Catálogo de publicaciones - libros

Performance measurements were extensively conducted to characterize parallel computer systems by using modelling and experiments. By analyzing them, we corroborate current models did not provide precise memory characterization. After detailed result observation, we conclude that the performance slowdown is linear when using the main memory, and exponential when using the virtual memory.

In this paper, we propose a characterization model composed of two regressions which represent the slowdown caused by memory usage. Experimental results confirm the memory slowdown model improves the quality of computing system characterization, allowing to carry out simulations and the use of such results as a way to design real systems, minimizing project design costs.

In VLIW processor design, clustered architecture becomes a popular solution for better hardware efficiency. But the inter-cluster communication (ICC) will cause the execution cycles overhead. In this paper, we propose a shared cluster register file (SCRF) architecture and a SCRF register allocation algorithm to reduce the ICC overhead. The SCRF architecture is a hybrid register file (RF) organization composed of shared RF (SRF) and clustered RFs (CRFs). By putting the frequently used variables that need ICCs on SRF, we can reduce the number of data communication of clusters and thus reduce the ICC overhead. The SCRF register allocation algorithm exploits this architecture feature to perform optimization on ICC reduction and spill codes balancing. The SCRF register allocation algorithm is a heuristic based on graph coloring. To evaluate the performance of the proposed architecture and the SCRF register allocation algorithm, the frequently used two-cluster architecture with and without the SRF scheme are simulated on Trimaran. The simulation results show that the performance of the SCRF architecture is better than that of the clustered RF architecture for all test programs in all measured metrics.

The paper considers the model of a parallel program, which can be effectively interpreted using an instrumental computer, allowing for fairly exact prediction of the actual runtime of a parallel program on a specific parallel computing system. The model has been developed for parallel Java-programs with explicit exchange of messages by means of the MPI library. The model is part of the ParJava environment. The model is derived by converting the program control tree, which, for the Java-program, can be built by modifying the abstract syntax tree. Communication functions are modeled by using the LogGP model, allowing to account for the specific features of a distributed computational system.

The architecture of a runtime system supporting parallel execution of fragmented library subroutines on multicomputers is proposed. The approach makes possible to develop the library of parallel subroutines and to provide automatically their dynamic properties such as dynamic load balancing. Usage of the MPI for communications programming provides good portability of an application.

The paper presents a simulation-based approach to creation of decision support system for IT Service Management. The presented approach includes monitoring of stochastic data IT Services and calculation of measure of alignment to business goals with its characteristic basing on SLA/SLO. The approach combines the benefits of two kinds of models: analytical and simulation ones. The key idea of the paper is to demonstrate how modern methods of stochastic process analysis may enhance trustworthiness and quality of decision making along business goals within IT Services.

An effective workload distribution has a prime rule on reducing the total execution time of a parallel application on heterogeneous environments, such as computational grids and heterogeneous clusters. Several methods have been proposed in the literature by many researchers in the last decade. This paper presents two approaches to workload distribution based on analytical models developed to performance prediction of parallel applications, named PEMPIs VRP (). The workload is distributed based on relative performance ratios, obtained by these models. In this work, we present two schemes, static and dynamic, in a research middleware for a heterogeneous network of computers. In the experimental tests we evaluated and compared them using two MPI applications. The results show that, using the VRP’s dynamic strategy, we can reduce the imbalance, among the execution time of the processes, in relation to average time from 25% to near of 5%.

A flash memory has write-once and bulk-erase properties so that an intelligent allocation algorithm is essential to providing applications efficient storage service. This paper first demonstrates that the online version of FLASH allocation problem is difficult, since we can find an adversary that makes every online algorithm to use as many number of blocks as a naive and inefficient algorithm. As a result we propose an offline allocation algorithm called (BestM) for allocating blocks in FLASH file systems. The experimental results indicate that BestM delivers better performance than a previously proposed First Re-arrival First Serve (FRFS) method.

The paper describes the OpenTS parallel programming system that provides the runtime environment for T++ language. T++ is an extension for C++ that adds a set of keywords to C++, allowing smooth transition from sequential to parallel applications. In this context the support of repeated assignments to a variable is an important feature. The paper focused on semantics and implementation of such variables in T++. Applications written in T++ can be run on computational clusters, SMPs and GRIDs, either in Linux or Windows OS.

A new highly parallelizable method of moving mesh construction based on the Kohonen’s Self-Organizing Maps (SOM) is proposed. This method belongs to a class of methods in which the mesh is an image under an appropriate mapping of a fixed mesh over a computational domain. Unlike the conventional methods of this class, the proposed method doesn’t require solving complicated systems of nonlinear partial differential equations and is able to work with arbitrary time-dependent mesh density function. High efficiency of parallelization is conditioned by the inherent parallelism of the underlying stochastic SOM algorithm. Sequential as well as detailed parallel algorithms for moving mesh construction are proposed.

Scientific applications are increasingly challenging computational platforms and software tools. In this scenario of improving performance demand, computer cluster users require for mechanisms that could reduce data transfer delay. To contribute to this question, we proposed a data transfer in advance mechanism that improved overall system performance by diminishing data-intensive job wait. We also designed and implemented the Integrated Scheduling System (ISS) to analyze and evaluate our proposal. This system automates the preparation, submission and tracking of job executions in a cluster. The mechanism is also combined with an I/O file operation and computation overlapping idea that results in significant improvement of performance rates, confirmed by some experiments.