PASCO 2015 Invited Speakers
Overview of Recent Algorithms, Trends and Libraries in Numerical
Dense Linear Algebra
Department of Mathematical and Statistical Sciences
University of Colorado Denver
We review developments of Numerical Dense Linear Algebra by focusing on algorithms, trends, and libraries. We present various concepts: vectorizing, BLAS, blocking, tiling, out-of-order execution, pipelining, recursion, 2D block cyclic, data flow, and how these concepts have made their ways in Numerical Dense Linear Algebra algorithms and libraries. Along this journey, we will review algorithms which are present in the libraries LINPACK, LAPACK, ScaLAPACK, PLASMA, elemental and DPLASMA.
Interactive computations and outsourcing
INRIA Team-Project MOAIS
Laboratoire d'Informatique de Grenoble (CNRS, INRIA, INPG, UJF, UPMF)
Ensimag / Grenoble-INP - Grenoble University
An interactive computation involves input/output communication with the external world (e.g. a distributed global computing platform) during computation. Interaction between different computing models appears successful in both complexity theory (e.g. program checkers, interactive proofs, PCP theorem, ...) and practice (e.g. hybrid computing that exploits both multicore and GPUs, remote computing on HPC infrastructures, crowdsourcing, ...).
We focus in this talk on interactive algorithms where a (parallel) client outsources most of the computation to an external server, typically an HPC platform. Our first goal is designing an efficient algorithm for the client (including I/O), while guaranteeing an acceptable cost on the server. Illustrative examples are provided for linear algebra. A first inherent problem with outsourcing is to trust the validity of the result: either by checking, or asking the server to provide a computation certificate, or using algorithm-based fault tolerance (ABFT). ABFT may be used on global computing platforms such as P2P to tolerate erasures or errors from some peers. Yet such open platforms raise the issue of input secrecy. While perfect secrecy is unreachable for some problems (e.g. NP-hard), fully homomorphic encryption provides a generic but expensive solution against computationally bounded adversary. Alternative efficient algorithms for exact linear algebra are considered.
Exploiting Tesla GPUs with Magma
University of Sydney
I will present an overview of some algorithms which I have developed within Magma which exploit an NVIDIA Tesla GPU when available. This will include a discussion of how Magma can directly solve a random 16-bit prime instance of the Cortois (11,9,8)-Minrank Challenge C in about a day's computation using just one CPU core and one Tesla C2075 GPU.