Franco Chichizola

Dynamic load balancing on non-homogeneous clusters

This paper discusses the dynamic and static balancing of non-homogenous cluster architectures, simultaneously analyzing the theoretical parallel speedup as well as the speedup experimentally obtained. A classical application (Parallel N-Queens) with a parallel solution algorithm, where processing predominates upon communication, has been chosen so as to go deep in the load balancing aspects (dynamic or static) without distortion of results caused by communication overhead. Four interconnected clusters have been used in which the machines within each cluster have homogeneous processors although different among clusters. Thus, the set can be seen as a N-processor heterogeneous cluster or as a multi-cluster scheme with 4 subsets of homogeneous processors. At the same time, three forms of load distribution in the processors (Direct Static, Predictive Static and Dynamic by Demand) have been studied, analyzing in each case parallel speedup and load unbalancing regarding problem size and the processors used.

Face recognition: reduced image Eigenfaces method

A new algorithm for automatic face recognition is presented: reduced image eigenfaces (RlE), based on the eigenface mode, improvements in the recognition percentage. The original eigenfaces method has been implemented in order to compare the results obtained by the new method under various conditions, such as quantity of people and quantity of photos of each of them. In the experimentation, we have used a limited-image data base which is internationally normalized. With RIE, two important advantages are achieved in relation to the previous model: improvement of the percentage of success in the recognition, and possibility of enhancing the set of images used for training

Mapping Tasks to Processors in Heterogeneous Multiprocessor Architectures: The MATEHa Algorithm

An automatic task-to-processor mapping algorithm is analyzed in parallel systems that run over loosely coupled distributed architectures. This research is based on the TTIGHa model that allows predicting parallel application performance running over heterogeneous architectures. In particular, the heterogeneity of both processors and communications is taken into consideration. From the results obtained with the TTIGHa model, the MATEHa algorithm for task-to-processors assignment is presented and its implementation is analyzed. Experimental results working on subsets of two-cluster heterogeneous machines are presented, analyzing the resulting mapping scheme with MATEHa and two previous mapping methods: MATE and HEFT. Finally, the algorithm robustness is considered based on the variation of model parameters: inter-process communication times and processing times.

Analysis of RAPL Energy Prediction Accuracy in a Matrix Multiplication Application on Shared Memory

In recent years, energy consumption has emerged as one of the biggest issues in the development of HPC applications. The traditional approach of parallel and distributed computing has changed its perspective from looking for greater computational efficiency to an approach that balances performance with energy consumption. As a consequence, different metrics and measurement mechanisms have been implemented to achieve this balance. The objective of this article focuses on monitoring and analyzing energy consumption for a given application through physical measurements and a software interface based on hardware counters. A comparison of the energy values gathered by Intel RAPL versus physical measurements obtained through the processor power source is presented. These measurements are applied during the execution of a classic matrix multiplication application. Our results show that, for the application being considered, the average power required by the processor has an error of up to 22% versus the values predicted by RAPL.

Experiences with Electronic Vote: Challenges and Solutions

This paper presents results of a research on electronic voting carried out by the Institute of Research in Computer Science LIDI of the School of Computer Science of the National University of La Plata, Argentina, as part of research in e-Government developed since 2003. In particular, the topic of electronic vote has generated various experiences that allowed developing different election models for public and private environments. These experiences have resulted in the transfer of technology to several sectors

PERFORMANCE COMPARISON OF PARALLEL PROGRAMMING PARADIGMS ON A MULTICORE CLUSTER

Currently, most supercomputers are multicore clusters. This type of architectures is said to be hybrid, because they combine distributed memory with shared memory. Traditional parallel programming paradigms (message passing and shared memory) cannot be naturally adapted to the hardware features offered by these architectures. A parallel paradigm that combines message passing with shared memory is expected to better exploit them. Therefore, in this paper the performance of two parallel programming paradigms (message passing and combination of message passing with shared memory) is analyzed for multicore clusters. The study case used is the construction of phylogenetic trees by means of the Neighbor-Joining method. Finally, conclusions and future research lines are presented.

Parallel processing puzzle N 2 -1 on cluster architectures performance analysis

An analysis of a parallel solution of N2-1 puzzle using clusters, is presented. This problem is interesting due to its complexity and related applications, particularly in the field of robotics. A variation of classic heuristics for forecasting the work to be done in order to reach a solution is analyzed, and it is shown that its use significantly improves the time of sequential algorithm A . Then, a parallel solution on a distributed architecture is presented and speedup is analyzed based on the number of processors, efficiency, and the possible superlinearity when scaling the problem.