2026-08-13 –, Room 3
While Julia’s Dagger.jl provides a productive framework for task-based parallelism using Directed Acyclic Graphs (DAGs), its default reliance on TCP-based Distributed.jl limits performance on low-latency HPC interconnects. To bridge this gap, we developed MPIAcceleration, a strategic extension that replaces standard transport with an MPI-aware backend. By leveraging MPI.jl and non-blocking communication, we enable Dagger to use specialized hardware such as InfiniBand and Slingshot while maintaining a simple, high-level API.
Dagger's MPIAcceleration works seamlessly with the current scheduler, allowing task graphs to be executed across MPI ranks with minimal modifications. It only requires a single line of code: Dagger.accelerate!(:mpi).
When this feature is enabled, each MPI rank is integrated into Dagger's Processor/Memory Space model, which ensures that tasks are executed close to where their data resides. This rank-aware placement helps to minimize communication overhead. Additionally, remote data transfers happen transparently, providing handles on the appropriate ranks.
Yan Guimarães is a Software Engineering student at the University of Brasília (UnB) researching high-performance computing and distributed task scheduling in the Julia Language. He focuses on developing an MPI-based backend for Dagger.jl, a runtime system that uses DAG-based scheduling to handle distributed workloads automatically. He integrates MPI into Dagger's scheduler with one line of code, which enables MPI-aware task placement while hiding communication complexity. He evaluates performance on the Aurora exascale supercomputer and AWS, using parallel Cholesky decomposition benchmarks, showing competitive results on HPC interconnects. This research, developed through Google Summer of Code 2025.
Consultant at MIT's JuliaLab, Co-maintainer of Dagger. My interests span from more broad topics such as the accessibility and educational initiatives for parallel computing to Applied Physics and Numerical Linear Algebra.