Title: Algorithms as Multilinear Tensor Equations
Presented by Edgar Solomonik, ETH Zurich
Abstract: Tensors provide a powerful abstraction for expressing algorithms on sparse or dense datasets in their natural dimensionality. Graph algorithms such as betweenness centrality, as well as recursive algorithms such as FFT and bitonic sort, can be succinctly written as tensor operations over a suitable algebraic structure. I will introduce communication and synchronization cost lower bounds for a general class of tensor algorithms, including sparse iterative methods and matrix factorizations. Then, I will present parallel algorithms that achieve minimal cost with respect to these bounds and obtain improved scalability on supercomputers. Additionally, I will describe new innovations in handling symmetry and sparsity in tensors. Some of the proposed algorithms are deployed in a massively-parallel tensor framework, whose development has been driven by applications in quantum chemistry. I will show the performance of the framework for algorithmic benchmarks as well as for coupled cluster methods, which model electronic correlation by solving tensor equations.
Bio: Edgar Solomonik is a postdoctoral fellow at ETH Zurich working in the field of parallel numerical algorithms. His research introduced more communication-efficient algorithms for numerical linear algebra and his software for tensor computations has been widely adopted in the field of electronic structure calculations. He obtained his BS from the University of Illinois, Urbana-Champaign and his PhD from the University of California, Berkeley, both in Computer Science. He was the recipient of the DOE Computational Science Graduate Fellowship, the David J. Sakrison Memorial Prize, and the ACM-IEEE George Michael HPC Fellowship.
![](/sites/default/files/styles/thumbnail/public/default_images/6KbA9AJs_400x400_0.png?itok=JZ3_TJqR)