Deinsum: Practically I/O Optimal Multi-Linear Algebra for Distributed Systems

Explore a conference talk on Deinsum, a framework for distributed multilinear algebra computations using Einstein notation. Delve into the challenges of optimizing data movement in massively-parallel systems and learn how Deinsum addresses these issues. Discover the framework's approach to deriving data movement-optimal tiling and generating distributed schedules, enhancing local computation performance. Examine the application of Deinsum to important tensor kernel classes, including Matricized Tensor Times Khatri-Rao Products and Tensor Times Matrix chains. Gain insights into the framework's workflow, binary operations, communication minimization techniques, and optimal tile size determination. Analyze experimental results showcasing Deinsum's performance on the Piz Daint supercomputer, demonstrating significant speedups over existing solutions.

Introduction
Modern Scientific Applications
Deinsum Workflow
Binary Operations
Minimize Communication
Example
Cartesian Process Grid
Optimal Tile Sizes
Annotation
Data Distribution
Code Generation
Experimental Results
Conclusion