CS 267 Final Report Reproducible Parallel Matrix-Vector Multiply

Parallel code can be difficult to verify due to inherently non-reproducible execution models. When debugging or writing tests, users could benefit from getting the same result on different runs of the simulation. This is the goal that the ReproBLAS project (Nguyen et al.) intends to achieve. ReproBLAS [3] has so far introduced a reproducible floating point type (indexed float) and associated algorithms underlying BLAS1 serial and parallel routines. A long-term goal of ReproBLAS is, for example, the implementation of a fullyfeatured reproducible PBLAS [6]. ReproBLAS has yet to define a formal interface for its underlying algorithms, and it is as yet unknown whether the existing low-level routines can be assembled together to form reasonably efficient higher-level routines. Here, we implement a reproducible matrix-vector multiply in order to gauge the feasibility of and identify challenges in building a more complex reproducible linear algebra library.