Distributed sparse computation

DISTAL: The Distributed Tensor Algebra Compiler

A compiler for distributed dense tensor algebra that lets users specify how data is distributed and how computation is distributed independently, then compiles to a task-based runtime.

C++ Legion

The problem

Running tensor computations across a cluster requires choosing both a data distribution and a computation distribution, and getting good performance usually means entangling those decisions with the kernel.

The idea

Express data distribution and computation distribution as separate, composable schedules over a machine model, and compile the combination to distributed tasks on the Legion runtime.

How it fits the group's work

DISTAL carries the TACO philosophy — separate concerns, then schedule them — into the distributed setting, adding machines as an explicit axis alongside data representation. SpDISTAL extends the approach to sparse tensors, and the ideas continue in the group’s work on task-based runtimes and Legate Sparse.

Key publications

All software