El ràpid creixement de la Web requereix una gran quantitat de recursos computacionals que han de ser utilitzats eficientment. Avui en dia, els servidors basats en hardware estendard son les plataformes preferides per executar els servidors web, ja que són les plataformes amb millor relació rendiment/cost. El treball presentat en aquesta tesi esta dirigit a millorar la eficàcia en la gestió de recursos dels servidors web actuals. Per assolir els objectius d'aquesta tesis s'ha caracteritzat el funcionament dels servidors web en diverses entorns representatius, per tal de identificar el problemes i coll d'ampolla que limiten el rendiment del servidor web. Amb l'estudi dels servidors web s'ha identificat dos problemes principals que disminueixen l'eficiència dels servidors web en la utilització dels recursos hardware disponibles. El primer problema identificat és la evolució del protocol HTTP per incorporar connexions persistents i seguretat, que disminueix el rendiment e incrementa la complexitat de configuració dels servidors web. El segon problema és la naturalesa de algunes aplicacions web, les quals estan limitades per la memòria física o l'ample de banda amb el disc, que impedeix la correcta utilització dels recursos presents en les maquines multiprocessadors. Per solucionar aquests dos problemes dels servidors web hem proposat dues tècniques. En primer lloc, l'arquitectura hibrida, una evolució de l'arquitectura multi-threaded que es pot implementar fàcilment el els servidor web actuals i que millora notablement la gestió de les connexions i redueix la complexitat de configuració de tot el sistema. En segon lloc, hem implementat en el kernel del sistema operatiu Linux un comprensió de memòria principal per millorar el rendiment de les aplicacions que tenen la memòria com ha coll d'ampolla, millorant així la utilització dels recursos disponibles. Els resultats d'aquesta tesis estan avalats per una avaluació experimental exhaustiva que ha provat la efectivitat i viabilitat de les nostres propostes. Cal destacar que l'arquitectura de servidor web hybrida proposada en aquesta tesis ha estat implementada recentment per coneguts servidors web com és el cas de Apache, Tomcat i Glassfish. / The unstoppable growth of the World Wide Web requires a huge amount of computational resources that must be used efficiently. Nowadays, commodity hardware is the preferred platform to run web server systems because it is the most cost-effective solution. The work presented in this thesis aims to improve the efficiency of current web server systems, allowing the web servers to make the most of hardware resources. To this end, we first characterize current web server system and identify the problems that hinder web servers from providing an efficient utilization of resources. From the study of web servers in a wide range of situations and environments, we have identified two main issues that prevents web servers systems from efficiently using current hardware resources. The first is the extension of the HTTP protocol to include connection persistence and security, which dramatically impacts the performance and configuration complexity of traditional multi-threaded web servers. The second is the memory-bounded or disk-bounded nature of some web workloads that prevents the full utilization of the abundant CPU resources available on current commodity hardware. We propose two novel techniques to overcome the main problems with current web server systems. Firstly, we propose a Hybrid web serverarchitecture which can be easily implemented in any multi-threaded web server to improve CPU utilization so as to provide better management of client connections. And secondly, we describe a main memory compression technique implemented in the Linux operating system that makes optimum use of current multiprocessor's hardware, in order to improve the performance of memory bound web applications. The thesis is supported by an exhaustive experimental evaluation that proves the effectiveness and feasibility of our proposals for current systems. It is worth noting that the main concepts behind the Hybrid architecture have recently been implemented in popular web servers like Apache, Tomcat and Glassfish.

abstract: Speech recognition and keyword detection are becoming increasingly popular applications for mobile systems. While deep neural network (DNN) implementation of these systems have very good performance, they have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this thesis, techniques to reduce the memory and computation cost of keyword detection and speech recognition networks (or DNNs) are presented. The first technique is based on representing all weights and biases by a small number of bits and mapping all nodal computations into fixed-point ones with minimal degradation in the accuracy. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Additional reduction in memory is achieved by a technique called weight pruning, where the weights are classified as sensitive and insensitive and the sensitive weights are represented with higher precision. A combination of these two techniques helps reduce the memory footprint by 81 - 84% for speech recognition and keyword detection networks respectively. Further reduction in memory size is achieved by judiciously dropping connections for large blocks of weights. The corresponding technique, termed coarse-grain sparsification, introduces hardware-aware sparsity during DNN training, which leads to efficient weight memory compression and significant reduction in the number of computations during classification without loss of accuracy. Keyword detection and speech recognition DNNs trained with 75% of the weights dropped and classified with 5-6 bit weight precision effectively reduced the weight memory requirement by ~95% compared to a fully-connected network with double precision, while showing similar performance in keyword detection accuracy and word error rate. / Dissertation/Thesis / Masters Thesis Computer Science 2016