MPI) for ghost exchange and mesh distribution.
• Partitioning: Built-in support for graph partitioning (Metis, custom algorithms).
• MPI Integration: Examples and tests for distributed mesh and data exchange using MPI.
• Extensive Testing: Serial, parallel, and property-based tests.

Getting Started

To build and run the project:

cargo build

cargo run

Running MPI Examples

Some features require MPI. To run an MPI example (e.g., mpi_complete.rs):

mpirun -n 2 cargo run --example mpi_complete

Other examples:

• mesh_distribute_two_ranks.rs: Demonstrates mesh distribution across two ranks.
• mpi_complete_no_overlap.rs, mpi_complete_multiple_neighbors.rs, mpi_complete_stack.rs: Test various parallel completion scenarios.

Project Structure

• src/: Library source code
  • topology/: Mesh topology and Sieve traits
  • data/: Atlas and Section for field data
  • overlap/: Overlap and Delta for ghost exchange
  • algs/: Algorithms for communication, distribution, completion, partitioning
  • partitioning/: Graph partitioning algorithms
• examples/: Usage and integration tests (serial and MPI)
• tests/: Unit and integration tests
• benches/: Benchmarks
• Cargo.toml: Project manifest and dependencies

Optional Features

• mpi-support: Enable MPI-based communication and parallel tests
• metis-support: Enable Metis-based partitioning (requires Metis and pkg-config)

Version 1.2.4 — Release Notes

Highlights

• Add distributed closures via completion: call closure_completed(...) to traverse across partitions. It fetches remote cones/supports on demand or upfront (policy-driven) and fuses them locally.
• Introduce a minimal communication layer (Communicator) with NoComm, RayonComm, and optional MPI backends for non-blocking send/recv and barriers.
• Ship a full section completion pipeline: neighbour_links → exchange_sizes(_symmetric) → exchange_data(_symmetric) with robust, symmetric handshakes that avoid deadlocks.
• Implement sieve completion (complete_sieve, complete_sieve_until_converged) to synchronize overlap arrows across ranks until convergence.
• Add stack completion helpers to mirror section completion for vertical stacks.

Performance & Determinism

• Replace boxed traversal closures with concrete iterators: closure_iter, star_iter, closure_both_iter (zero dyn-dispatch in hot loops).
• Add point-only adapters (cone_points, support_points) to eliminate payload clones during traversal.
• Provide deterministic traversal order via a height-major chart (index ↔ point) and bitset visited sets.
• Make InMemorySieve mutators degree-local and prealloc-aware: reserve_cone, reserve_support, and incremental mirror updates for set_cone/set_support (no global rebuilds).

Correctness Improvements

• Preserve the remote mesh point id in overlap completion wires; reconstruct Remote { rank, remote_point } exactly on receive.
• Tighten Overlap::add_link to respect closure-of-support and avoid duplicate links efficiently.
• Enforce minimality in meet/join (return only minimal shared faces / minimal cofaces).

API Additions

• closure_completed(...) with CompletionPolicy (Pre/OnDemand; Cone/Support/Both).
• Communicator trait + NoComm, RayonComm, and (feature-gated) MpiComm.
• Completion utilities: neighbour_links, size_exchange, data_exchange, sieve_completion, stack_completion.
• Deterministic helpers: chart_points(), chart_index() (height-major chart).

Optional/Advanced

• Hooks for DAG reachability indices (e.g., GRAIL / chain-labels) to speed redundancy checks and transitive reduction.
• Oriented closure plumbing for assembly into sections (sign-aware accumulation).

Breaking Changes

• None. All new capabilities are additive. Existing calls continue to work; new fast paths are opt-in.

Upgrade Notes

• Enable MPI backend with --features mpi-support.
• Use closure_local(...) for legacy behavior; switch to closure_completed(...) to traverse globally with overlap-driven completion.
• For best performance, prefer *_iter and cone_points/support_points in custom algorithms.

Testing & Docs

• Add parity tests for old vs. new traversals.
• Add symmetric comms tests for section/sieve/stack completion.
• Update docs with examples for closure_completed, communicator setup, and overlap construction.

License

MIT License. See LICENSE for details.