+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Monday, 23. January 2023, 15:30 Uhr
Ort: Online (Zoom: https://umu.zoom.us/my/pauldj)
Referent:
Christos Psarras, M.Sc.
International Research Training Group (IRTG-2379)
Thema:
Beyond the rigid interfaces of super-optimized building-block libraries.
Our experiences in Chemometrics
Abstract:
The efficient computation of linear algebra expressions is a challenging
task faced by many practitioners in scientific fields, such as engineering,
image processing, and computational chemistry, to name a few. For most
applications, mapping a target expression into a sequence of
highly-optimized library routines (often referred to as "building-block"
libraries, e.g., BLAS, LAPACK), is an approach that offers good
computational performance as well as accuracy. However, in other
applications, this approach inherently results in a vast under-utilization
of the available computational resources, and thus reduced performance. In
this talk, we emphasize on these, latter, applications, showcasing two
occurrences that routinely arise in Chemometrics: the Canonical Polyadic
Decomposition (CP) and Jackknife resampling of CP models. For the first
occurrence, we describe the limitations of "mapping to building-blocks"
when computing multiple, low-rank CP decompositions. After close
collaboration with Chemometrics practitioners, we present a method (and
algorithm), CP-CALS, which leverages information about their workflow, to
overcome said limitations and achieve better performance. For the second
occurrence, we describe the unique challenge of Jackknife resampling. We
present a solution that addresses this challenge by making it possible to
use CP-CALS to significantly increase performance, at the cost of slightly
increasing the total amount of required computation. Through extensive
experimentation with synthetic and real datasets on single-threaded and
multi-threaded architectures, as well as on accelerators, we illustrate the
improved efficiency and performance of our methods.
Es laden ein: die Dozentinnen und Dozenten der Informatik