No abstract

This paper shows a new hardware solution for implementing barrier synchronization in 2D-mesh multicomputers. The synchronization process is executed in n-I sequential stages where n is the largest dimension of the nxm mesh configuration. Each stage establishes simultaneous communication with maximum 4 nodes through additional wired-AND wires, allowing the propagation of information among the adjacent cells every time. This solution presents the following characteristics: a ) fast barrier synchronization, b) allows that any subset of nodes participate in each barrier whenever sequential barriers are considered, and e) allows simultaneous execution of multiple barriers whenever subsets of nodes belonging to non-overlapping rectangles are considered. The control logic and number of wired-AND wires are the same per node. Therefore, the system size is easily increased, being only limited by the expansion capacity of the 2D-mesh network. Analytical results indicate that the new approach should perform almost as well as efficient barrier synchronizers [I-31.

O SPLASH é um conjunto de programas aplicativos paralelos coletado pela Universidade de Stanford para uso em projeto e avaliação de sistemas multiprocessadores paralelos, distribuído livremente para fins acadêmicos. Utilizando simuladores baseados no MINT, será feita uma avaliação do comportamento de algumas aplicações do SPLASH. Três características serão estudadas, procurando estender os trabalhos anteriores da área e fornecer subsídios para a escolha de parâmetros arquiteturais do multiprocessador SPADE: i) o custo da sincronização; ii) a hierarquia de conjuntos de trabalho; e iii) o número de cópias invalidadas em cada operação de escrita em caches que seguem a política de invalidação em escrita.

Phylogenetics determines the evolutionary relationships between groups of species, through a phylogenetic tree. PhyML is among the main programs for the reconstruction of phylogenetic trees. Bootstrap is a statistical method used to measure the confidence of a given data set, which is usually applied in the analysis of inferred phylogenetic trees. In PhyML this method has two MPI parallel implementations: with point-to-point operations and collective operations. The second version is more efficient than the first, however it has a limitation on the number of bootstrap to be used due to the increase in memory consumption. In order to solve this problem, three proposals were developed. The objectives of this work were to carry out the validation of these versions together with performance tests. The validation showed that the proposed solutions present results equivalent to the point-to-point version. In the performance simulations, two solutions were shown to be superior to the point-to-point version, with the best one achieving gains of 28.46% and 39.64% for 32 and 64 processes, respectively. Therefore, the enhancements allow alternatives to the point-to-point version without limitingmemory.