Catálogo de publicaciones - libros

Wireless technology is increasingly finding its way into industrial communication because of the tremendous advantages it is capable of offering. However, the high bit error rate characteristics of wireless channel due to conditions like attenuation, noise, channel fading and interference seriously impact the timeliness and guarantee that need to be provided for real-time traffic. Existing wired protocols including the popular PROFIBUS perform unfavorably when extended or adapted to the wireless context. Other wireless protocols proposed either do not adapt well to erroneous channel conditions or do not provide real-time guarantees. In this paper, we present a novel real-time MAC (Medium Access Control) protocol that is specifically tailored to the message characteristics and requirements of the industrial environments. The protocol exploits both the spatial and temporal diversity of the wireless channel to effectively schedule real-time messages in the presence of bursty channel error conditions. Simulation results show that the proposed protocol achieves much better loss rate compared to baseline protocols under bursty channel conditions.

Users of parallel machines need to have a good grasp for how different communication patterns and styles affect the performance of message-passing applications. MPI Collective communications involve multiple processors, and their performance prediction is a tricky task to perform. In order to evaluate the performance of collective communications, we attempt to extend LogGP and P-LogP standard point-to-point models. Our objective is to compare these models with the empirical data, and identify the most suitable for performance characterization of collective communications. The models proposed are related with the implemented algorithms in MPICH. The experimental results performed on clusters of 16 and 64 processors connected by Infiniband and Gigabit Ethernet networks respectively, show the same trend. For any collective operation, given a number of processors and a range of message sizes, there is at least one model that predicts the performance precisely.

The classification phase is computationally intensive and frequently recurs in tracking applications in sensor networks. Most related work uses traditional signal processing classifiers, such as Maximum A Posterior (MAP) classifier. Naïve formulations of MAP are not feasible for resource constraint sensornet nodes. In this paper, we study computationally efficient methods for classification. We propose to use one-sided Jacobi iterations for eigen value decomposition of the covariance matrices, the inverse of which are needed in MAP classifier. We show that this technique greatly simplifies the execution of MAP classifier and makes it a feasible and efficient choice for sensornet applications.

Shared Memory Multiprocessor (SMP) systems based on processors with Chip-level MultiThreading (CMT) technology are becoming mainstream servers in High Performance Computing (HPC) applications and commercial business applications as well. With multiple threads executing on a processor chip at the same time, CMT servers promise to deliver higher aggregate performance than servers without CMT technology. However, resource sharing among the threads executing on the same processor chip can cause conflicts and hurt the performance. Thus in order to obtain high performance and scalability on CMT servers, it is crucial to understand the performance impact that the CMT processors have on the target applications. In this paper, we evaluate the performance of an example high-end CMT server, Sun Fire E25K, using HPC applications parallelized with OpenMP standard, SPEC OMPL (standard OpenMP benchmark suite). We also study the performance impact of the resource conflicts on the CMT processor for each benchmark program.

Locality behavior study is crucial for achieving good performance for irregular problems. Graph algorithms with large, sparse inputs, for example, oftentimes achieve only a tiny fraction of the potential peak performance on current architectures. Compared with most numerical algorithms graph algorithms lay higher pressure on the memory system. In this paper, using the minimum spanning tree problem as an example, we study the locality behavior of graph algorithms, both sequential and parallel, for arbitrary, sparse instances. We show that the inherent locality of graph algorithms may not be favored by the current architecture, and parallel graph algorithms tend to have significantly poorer locality behaviors than their sequential counterparts. As memory hierarchy gets deeper and processors start to contain multi-cores, our study suggests that architectural support and new parallel algorithm designs are necessary for achieving good performance for irregular graph problems.

We propose a precomputation-based scheme offering Pareto optimal solutions to the network optimization problem. This scheme precomputes bandwidth allocation (rate vector) and end-to-end paths with QoS guarantees. It prepares a routing database, identifying an optimal path upon each connection request. We propose a mixed-integer optimization model with nonlinear utility functions. The purpose of this work is to choose the optimal solutions and to provide decision makers a set of solutions satisfying users’ preferences with fairness. It prepares for a proportionally fair treatment of all competing connections. Numerical results show that the proposed model can provide each connection with its fair share of the bandwidth which is proportional to the target rate.

Flash Dissemination is a particularly useful form of data broadcast that arises in many mission-critical applications. The goal is rapid distribution of medium amounts of data in as short a time period as possible. While optimal algorithms are available for a highly constrained case (all nodes having the same bandwidth and latency), there is relatively little work in the context of heterogenous networks. Most systems and protocols today either use trees or randomized mesh-based techniques to deal with heterogeneity and work with local knowledge. We argue that a protocol with global knowledge can perform much better. In this paper, we propose two centralized heuristics – DIM-Rank and DIM-Time that use global knowledge to schedule data transfer between nodes. The heuristics are based upon insights from broadcast theory. We perform experimental evaluation of these two heuristics with decentralized randomized approaches and show that DIM-Rank achieves faster dissemination than decentralized approaches across a range of heterogeneity metrics.

With the recent proliferation of video-on-demand services, caching in a multimedia streaming server is becoming increasingly important. Previous studies have shown that request interval based caching and its extension for considering different video popularity performs well for various streaming environments. In this paper, we show that block level refinement of this existing scheme can further improve the performance of streaming servers. Trace driven simulations with real world VOD traces have shown that the proposed scheme improves the cache hit rate and the startup latency.

In a grid, data is stored in geographically-dispersed virtual organizations with varying administrative policies and structures. Current grid middleware provide basic data-management services including data access, transfer and simple replica management. Grid applications often require much more sophisticated and flexible mechanisms for manipulating data than these, including logical hierarchical namespace, automatic replica management and automatic latency management. We propose a view-oriented framework that builds on top of existing middleware and provides global and application-specific logical hierarchical views. Specifically, we developed mechanisms to create, maintain, and update these views. The views are synchronized using an efficient group communication protocol. Gvu (pronounced G-view) is built as a distributed set of synchronized servers and scales much better than the existing grid services. We conducted experiments to measure various aspects of Gvu and report on the results, showing Gvu to outperform existing grid services, thanks to its distributed nature.