Luz Marina Moreno

MALLBA: a library of skeletons for combinatorial optimisation

The MALLBA project tackles the resolution of combinatorial optimization problems using algorithmic skeletons implemented in C++. mallba offers three families of generic resolution methods: exact, heuristic and hybrid. Moreover, for each resolution method, MALLBA provides three different implementations: sequential, parallel for local area networks, and parallel for wide area networks (currently under development). This paper explains the architecture of the MALLBA library, presents some of its skeletons, and offers several computational results to show the viability of the approach.

The master-slave paradigm on heterogeneous systems: a dynamic programming approach for the optimal mapping

We study the master-slave paradigm over heterogeneous systems. According to an analytical model, we develop a dynamic programming algorithm that allows to solve the optimal mapping for such paradigm. Our proposal considers heterogeneity due both to computation and also to communication. The optimization strategy used allows us to obtain the set of processors for an optimal computation. The computational results show that considering heterogeneity also on the communication increases the performance of the parallel algorithm.

Towards the automatic optimal mapping of pipeline algorithms

The assignment of computations to processors is a crucial factor determining the effectiveness of a parallel algorithm. The portability of parallel programs has involved lot of effort during the last decade. However, the performance of a parallel code suffers, in many cases, from inherent effects of the target architectures. The optimal mapping of a parallel program is strongly dependent on the granularity and network architecture. We focus on the problem of finding the optimal mapping of pipeline algorithms on a ring of processors. We propose an analytical model that allows an easy estimation of the parameters needed to obtain the mapping. The model can be introduced in a suitable tool to automatically produce this mapping. Both the accuracy of the model and the optimal efficiency of the algorithm found are contrasted on pipeline algorithms for the knapsack problem, for the resource allocation problem and for the path planning problem.

On the prediction of master-slave algorithms over heterogeneous clusters

We study the performance of master-slave algorithms on heterogeneous networks. The word heterogeneity refers here to both the processing and communication capabilities. Following an inductive approach, we derive a formula that predicts the performance for the general case. The computational results carried out on a heterogeneous cluster of PCs prove the effectiveness of the approach and the accuracy of the predictions. The numerical minimization of this function provides an efficient approach for an optimal distribution of the work.

Pipelines on heterogeneous systems: models and tools

We study the performance of pipeline algorithms in heterogeneous networks. The concept of heterogeneity is not only restricted to the differences in computational power of the nodes, but also refers to the network capabilities. We develop a skeleton tool that allows us an efficient block‐cyclic mapping of pipelines on heterogeneous systems. The tool supports pipelines with a number of stages much larger than the number of physical processors available. We derive an analytical formula that allows us to predict the performance of pipelines in heterogeneous systems. According to the analytical complexity formula, numerical strategies to solve the optimal mapping problem are proposed. The computational results prove the accuracy of the predictions and effectiveness of the approach. Copyright

An Analytical Model for Pipeline Algorithms on Heterogeneous Clusters

The performance of a large class of virtual pipeline algorithms on heterogeneous networks is studied. The word heterogeneity refers here both to the processing and communication capabilities. To balance the differences among processors requires a vectorial combination of block and cyclic mappings, assigning different number of virtual processes per processor. Following a progressive approach, we derive a formula that predicts the performance for the general case. The computational results carried out on a heterogeneous cluster of PCs prove the effectiveness of the approach and the accuracy of the predictions.

The Tuning Problem on Pipelines

Performance analysis and prediction is an important factor determining the efficiency of parallel programs. Considerable efforts have been made both in pure theoretical analysis and in practical automatic profiling. Unfortunately, contributions in one area seem to ignore the results of the other. We introduce a general performance prediction methodology based on the integration of analytical models and profiling tools. According to this approach we have developed a tool that automatically solves the prediction of the parameters for optimal executions of parallel pipeline algorithms. The accuracy of the proposal has been tested on a CRAY T3E for pipeline algorithms solving combinatorial optimization problems. The results obtained suggest that the technique could be successfully ported to other paradigms.

Optimal Mapping of Pipeline Algorithms

The optimal assignment of computations to processors is a crucial factor determining the effectiveness of a parallel algorithm. We analyze the problem of finding the optimal mapping of a pipeline algorithm on a ring of processors. There are too many variables to consider, the number of virtual processes to be simulated by a physical processor and the size of the packets to be communicated. We provide an analytical model for an optimal approach to these elements. The low errors observed and the simplicity of our proposal makes this mechanism suitable for its introduction in a parallel tool that compute the parameters automatically at running time.

On the parallel prediction of the RNA secondary structure

Publisher Summary This chapter discusses the parallelization of the Vienna package algorithm that predicts the secondary structure of the RNA. Before the tool can be effectively used, it must be tuned on the target architecture. This tune consists of finding the tile sizes for optimal executions. The analytical model is developed to find these optimal tile sizes. The validity of the analytical model developed for the parallel algorithms presented is studied through an extensive set of experiments performed on a Origin 3000. A statistical model is presented, to deal with the cases where the hypothesis of the analytical model is not satisfied. The algorithms are satisfactorily applied to real sequences. The algorithm is applied to real sequences. The speedups, although satisfactory, do not scale for some of those instances— that is, p 3 8. This is probably due to the use of a common static tile size. A dynamic variable tile size on every macro-rectangle is introduced in the chapter.

Adaptive Execution of Pipelines

Given an algorithm and architecture a tuning parameter is an input parameter that has consequences in the performance but not in the output. The list of tuning parameters in parallel computing is extensive: some depending on the architecture, as the number of processors and the size of the buffers used during data exchange and some depending on the application. We formalize the General Tuning Problem and propose a generic methodology to solve it. The technique is applied to the special case of pipeline algorithms. A tool that automatically solves the prediction of the tuning parameters is presented. The accuracy is tested on a CRAY T3E. The results obtained suggest that the technique could be successfully ported to other paradigms.