IAS Seminar "Communication avoiding Householder tridiagonalization"

Recently, communication avoiding (CA) is thought as one of the key technologies to overcome drawbacks on parallel computing. CA is based on data replication, exchange of the calculation order, improvement of locality, and so on. Typical examples of CA are the TSQR (Tall Skinny QR decomposition) algorithm and the Matrix Power Kernels (MPK) algorithm. They contribute reducing communication overhead for the orthonormalization operation and the Krylov subspace iteration method, respectively.

In this talk, the proposal of a novel type of Communication Avoiding algorithm for the Householder Tridiagonalization is addressed. The algorithm reduces the number of collective communication while the extra cost of calculation increases. We confirm the effect of the CA-based Householder tridiagonalization on massively parallel computers, especially K-computer and JUQUEEN systems. This talk is an early and preliminary benchmark report.

Anyone interested is cordially invited to participate in this seminar.