Milena Lazarova

Comparison of parallel metaheuristics for solving the TSP

The paper compares the efficiency of several metaheuristics for solving the travelling salesman problem. Parallel computational models are suggested for ant colony optimization, simulated annealing and genetic algorithm. The models utilize multiple independent runs on multicomputer platform. The performance parameters and parallelism profiling of the metaheuristics are made based on hybrid (MPI+OpenMP) implementations of the suggested models for solving several instances of TSP from the TSPLIB.

Parallel simulated annealing for solving the room assignment problem on shared and distributed memory platforms

The paper is aimed at investigating the efficiency and quality of solutions of parallel simulated annealing on shared memory and distributed memory computer platforms for solving the room assignment problem. The parallel computational model for shared memory system utilizes concurrent generation and evaluation of moves by OpenMP based multithreading. Asynchronous and synchronous moves generation based on combination of functional and data parallelism are utilized for parallel computations using message passing on a distributed memory platform. Parallelism profiling and scalability analyses both in respect of the size of the parallel platform and the problem workload are presented.

Parallel models for sequence alignment on CPU and GPU

The paper presents parallel computational models of Smith-Waterman algorithm for CPU and GPU. An investigation is made of the performance parameters of computing similarity indexes between query sequences and a reference sequence using the suggested parallel programming models. Implementations for GPU based sequence alignment using nVIDIA CUDA and OpenCL as well as CPU based sequence alignment using OpenMP multithreaded implementation are presented. The experimental analyses are aimed at searching for similarities of the human gamma interferon protein and influenza virus.

Token-based adaptive load balancing for dynamically parallel computations on multicomputer platforms

The paper suggests an algorithm for token-based adaptive load balancing for dynamically parallel computations on multicomputer platforms. The proposed algorithm for load balance is initiated and performed by the idle or under-loaded processes and requires token message circulating among the parallel processes and bearing information about the load distribution throughout the system. The efficiency of the algorithm is estimated for the case study of Sam Loyds puzzle utilizing parallel version of branch-and-bound search algorithm with depth-first search strategy. The experimental study is based on flat parallel program implementations. Speedup and efficiency of the parallel system are estimated as well as scalability of the application workload and the multicomputer size.

PARMETAOPT — Parallel metaheuristics framework for combinatorial optimization problems

The paper presents an experimental parallel metaheuristics framework for solving combinatorial optimization of grand challenge scientific and engineering problems that has been developed based on biologically inspired metaheuristics, modeling of social behavior and cultural evolution as well as trajectory-based methods. A prototype class library for metaheuristics is developed and several parallel computational models of metaheuristics for solving combinatorial optimization problems are implemented. The library contains implementations in C++ of parallel computational models for both population based and trajectory based metaheuristics. Some improvements in the parallel models are suggested and implemented in the library PARMETAOPT. The influence of the parameters on the performance of some of the parallel algorithms is analyzed using the developed parallel metaheuristics framework and performance tuning rules are suggested. The implementations are based on message passing with MPICH2 for the flat programming models and OpenMP API is used for multithreading in the hybrid programming models.

Real Time GPU Accelerated Radar Scan Conversion and Visualization

The paper presents an approach for real time acquisition and scan conversion of high resolution radar data on a modern computer system using GPU based acceleration. The suggested data processing for radar scan conversion and visualization is based on an OpenCL and OpenGL data sharing model, on asynchronous data acquisition and on texture based rendering. The experimental results show a stable number of frames per second being obtained using the proposed approach for GPU accelerated data acquisition and processing. This provides relevant performance and sufficient frame rate for real time radar data visualization.

Parallel software framework for high-performance multispectral analysis for earth monitoring

The paper presents a software package BlueVision for parallel processing of multispectral data from remote sensing of the Earth. The purpose is to provide a software framework for high-performance analysis aimed at detection and monitoring of different natural hazards. The package consists of five modules that implement parallel computational modules for fire detection, deforestation, soil salinity, water pollution, flooding. Additional module visualizes the detected areas on a geographic map. Performance analysis are based on the experimental results on a heterogeneous multicomputer cluster and the Bulgarian supercomputer BlueGene/P.