COMPUTER ARCHITECTURE (3)

Author: Andrade Canosa Diego.
Year: 2006.
University: A CORUÑA [www.udc.es].
Place of defense: Facultad de Informática.
Place of preparation: Facultad de Informática.

Summary: The performance of memory hierarchies, in which the cache plays a key role, is critical in the current general purpose computers and embedded systems, due to the growing problem of bottleneck in the system memory. Unfortunately, the behavior of the cache is very volatile and difficult to predict. This is especially true in the presence of irregular patterns of access, which exhibited little town. Such patterns are very common in such applications in which some references are affected by suspended sentences or in the compressed storage arrays scattered leads to the emergence of indirecciones. SIn hand, the cache behavior in the presence of irregular patterns of access has not been studied extensively. In this thesis we present extensions of a systematic analytical modeling technique based SMEs (Equations probability of failures) that allow automatic behavioral analysis for cache tags with suspended sentences whose true value can not be determined at compile time and codes references irregular due to indirecciones respectively. The model generates very precise predictions despite the irregularity and has a low computational cost is the first model that meets these two characteristics can automatically analyze such codes. These properties make the model suited to serve as a guide compiler optimizations. The extension of the model codes for irregular indirecciones has been integrated into the compiler XARK, a compiler - oriented automatic recognition of kernels on scientific applications. We show how to make use of the powerful capabilities of extracting information about this compiler to allow automatic modeling codes based on scientific loops.

Author: Pardines Lence Inmaculada.
Year: 2006.
University: SANTIAGO DE COMPOSTELA [www.usc.es].
Place of defense: Escola Técnica Superior de Enxeñería.
Place of preparation: Facultad de Física.

Summary: In this thesis is a study of strategies for paralelizacion best suited to the efficient execution of the routines commonly used methods of non-linear optimization, and in particular algorithms cuasi-Newton on distributed memory systems. To evaluate the efficiency of the proposed solutions has been used code optimization MINOS as a testing ground.

Author: FERRANDEZ PASTOR FRANCISCO JAVIER.
Year: 2006.
University: ALICANTE [www.ua.es].
Place of defense: ESCUELA POLITÉCNICA SUPERIOR UNIVERSIDAD DE ALICANTE.
Place of preparation: ESCUELA POLITÉCNICA.

Summary: The thesis proposes a method for differential treatment in receiving electromagnetic waves and applications arising therefrom. The phenomena of interaction in the course of transmission of electromagnetic waves are used in different disciplines which features state of the environment or crossing. Among them, from the observation of these phenomena, it announced a working hypothesis that seeks to take advantage of an act contrasted: different frequencies induce differences in the interaction for the same medium. It has developed a simulation tool that solves the equations with electromagnetic numerical methods for quantifying the interaction between the wave and the environment. Its justification is that is a prelude to construction of any specialized hardware solution. The simulator also has been designed in line with the approaches of the current methodological called Science, Technology and Society (CTS), which acknowledges the influence of various social aspects in the development of science and as a consequence, also appreciates the transmission and access to conocimeintos that can be understood and used universally. In this sense, the simulation tool that incorporates utilities are accessible both the understanding of physical phenomena which occurred interaction or modification of algorithms that solve the problem of interaction between wave and physical environment. Systems posiconamiento satellite, geology, geodesy and meteorology are examples of the different scenarios possible working. Generally working with frequency that maximize the interaction onda-medio for, thus able to measure at the reception effects. Thus, from the range of the visible spectrum microwave until the election depends on factors such as the characteristics of the medium (permitivity, conductivity, dielectric constant), the pervasiveness of the signal or even the geometry or the distance between sender and receiver. For each application using a model that represents the phenomena and mechanisms of interaction which applies the technique of measuring differential multifrecuencial through the corresponding core simulation adaptado.El proposed method uses measuring technique differential multifrecuencial into the receiver with contrasting systems current reception on a frequency of work. The measure represents a new proposal differential treatment of the signal which introduces advantages in synchronization devices and as a consequence, simplifies installation and manteniento of reference systems. The goal is to develop systems more accessible, both from a technical standpoint and the economy. As an example of use are discussed scenarios in which the speed of propagation of the signal is a function of how often the same. Its use in the characterization of specific means for applications in remote sensing and measuring distances to positioning systems.

Author: Amancio Duarte Angelo.
Year: 2006.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: Escola Tecnica Superior d'Enginyeria.
Place of preparation: Escola Tecnica Superior d'Enginyeria.

Summary: The fault tolerance has become an important requirement for computer engineers and software developers, because the occurrence of failures increases the operating cost of a parallel computer. On the other hand, the activities of the fault tolerance mechanism reduces the benefits of the system from the user's perspective. This thesis presents a fault-tolerant architecture for parallel computers, called RADIC (Redundant Array of Distributed Fault Tolerance Controllers,), which is both transparent, decentralized, flexible and scalable. RADIC is a fault-tolerant architecture that is based controller distributed to handle bugs. The controller is based on processes involved, which share the resources of the computer user in parallel. To validate the operation of the architecture RADIC, we performed an implementation that follows the standard MPI-1 and contains elements of the architecture. The implementation, called RADICMPI allows verify the functionality of RADIC in situations without judgment or under fault conditions. The tests were conducted using an injector failures, involved in the code RADICMPI, enabling all the conditions necessary to validate the operation of the controller distributed RADIC. It also used the same implementation to study the implications of using RADIC in a real environment. This allowed to evaluate the operation of the architecture in practical situations, and to study the influence of parameters RADIC on system performance. The results proved that the architecture of RADIC working properly and that is flexible, scalable, transparent and decentralized. In addition, RADIC established an architecture for fault tolerance systems based on message passing.

Author: FERNANDEZ ESCRIBANO GERARDO.
Year: 2006.
University: CASTILLA-LA MANCHA [www.uclm.es].
Place of defense: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.
Place of preparation: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.

Summary: The underlying problem in the doctoral thesis was resolved through three proposals math: 1. The employment ratio of continuous (DC) of the Transformed Discreta Cosine (DCT). The DC coefficient blocks 8x8 of the DCT (frequency domain) represents the sum of all values in the array (domain space), divided by 8. Therefore, if you calculate the difference in absolute value of the DC coefficients of two matrices 8x8, we get a quick estimate of the difference in the mean values of the parent. This principle, acquired a large representation in the domain of frequencies, as most of the standards for encoding images often employ the DCT, and more specifically the DC coefficient values close to it for storing images. Moreover, the calculation of the DC coefficient matrix 16x16 is a calculated quickly through the use of DC coefficients of four arrays 8x8. 2. The use of statistics (mean and variance) for the location in images of areas with a lot of information and detail (contours), and areas with little information (funds). Once this classification, the information is processed using statistical techniques, to create decision trees based on rules (Data Mining and Machine Learning). 3. The use of the theorem of Pythagoras in defining spaces spherical search for the optimization of motion vectors in video encoding standards. Means motion vector to vector indicating the direction and the direction of movement between two consecutive frames in a video sequence.

Author: MARTINEZ MORAIS RAUL.
Year: 2006.
University: CASTILLA-LA MANCHA [www.uclm.es].
Place of defense: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.
Place of preparation: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.

Summary: Advanced Switching (AS) is a network technology based on PCI Express. PCI Express is the next generation PCI, which is already used extensively by replacing the PCI bus. AS is an extrapolation of PCI Express that takes lent its two layers of architectural lowest level, and includes a transaction layer optimized to enable new capabilities such as peer-to-peer communication. While PCI Express has already begun to reform a new generation of personal computers and servers traditional network interconnection common with the communications industry seems logical and necessary. Thus, AS was designed to proliferate in multiprocessing environments, peer-to-peer systems in communications, storage, network interconnection, servers and embedded platforms. Moreover, the quality of service (Quality of Service, QoS) is becoming an important feature for high performance networks. Providing QoS in computing and communications is currently the focus of many research efforts by industry and academia. AS incorporates mechanisms that can be used to provide QoS. Specifically, AS allows use Virtual Channels (Virtual Channels, VCs), arbitration on the output ports and a control mechanism for admission. In addition, AS provides a flow control at the level of liaison and VC. These mechanisms allow us to add traffic with similar features in a single VC and VC provide each treated differently based on their requirements. The main aim of the thesis was to study the various mechanisms AS with a view to proposing a general framework to provide QoS to applications on the network technology. Here, the main focus of work, given its importance in providing QoS, has been studying the mechanisms of planning AS. Our goal has been implemented in an efficient manner, taking into account both its benefits and its complexity. To achieve these goals, we have proposed several possible implementations planner minimum bandwidth of AS. We proposed to amend the planner based AS table with the aim of solving the problems of this requirement to provide QoS with variable packet sizes. We have also proposed setting scheduler based on table for decoupling resulting allocations for bandwidth and latency. In addition, we have carried out a design hardware in different planners to obtain estimates on the time of arbitration and the silicon area required. In addition, we have developed our own simulator to evaluate the benefits of our proposals.

Author: MARTINEZ VICENTE ALEJANDRO.
Year: 2006.
University: CASTILLA-LA MANCHA [www.uclm.es].
Place of defense: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.
Place of preparation: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.

Summary: The interconnection networks are a key component in a large number of systems. The mechanisms for quality of service (QoS) are responsible for ensuring that reaches a certain performance in the network. The traditional solutions to provide QoS in networks to interconnect high performance usually are based on complex architectures. The main objective of this thesis is to investigate whether we can provide efficient mechanisms for QoS. Our aim is to achieve full support QoS with minimal resources. To do so, identify overlaps in the proposed mechanisms for QoS and are eliminated without affecting performance. This thesis consists of three parts. The first began with the proposals traditional QoS level class of traffic. In the second part, we propose how to adapt the mechanisms for QoS-based interconnection networks deadlines for high performance. Finally, we investigate the interaction of QoS mechanisms to control congestion.

Author: VILLALON MILLAN JOSE MIGUEL.
Year: 2006.
University: CASTILLA-LA MANCHA [www.uclm.es].
Place of defense: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.
Place of preparation: ESCUELA POLITECNICA SUPERIOR DE ALBACETE.

Summary: This Doctoral Thesis has as its main aim to improve the performance of multimedia communications over wireless networks IEEE 802.11. Therefore, we study the major problems that arise with regard to support QoS and multicast traffic transmission to tackle the problem of multimedia communications over such networks. The main problems are that we find that the mechanism EDCA proposed IEEE 802.11e to try to provide QoS not get adequate benefits in some scenarios hybrids, which are both stations using the IEEE Standard 802.11 (DCF) to access the channel stations as IEEE 802.11e (EDCA), and that the multicast support such networks is inadequate for the transmission of video. Based on these problems encountered in this Doctoral Thesis aims on the one hand the mechanism B-EDCA to improve the performance of QoS achieved by the way of access to the canal EDCA, and on the other the ARSM, and its expanded version H-ARSM to improve transmission multicast traffic. The scheme B-EDCA has been designed to enhance the benefits offered to multimedia applications both in the presence and absence of stations in IEEE 802.11. Because of the high incidence of this type of network, maintaining compatibility with the IEEE Standard 802.11 is one of the starting points of our scheme. For the definition of the proposal was made an exhaustive study of the functions of DCF access to the canal and EDCA, which found that the BPI is the most important parameter in both functions to provide good levels of QoS. This timer is used in two different times, and in both cases have the same values. We propose the division of this timer into two, so that one hand will continue to maintain compatibility between standards, and continue to give guaranteed access to AP to access the channel, and other improved benefits package multimedia applications when using more appropriate values in the second timer. The results have shown that this scheme is able to enhance the benefits offered to multimedia applications both in the presence and absence of stations in IEEE 802.11. The results also have been found as the new proposed scheme also get improved benefits achieved by the rest of improvement schemes in the literature. The second objective of this PhD thesis has been to improve communications multicast networks IEEE 802.11. In this standard, sending multicast packets of data is performed through a simple mechanism broadcast in which the issuer does not expect the confirmation of the data packets. This simple scheme used to avoid the avalanche of collisions on surveys of multicast packet data presents two major problems. The first is the impossibility of providing a standard reliable multicast communication. Because the data packets are not confirmed, none of the packages sent in error, whether by collisions or by the use of a wealth of sending inappropriate, it will be broadcast, which will impact on an unacceptable quality of the multimedia content delivered in the receivers. The second problem is the lack of adaptation of the flow of shipment to the changing conditions of the wireless channel. Due to lack of information feedback, the flow of sending data packets multicast must be set by the system administrator depending on the coverage area for it to the network. Therefore, in this Doctoral Thesis proposed mechanisms ARSM and H-ARSM as mechanisms for efficient multicast support for the transmission 8 video 49d either version is not scalable as scalable. ARSM and H-ARSM provide an automatic selection flow at the same time providing reliable transmission networks 802.11 capable multicaudal. The proposals in this Doctoral Thesis can obtain adequate video quality experienced by the end user. The results shown in this Doctoral Thesis have shown that both schemes improve the transmission of multicast traffic, making this shipment more efficiently.

Author: Fritzsche Paula Cecilia.
Year: 2006.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: Escuela Técnica Superior de Ingeniería.
Place of preparation: Facultad de Veterinaria.

Summary: This thesis tries to take a step forward in a little explored area. The same practice presents a new methodology for predicting performance of applications dependent on the data in a parallel computer, which reflects a non determinista.Dichas applications must meet certain characteristics such as requiring massive calculations and be ready to work parallel with all that this involves (synchronization, communication). The idea of working in this direction arises from the need to respond to a number of parallel applications that are normally frequent left to one side. Any contribution in the area not to be minimal by deterministic represents a major step forward in the absence of general knowledge. Not only must we think that the problem is solved but it represents for the benefit of family problems containing the problem resuelto.La proposal is based on certain assumptions and uses computer science to extract knowledge from experiments designed (input ), in a first phase, and then provides a mechanism for predicting which is quite promising, in a second phase. The formulation of the assumptions, the composition of the experiments, the execution of the application studied, the process of discovering knowledge, interpretation of the model and quality assessment are the steps that are part of phase extraction of knowledge, while the phase comparison of data belongs to the phase of prediction. A representative implementation dependent on the input data is the problem of traveler (TSP), where it is difficult to predict the run-time depend heavily on whole cities of entry. This problem is of considerable significance not only from a theoretical point of view. There are many cases of major practical problems can be formulated as TSP problems and many other problems are generalizations of this problem. Three parallel algorithms TSP euclideano (exhaustive algorithm, the algorithm pruning local and global pruning algorithm) are presented and then studied in detail to show the usefulness and benefits of the methodology presented. The algorithms have TSP as input the number of cities and their coordinates, the city start and the number of processors among other data, process them and generate a large volume of information between them and the way minimum distance (optimal solution ). The algorithms were developed in order to analyze the influence of varying degrees of indeterminismo in predicting performance. All follow the paradigm of programming teacher worker, one of the paradigms used more extensively in the field of parallelism. It is also important to mention that the algorithms TSP recursively searching all possible paths (permutations) through a branching process and differ only in the strategy of pruning used. Certainly, that the order of execution time of parallel algorithms TSP pruning local and global pruning depends on the number of cities (C), the number of processors (P) and other parameters associated with data entry, OR (C, P?). Finding these other parameters is key to getting a good or acceptable prediction of performance. Currently, the assumption is that workers are the sum of the distances (SD) of a city to the other cities and diversion half of these amounts (MDSD), play a major role in the execution times of the algorithms mentioned, OR (C, P, SD, MDSD). However, neither the sum of the distance from one city to other cities or diversion half of these amounts, affect weather final performance in the comprehensive algorithm, O (C, P).

Author: CAYMES SCUTARI PAOLA GUADALUPE.
Year: 2006.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: DEP. DE ARQUITE.DE COMPU.Y SISTEMAS OPERATIVOS.
Place of preparation: ESCUELA TECNICA SUPERIOR DE INGENIERIA.

Summary: In the last years, computing performance demand has been in increase. This necessity appeared specially in different scientific areas that have to solve complex problems. Thus, Biology, Physics and Chemistry are becoming the main producers and users of applications with high performance computing requirements. There are many applications that differ from the functional point of view, such as the determining of the human genome, the simulation of the universe, nature models study, etc. However, in general the data set size and the complexity of the operations over them require the use of very powerful systems in order to solve the problem as fast as possible and using the resources in an efficient way. Thus, the increasing necessity for high performance systems/computing has been directing the attention of the scientific field towards the parallel/distributed paradigm. The development of parallel applications has to follow a specific manner to allow for their execution in a parallel system. In addition, once the application has been implemented, it has to be systematically tested in order to guarantee its correctness. Following that, the application has to be adjusted to ensure that no bottlenecks exist in the execution. The optimization process, so-called tuning process, is the process followed in order to adapt and improve the behaviour of the applications by modifying their critical parameters. It includes several and successive phases. Firstly, during a monitoring phase the information about the behaviour of the application is captured. Next, the information is analyzed, by looking for bottlenecks, deducing their causes and trying to determine the adequate actions to eliminate them. Finally, appropriate changes have to be applied to the code to solve the problems and improve the performance. As a consequence, the developers are forced to know very well the application, the different involved software layers and the behaviour of the distributed system. All these issues make difficult and costly the performance tuning process, specially for non-expert users, due to a high degree of expertise is required in order to significantly improve the behaviour of the application. Fortunately, through the years different approaches and tools have been developed with the aim of helping the user during some of the optimization phases (monitoring, analysis or tuning phases). However, in general all of them require the user to know parallel programming in depth and take an active part in tuning the application. In consequence, tools capable of automating the tuning of parallel programs in a user-friendliness way need to be provided. One of the available tools is MATE (Monitoring, Analysis and Tuning Environment), which is an automatic and dynamic tuning environment for parallel applications. As its name indicates, MATE works in three continuous and iterative phases in order to adapt the deployment of the application according to the current state of the execution environment. MATE includes the knowledge to tune performance problems in pieces of software called ``tunlets''. Each tunlet includes the logic to collect behavioural information, analyze it on the fly and decide what the required tuning actions are. The objective of this work is to extend the usability of MATE. Our work covers two different aspects of MATE: * the improvement of the performance reached by the centralized analysis executed by MATE, due to it turns in a bottleneck as the size of the application increases. Thus, we provide an alternative to provide MATE with scalability properties. * the increase in the user-friendliness of MATE in order to facilitate the inclusion of new performance knowl 8 edge in 9b4 it. Thus, we make MATE transparent for the users. According to the first aspect, we propose a novel approach to execute the analysis process, called Distributed-Hierarchical Collecting-Preprocessing Approach. This approach is based on the distributed collection of events which alleviates the centralized old-fashion in which collection was done, and in the preprocessing of cumulative or comparative operations as possible. Thus, the Global Analyzer receives just the necessary information condensed in a unique message from each Collector-Preprocessor, which considerably reduces the overload of Global Analyzer. In this way, MATE is provided with scalability properties. According to the second aspect, we provide a methodology, including a designed language and a developed translator to automatically insert tunlets (knowledge) in MATE. When some problem has to be tuned in a parallel application the user has to develop the corresponding tunlet. By using our methodology, the user is exempted from being involved in implementation details of MATE. Thus, by defining a set of abstractions about the application and the performance model, such abstractions can be formalized in a tunlet specification using the provided language. Such specification will be automatically translated in a tunlet ready to be used in MATE. Both scalability and transparency of MATE are qualities necessary to make MATE a more useful and user-friendly tool. The proposals and developments presented attempt to provide MATE with such characteristics. Experiments showed the viability of the proposals. We trust our work is an important contribution from automatic and dynamic tuning to overcome some problems. We hope investigation continues on this area, then we will be able to see in an immediate future the results of such advances.

Author: Santos Coelho de Souza Leandro.
Year: 2006.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: Escola Tècnica Superior d' Enginyeria.
Place of preparation: ETSE-Escola Tècnica Superior d Enginyeria.

Summary: Advances in network technology will provide access to new generation, fully-interactive and client-oriented services such as the Video-on-Demand (VoD). Through these services, users will be able to down-stream video data from remote sites at any time on-demand. During a streaming session, VoD servers provide a continuous visualization of multimedia content, guaranteeing QoS (Quality of Service) levels. VoD services are proving critical for education, entertainment, business, and video-conference proposals. Streaming Video-on-Demand over the Internet is the next major step in the evolution of media content delivery. The global extension of the Internet network makes this environment very attractive for providing VoD services to millions of potential clients. In order to provide Internet VoD services that accommodate a huge number of concurrent users, the design of Large-Scale Video-on-Demand systems (LVoD) is required. The development of these systems requires a highly scalable, fault-tolerant and reliable LVoD architecture capable of providing high performance with limited costs. Nowadays the main alternative that has emerged for deploying VoD services on the Internet is the utilization of the peer-to-peer (P2P) paradigm. A P2P streaming architecture can offer better streaming capacity than conventional server-based architectures, bringing clients' computational and storage resources into the system, hence reducing the workload placed on the server and thereby increasing the overall system scalability. However, the Internet environment imposes further restrictions on P2P streaming schemes in order to provide VoD service. Problems such as quality of service (QoS) support, packet loss, limited bandwidth over peer resources, different Internet Service Providers' (ISP) technology support and content copyright protection are examples of Internet restrictions. This thesis proposes a new dynamic Peer-to-Peer VoD system in the Internet environment, called DynaPeer. DynaPeer resolves the Internet restrictions through the implementation of several techniques for effective client management. In order to overcome non-dedicated network and heterogeneity problems, DynaPeer is based on delivery policies capable of dynamically adapting themselves to network conditions. In this way, all requests for service will be monitored, analyzing traffic and performance to allow the system to react quickly if the QoS level is in danger. Furthermore, to deal with the heterogeneous network characteristics of ISPs, DynaPeer provides two basic delivery policies. The first policy is restricted to Unicast environments while the second takes advantage of the multicast capabilities of the clients' networks to dynamically exploit the IP multicast mechanism. The peers' asymmetrical bandwidth feature is addressed by a multi-peer collaboration process. Aggregating resources of several peers in a coordinated way will make it possible to accumulate enough resources (output bandwidth, buffer storage, etc) to generate a stream service to new incoming requests. The copyright issue is addressed by the incorporation of a sliding window technique for video information. This technique works by allowing peers to store and deliver a portion of content for different periods of time. In this way, peers never have complete access to full video data. To implement the new peer-to-peer VoD system DynaPeer defines a Virtual Server, a logic entity, responsible for managing client collaboration by guaranteeing the LVoD requirements and solving Internet constraints. Thus, Virtual Server implements key policies for P2P VoD-streaming: control admission, scheduling, delivery, flow control and fault tolerance.