RustMP

This repository contains the source code for Alex Bowman, Paul Ouellette, Raffi Sanna, and Jack Yu's research project for CSC200H: an implicit parallelism library written in Rust. For more details, please see the project report.

Authors

• Alex Bowman (abowman6@u.rochester.edu)
• Paul Ouellette (pouellet@u.rochester.edu)
• Raffi Sanna (rsanna@u.rochester.edu)
• Jack Yu (yyu57@u.rochester.edu)

Building/Running

RustMP requires the following dependencies to run:

• hwloc2-v2.2.0: C HWLoc wrapper used for NUMA aware process pinning
• lazy_static-v1.4.0: Lazy static macro used for delayed singleton init
• num-v0.3: Needed to run the Rust factorial benchmarks
• rand-v0.8.3: Random number generation used for benchmarking examples
• rayon-v1.5.0: Rayon parallel iterator library used in RustMP/Rayon.. comparision tests

src/ contains all code for the RustMP library, including lib.rs, sysinfo.rs, and threadpool.rs. src/bin contains benchmarking programs demonstrated in our paper. To run one of the benchmarks with cargo, execute the following command:

# Set the RustMP, OpenMP, and Rayon thread pool thread count to <n>
$ export RMP_NUM_THREADS=<n>
$ export OMP_NUM_THREADS=<n>
$ export RAYON_NUM_THREADS=<n>

# Run the test
$ cargo run --release --bin <testname>

C comparison benchmarks can be found in omp/. These can be compiled using make. All matrix multiplication tests support one integer input as for the square matrix dimensions (i.e. ./matmul <nsize>).

Additional benchmarks are located in benches. These require Rust nightly and can be run with the following command, note that RustMP thread pool lazy startup cost is factored into these tests:

$ cargo +nightly bench

Known issues

Due to Rust compiler limitations, Rust is only able to support up to 32 layers of nested macros by default. If an alternative test program that exceeds this nested depth for macros is used, please consider increasing this limit in Cargo.toml.