Frédéric Krüger

EASEA: specification and execution of evolutionary algorithms on GPGPU

EASEA is a framework designed to help non-expert programmers to optimize their problems by evolutionary computation. It allows to generate code targeted for standard CPU architectures, GPGPU-equipped machines as well as distributed memory clusters. In this paper, EASEA is presented by its underlying algorithms and by some example problems. Achievable speedups are also shown onto different NVIDIA GPGPUs cards for different optimization algorithm families.

Speedups between ×70 and ×120 for a generic local search (memetic) algorithm on a single GPGPU chip

This paper presents the first implementation of a generic memetic algorithm on one of the two GPU (Graphic Processing Unit) chips of a GTX295 gaming card. Observed speedups range between ×70 and ×120, mainly depending on the population size. An automatic parallelization of a memetic algorithm is provided through an upgrade of the EASEA language, so that the EC community can benefit from the extraordinary power of these cards without needing to program them.

Automatic Parallelization of EC on GPGPUs and Clusters of GPGPU Machines with EASEA and EASEA-CLOUD

GPGPU cards are very difficult to program efficiently. This chapter explains how the EASEA and EASEA-CLOUD platforms can implement different evolution engines efficiently in a massively parallel way that can also serve as a starting point for more complex projects.

Blind signal separation applied to power load profiles computation

Electrical distribution companies struggle to find precise energy demand for their networks. They have at their disposal statistical tools such as power load profiles, which are however usually not precise enough and do not take into account factors such as the presence of electrical heating devices or the type of housing of the end users. In this paper, we show how the determination of accurate load profiles can be considered as a noisy blind source separation problem solved by an evolutionary algorithm. The power load profiles obtained demonstrate considerable improvement in the load curve forecasts of 20kV/400V substations.

Generic Local Search (Memetic) Algorithm on a Single GPGPU Chip

Memetic algorithms (MAs), evolutionary algorithms coupled with a local search routine, have been shown to be very efficient in solving a great variety of problems. This chapter presents the first implementation of a generic parallel MA on a general-purpose graphics processing unit card. An upgrade of the EASEA platform provides an automatic generation and parallelization of an MA for both novice and experienced users. Experiments on a benchmark function and a real-world problem reveal speedups ranging between × 70 and × 120, depending on population size and number of local search iterations.

Using Large-Scale Parallel Systems for Complex Crystallographic Problems in Materials Science

Very recently, the design and understanding of materials synthesis have received a huge effort in which modeling approaches are decisive. Here, we focus on the generation of crystalline inorganic frameworks. Despite the high-throughput (HT) methods that have proved to be useful for the discovery of zeolites, the determination of the new phase’s structure takes up a large part of the entire process. Therefore, we show how graphic processing units or GPUs can be used in order to speed up this mandatory step. We describe GPUs and predictive methods for phase determination. Then, we show all the details that allowed us to reach a stable and robust solution with benchmark analysis and real applications to zeolites.

Massively Parallel Generational GA on GPGPU Applied to Power Load Profiles Determination

Evolutionary algorithms are capable of solving a wide range of different optimization problems including real world ones. The latter, however, often require a considerable amount of computational power. Parallelization over powerful GPGPU cards is a way to tackle this problem, but this remains hard to do due to their specificities. Parallelizing the fitness function only yields good results if it dwarfs the rest of the evolutionary algorithm. Otherwise, parallelization overhead and Amdahl’s law may ruin this effort.

High-Performance Computing for Accelerated Zeolitic Materials Modeling

Very recently, the design and understanding of materials synthesis have received considerable attention where modeling approaches are decisive. Here, we focus on the generation of crystalline inorganic frameworks. Despite high-throughput (HT) methods having proved to be useful for the discovery of zeolites, the determination of the new phases’ structure takes up a large part of the entire process. Therefore, we show how graphic processing units (GPUs) can be used in order to speed up this mandatory step. We describe GPUs and predictive methods for phase determination. Then, we show all the details that allow us to reach a stable and robust solution with benchmark analysis and real applications for zeolites.