Simplifying the Parallelization of Highly Concurrent Workloads
Ruediger Berlich, Sven Gabriel, Ariel Oscar Garcia · 2016
This paper discusses strategies for decreasing execution times of highly concurrent workloads in parametric optimization, on the example of version 1.6.1 of Geneva, a collection of optimization algorithms that focusses on problem domains with particularly long running evaluation functions.Particular emphasis lies on the Courtier broker architecture used for parallelization in environments ranging from many-core systems and GPGPU to Grids and Clouds.The paper also introduces other means that have proven to be useful for the reduction of overall compute time, such as acyclic creation of random numbers or avoidance of regions of parameter space that are known to lead to invalid results.