Friday,
November 18, 12:30 p.m. GWC 487
Yuji Shinano
Konrad Zuse Institute, Berlin, Germany
Parallelizing MIP Solvers using the UG Framework
Abstract
This talk will describe the Ubiquity Generator (UG), a parallel
framework for solving mixed-integer programming (MIP) problems. This
framework enables base MIP solvers to be run in parallel in an
external fashion, exploiting the high performance of the base MIP
solvers and directly benefitting from the frequent upgrades to the
individual base solvers. The base MIP solver could be SCIP, CPLEX,
etc., and the target computing environment could be a desktop
computer, a cluster computer or a supercomputer. We currently have
two parallel MIP solvers using the UG framework. One is ParaSCIP,
which is designed to run on a large scale distributed memory computing
environment. The other is FiberSCIP, a multi-threaded version of
SCIP. This talk will give a brief introduction to the UG framework
and the current state of parallel MIP solvers using it. Computational
results on a wide range of computing environments will be presented;
from a 4-core desktop computer to the 7,168 core supercomputer HLRN
II.
