Date |
Jun 21, 2016 |
Speaker |
Dr. Yuji Shinano, Zuse Institute Berlin
(Visiting Associate Professor, Institute of Statistical Mathematics)
|
Title |
Towards Using over a Million CPU Cores to Solve Previously
Unsolved Mixed Integer Programming Problem Instances
|
Abstract |
The Ubiquity Generator (UG) is a framework for the external parallelization of
MIP solvers. UG provides a systematic way to develop a parallel solver that can
run on large-scale distributed memory computing environments. It was used to
develop ParaSCIP, a distributed memory, massively parallel version of the open
source academic solver SCIP. In this talk we present a success story where we
solve 14 open MIP instances from MIPLIB2003 and MIPLIB2010 using ParaSCIP on up
to 80,000 cores of supercomputers. Finally, we introduce ParaXpress, for which
one of the fastest commercial MIP solvers, the FICO Xpress-Optimizer, has been
parallelized by UG. Combining the internal shared-memory parallelization of
Xpress and the external parallelization of UG, we aim at a new order of
magnitude for supercomputer core-usage in MIP solving.
|
|