Real-Time Application Mapping for Many-Cores Using a Limited Migrative Model

Many-core platforms are an emerging technology in the real-time embedded domain. These devices offer various options for power savings, cost reductions and contribute to the overall system flexibility, however, issues such as unpredictability, scalability and analysis pessimism are serious challenges to their integration into the aforementioned area. The focus of this work is on many-core platforms using a limited migrative model (LMM). LMM is an approach based on the fundamental concepts of the multi-kernel paradigm, which is a promising step towards scalable and predictable many-cores. In this work, we formulate the problem of real-time application mapping on a many-core platform using LMM, and propose a three-stage method to solve it. An extended version of the existing analysis is used to assure that derived mappings (i) guarantee the fulfilment of timing constraints posed on worst-case communication delays of individual applications, and (ii) provide an environment to perform load balancing for e.g. energy/thermal management, fault tolerance and/or performance reasons.

Use of shared memory in the context of embedded multi-core processors: exploration of the technology and its limits

Modern embedded systems embrace many-core shared-memory designs. Due to constrained power and area budgets, most of them feature software-managed scratchpad memories instead of data caches to increase the data locality. It is therefore programmers’ responsibility to explicitly manage the memory transfers, and this make programming these platform cumbersome. Moreover, complex modern applications must be adequately parallelized before they can the parallel potential of the platform into actual performance. To support this, programming languages were proposed, which work at a high level of abstraction, and rely on a runtime whose cost hinders performance, especially in embedded systems, where resources and power budget are constrained. This dissertation explores the applicability of the shared-memory paradigm on modern many-core systems, focusing on the ease-of-programming. It focuses on OpenMP, the de-facto standard for shared memory programming. In a first part, the cost of algorithms for synchronization and data partitioning are analyzed, and they are adapted to modern embedded many-cores. Then, the original design of an OpenMP runtime library is presented, which supports complex forms of parallelism such as multi-level and irregular parallelism. In the second part of the thesis, the focus is on heterogeneous systems, where hardware accelerators are coupled to (many-)cores to implement key functional kernels with orders-of-magnitude of speedup and energy efficiency compared to the “pure software” version. However, three main issues rise, namely i) platform design complexity, ii) architectural scalability and iii) programmability. To tackle them, a template for a generic hardware processing unit (HWPU) is proposed, which share the memory banks with cores, and the template for a scalable architecture is shown, which integrates them through the shared-memory system. Then, a full software stack and toolchain are developed to support platform design and to let programmers exploiting the accelerators of the platform. The OpenMP frontend is extended to interact with it.

Parametric and structural self-adaptation of embedded systems using evolvable hardware

Los sistemas empotrados han sido concebidos tradicionalmente como sistemas de procesamiento específicos que realizan una tarea fija durante toda su vida útil. Para cumplir con requisitos estrictos de coste, tamaño y peso, el equipo de diseño debe optimizar su funcionamiento para condiciones muy específicas. Sin embargo, la demanda de mayor versatilidad, un funcionamiento más inteligente y, en definitiva, una mayor capacidad de procesamiento comenzaron a chocar con estas limitaciones, agravado por la incertidumbre asociada a entornos de operación cada vez más dinámicos donde comenzaban a ser desplegados progresivamente. Esto trajo como resultado una necesidad creciente de que los sistemas pudieran responder por si solos a eventos inesperados en tiempo diseño tales como: cambios en las características de los datos de entrada y el entorno del sistema en general; cambios en la propia plataforma de cómputo, por ejemplo debido a fallos o defectos de fabricación; y cambios en las propias especificaciones funcionales causados por unos objetivos del sistema dinámicos y cambiantes. Como consecuencia, la complejidad del sistema aumenta, pero a cambio se habilita progresivamente una capacidad de adaptación autónoma sin intervención humana a lo largo de la vida útil, permitiendo que tomen sus propias decisiones en tiempo de ejecución. Éstos sistemas se conocen, en general, como sistemas auto-adaptativos y tienen, entre otras características, las de auto-configuración, auto-optimización y auto-reparación. Típicamente, la parte soft de un sistema es mayoritariamente la única utilizada para proporcionar algunas capacidades de adaptación a un sistema. Sin embargo, la proporción rendimiento/potencia en dispositivos software como microprocesadores en muchas ocasiones no es adecuada para sistemas empotrados. En este escenario, el aumento resultante en la complejidad de las aplicaciones está siendo abordado parcialmente mediante un aumento en la complejidad de los dispositivos en forma de multi/many-cores; pero desafortunadamente, esto hace que el consumo de potencia también aumente. Además, la mejora en metodologías de diseño no ha sido acorde como para poder utilizar toda la capacidad de cómputo disponible proporcionada por los núcleos. Por todo ello, no se están satisfaciendo adecuadamente las demandas de cómputo que imponen las nuevas aplicaciones. La solución tradicional para mejorar la proporción rendimiento/potencia ha sido el cambio a unas especificaciones hardware, principalmente usando ASICs. Sin embargo, los costes de un ASIC son altamente prohibitivos excepto en algunos casos de producción en masa y además la naturaleza estática de su estructura complica la solución a las necesidades de adaptación. Los avances en tecnologías de fabricación han hecho que la FPGA, una vez lenta y pequeña, usada como glue logic en sistemas mayores, haya crecido hasta convertirse en un dispositivo de cómputo reconfigurable de gran potencia, con una cantidad enorme de recursos lógicos computacionales y cores hardware empotrados de procesamiento de señal y de propósito general. Sus capacidades de reconfiguración han permitido combinar la flexibilidad propia del software con el rendimiento del procesamiento en hardware, lo que tiene la potencialidad de provocar un cambio de paradigma en arquitectura de computadores, pues el hardware no puede ya ser considerado más como estático. El motivo es que como en el caso de las FPGAs basadas en tecnología SRAM, la reconfiguración parcial dinámica (DPR, Dynamic Partial Reconfiguration) es posible. Esto significa que se puede modificar (reconfigurar) un subconjunto de los recursos computacionales en tiempo de ejecución mientras el resto permanecen activos. Además, este proceso de reconfiguración puede ser ejecutado internamente por el propio dispositivo. El avance tecnológico en dispositivos hardware reconfigurables se encuentra recogido bajo el campo conocido como Computación Reconfigurable (RC, Reconfigurable Computing). Uno de los campos de aplicación más exóticos y menos convencionales que ha posibilitado la computación reconfigurable es el conocido como Hardware Evolutivo (EHW, Evolvable Hardware), en el cual se encuentra enmarcada esta tesis. La idea principal del concepto consiste en convertir hardware que es adaptable a través de reconfiguración en una entidad evolutiva sujeta a las fuerzas de un proceso evolutivo inspirado en el de las especies biológicas naturales, que guía la dirección del cambio. Es una aplicación más del campo de la Computación Evolutiva (EC, Evolutionary Computation), que comprende una serie de algoritmos de optimización global conocidos como Algoritmos Evolutivos (EA, Evolutionary Algorithms), y que son considerados como algoritmos universales de resolución de problemas. En analogía al proceso biológico de la evolución, en el hardware evolutivo el sujeto de la evolución es una población de circuitos que intenta adaptarse a su entorno mediante una adecuación progresiva generación tras generación. Los individuos pasan a ser configuraciones de circuitos en forma de bitstreams caracterizados por descripciones de circuitos reconfigurables. Seleccionando aquellos que se comportan mejor, es decir, que tienen una mejor adecuación (o fitness) después de ser evaluados, y usándolos como padres de la siguiente generación, el algoritmo evolutivo crea una nueva población hija usando operadores genéticos como la mutación y la recombinación. Según se van sucediendo generaciones, se espera que la población en conjunto se aproxime a la solución óptima al problema de encontrar una configuración del circuito adecuada que satisfaga las especificaciones. El estado de la tecnología de reconfiguración después de que la familia de FPGAs XC6200 de Xilinx fuera retirada y reemplazada por las familias Virtex a finales de los 90, supuso un gran obstáculo para el avance en hardware evolutivo; formatos de bitstream cerrados (no conocidos públicamente); dependencia de herramientas del fabricante con soporte limitado de DPR; una velocidad de reconfiguración lenta; y el hecho de que modificaciones aleatorias del bitstream pudieran resultar peligrosas para la integridad del dispositivo, son algunas de estas razones. Sin embargo, una propuesta a principios de los años 2000 permitió mantener la investigación en el campo mientras la tecnología de DPR continuaba madurando, el Circuito Virtual Reconfigurable (VRC, Virtual Reconfigurable Circuit). En esencia, un VRC en una FPGA es una capa virtual que actúa como un circuito reconfigurable de aplicación específica sobre la estructura nativa de la FPGA que reduce la complejidad del proceso reconfiguración y aumenta su velocidad (comparada con la reconfiguración nativa). Es un array de nodos computacionales especificados usando descripciones HDL estándar que define recursos reconfigurables ad-hoc: multiplexores de rutado y un conjunto de elementos de procesamiento configurables, cada uno de los cuales tiene implementadas todas las funciones requeridas, que pueden seleccionarse a través de multiplexores tal y como ocurre en una ALU de un microprocesador. Un registro grande actúa como memoria de configuración, por lo que la reconfiguración del VRC es muy rápida ya que tan sólo implica la escritura de este registro, el cual controla las señales de selección del conjunto de multiplexores. Sin embargo, esta capa virtual provoca: un incremento de área debido a la implementación simultánea de cada función en cada nodo del array más los multiplexores y un aumento del retardo debido a los multiplexores, reduciendo la frecuencia de funcionamiento máxima. La naturaleza del hardware evolutivo, capaz de optimizar su propio comportamiento computacional, le convierten en un buen candidato para avanzar en la investigación sobre sistemas auto-adaptativos. Combinar un sustrato de cómputo auto-reconfigurable capaz de ser modificado dinámicamente en tiempo de ejecución con un algoritmo empotrado que proporcione una dirección de cambio, puede ayudar a satisfacer los requisitos de adaptación autónoma de sistemas empotrados basados en FPGA. La propuesta principal de esta tesis está por tanto dirigida a contribuir a la auto-adaptación del hardware de procesamiento de sistemas empotrados basados en FPGA mediante hardware evolutivo. Esto se ha abordado considerando que el comportamiento computacional de un sistema puede ser modificado cambiando cualquiera de sus dos partes constitutivas: una estructura hard subyacente y un conjunto de parámetros soft. De esta distinción, se derivan dos lineas de trabajo. Por un lado, auto-adaptación paramétrica, y por otro auto-adaptación estructural. El objetivo perseguido en el caso de la auto-adaptación paramétrica es la implementación de técnicas de optimización evolutiva complejas en sistemas empotrados con recursos limitados para la adaptación paramétrica online de circuitos de procesamiento de señal. La aplicación seleccionada como prueba de concepto es la optimización para tipos muy específicos de imágenes de los coeficientes de los filtros de transformadas wavelet discretas (DWT, DiscreteWavelet Transform), orientada a la compresión de imágenes. Por tanto, el objetivo requerido de la evolución es una compresión adaptativa y más eficiente comparada con los procedimientos estándar. El principal reto radica en reducir la necesidad de recursos de supercomputación para el proceso de optimización propuesto en trabajos previos, de modo que se adecúe para la ejecución en sistemas empotrados. En cuanto a la auto-adaptación estructural, el objetivo de la tesis es la implementación de circuitos auto-adaptativos en sistemas evolutivos basados en FPGA mediante un uso eficiente de sus capacidades de reconfiguración nativas. En este caso, la prueba de concepto es la evolución de tareas de procesamiento de imagen tales como el filtrado de tipos desconocidos y cambiantes de ruido y la detección de bordes en la imagen. En general, el objetivo es la evolución en tiempo de ejecución de tareas de procesamiento de imagen desconocidas en tiempo de diseño (dentro de un cierto grado de complejidad). En este caso, el objetivo de la propuesta es la incorporación de DPR en EHW para evolucionar la arquitectura de un array sistólico adaptable mediante reconfiguración cuya capacidad de evolución no había sido estudiada previamente. Para conseguir los dos objetivos mencionados, esta tesis propone originalmente una plataforma evolutiva que integra un motor de adaptación (AE, Adaptation Engine), un motor de reconfiguración (RE, Reconfiguration Engine) y un motor computacional (CE, Computing Engine) adaptable. El el caso de adaptación paramétrica, la plataforma propuesta está caracterizada por: • un CE caracterizado por un núcleo de procesamiento hardware de DWT adaptable mediante registros reconfigurables que contienen los coeficientes de los filtros wavelet • un algoritmo evolutivo como AE que busca filtros wavelet candidatos a través de un proceso de optimización paramétrica desarrollado específicamente para sistemas caracterizados por recursos de procesamiento limitados • un nuevo operador de mutación simplificado para el algoritmo evolutivo utilizado, que junto con un mecanismo de evaluación rápida de filtros wavelet candidatos derivado de la literatura actual, asegura la viabilidad de la búsqueda evolutiva asociada a la adaptación de wavelets. En el caso de adaptación estructural, la plataforma propuesta toma la forma de: • un CE basado en una plantilla de array sistólico reconfigurable de 2 dimensiones compuesto de nodos de procesamiento reconfigurables • un algoritmo evolutivo como AE que busca configuraciones candidatas del array usando un conjunto de funcionalidades de procesamiento para los nodos disponible en una biblioteca accesible en tiempo de ejecución • un RE hardware que explota la capacidad de reconfiguración nativa de las FPGAs haciendo un uso eficiente de los recursos reconfigurables del dispositivo para cambiar el comportamiento del CE en tiempo de ejecución • una biblioteca de elementos de procesamiento reconfigurables caracterizada por bitstreams parciales independientes de la posición, usados como el conjunto de configuraciones disponibles para los nodos de procesamiento del array Las contribuciones principales de esta tesis se pueden resumir en la siguiente lista: • Una plataforma evolutiva basada en FPGA para la auto-adaptación paramétrica y estructural de sistemas empotrados compuesta por un motor computacional (CE), un motor de adaptación (AE) evolutivo y un motor de reconfiguración (RE). Esta plataforma se ha desarrollado y particularizado para los casos de auto-adaptación paramétrica y estructural. • En cuanto a la auto-adaptación paramétrica, las contribuciones principales son: – Un motor computacional adaptable mediante registros que permite la adaptación paramétrica de los coeficientes de una implementación hardware adaptativa de un núcleo de DWT. – Un motor de adaptación basado en un algoritmo evolutivo desarrollado específicamente para optimización numérica, aplicada a los coeficientes de filtros wavelet en sistemas empotrados con recursos limitados. – Un núcleo IP de DWT auto-adaptativo en tiempo de ejecución para sistemas empotrados que permite la optimización online del rendimiento de la transformada para compresión de imágenes en entornos específicos de despliegue, caracterizados por tipos diferentes de señal de entrada. – Un modelo software y una implementación hardware de una herramienta para la construcción evolutiva automática de transformadas wavelet específicas. • Por último, en cuanto a la auto-adaptación estructural, las contribuciones principales son: – Un motor computacional adaptable mediante reconfiguración nativa de FPGAs caracterizado por una plantilla de array sistólico en dos dimensiones de nodos de procesamiento reconfigurables. Es posible mapear diferentes tareas de cómputo en el array usando una biblioteca de elementos sencillos de procesamiento reconfigurables. – Definición de una biblioteca de elementos de procesamiento apropiada para la síntesis autónoma en tiempo de ejecución de diferentes tareas de procesamiento de imagen. – Incorporación eficiente de la reconfiguración parcial dinámica (DPR) en sistemas de hardware evolutivo, superando los principales inconvenientes de propuestas previas como los circuitos reconfigurables virtuales (VRCs). En este trabajo también se comparan originalmente los detalles de implementación de ambas propuestas. – Una plataforma tolerante a fallos, auto-curativa, que permite la recuperación funcional online en entornos peligrosos. La plataforma ha sido caracterizada desde una perspectiva de tolerancia a fallos: se proponen modelos de fallo a nivel de CLB y de elemento de procesamiento, y usando el motor de reconfiguración, se hace un análisis sistemático de fallos para un fallo en cada elemento de procesamiento y para dos fallos acumulados. – Una plataforma con calidad de filtrado dinámica que permite la adaptación online a tipos de ruido diferentes y diferentes comportamientos computacionales teniendo en cuenta los recursos de procesamiento disponibles. Por un lado, se evolucionan filtros con comportamientos no destructivos, que permiten esquemas de filtrado en cascada escalables; y por otro, también se evolucionan filtros escalables teniendo en cuenta requisitos computacionales de filtrado cambiantes dinámicamente. Este documento está organizado en cuatro partes y nueve capítulos. La primera parte contiene el capítulo 1, una introducción y motivación sobre este trabajo de tesis. A continuación, el marco de referencia en el que se enmarca esta tesis se analiza en la segunda parte: el capítulo 2 contiene una introducción a los conceptos de auto-adaptación y computación autonómica (autonomic computing) como un campo de investigación más general que el muy específico de este trabajo; el capítulo 3 introduce la computación evolutiva como la técnica para dirigir la adaptación; el capítulo 4 analiza las plataformas de computación reconfigurables como la tecnología para albergar hardware auto-adaptativo; y finalmente, el capítulo 5 define, clasifica y hace un sondeo del campo del hardware evolutivo. Seguidamente, la tercera parte de este trabajo contiene la propuesta, desarrollo y resultados obtenidos: mientras que el capítulo 6 contiene una declaración de los objetivos de la tesis y la descripción de la propuesta en su conjunto, los capítulos 7 y 8 abordan la auto-adaptación paramétrica y estructural, respectivamente. Finalmente, el capítulo 9 de la parte 4 concluye el trabajo y describe caminos de investigación futuros. ABSTRACT Embedded systems have traditionally been conceived to be specific-purpose computers with one, fixed computational task for their whole lifetime. Stringent requirements in terms of cost, size and weight forced designers to highly optimise their operation for very specific conditions. However, demands for versatility, more intelligent behaviour and, in summary, an increased computing capability began to clash with these limitations, intensified by the uncertainty associated to the more dynamic operating environments where they were progressively being deployed. This brought as a result an increasing need for systems to respond by themselves to unexpected events at design time, such as: changes in input data characteristics and system environment in general; changes in the computing platform itself, e.g., due to faults and fabrication defects; and changes in functional specifications caused by dynamically changing system objectives. As a consequence, systems complexity is increasing, but in turn, autonomous lifetime adaptation without human intervention is being progressively enabled, allowing them to take their own decisions at run-time. This type of systems is known, in general, as selfadaptive, and are able, among others, of self-configuration, self-optimisation and self-repair. Traditionally, the soft part of a system has mostly been so far the only place to provide systems with some degree of adaptation capabilities. However, the performance to power ratios of software driven devices like microprocessors are not adequate for embedded systems in many situations. In this scenario, the resulting rise in applications complexity is being partly addressed by rising devices complexity in the form of multi and many core devices; but sadly, this keeps on increasing power consumption. Besides, design methodologies have not been improved accordingly to completely leverage the available computational power from all these cores. Altogether, these factors make that the computing demands new applications pose are not being wholly satisfied. The traditional solution to improve performance to power ratios has been the switch to hardware driven specifications, mainly using ASICs. However, their costs are highly prohibitive except for some mass production cases and besidesthe static nature of its structure complicates the solution to the adaptation needs. The advancements in fabrication technologies have made that the once slow, small FPGA used as glue logic in bigger systems, had grown to be a very powerful, reconfigurable computing device with a vast amount of computational logic resources and embedded, hardened signal and general purpose processing cores. Its reconfiguration capabilities have enabled software-like flexibility to be combined with hardware-like computing performance, which has the potential to cause a paradigm shift in computer architecture since hardware cannot be considered as static anymore. This is so, since, as is the case with SRAMbased FPGAs, Dynamic Partial Reconfiguration (DPR) is possible. This means that subsets of the FPGA computational resources can now be changed (reconfigured) at run-time while the rest remains active. Besides, this reconfiguration process can be triggered internally by the device itself. This technological boost in reconfigurable hardware devices is actually covered under the field known as Reconfigurable Computing. One of the most exotic fields of application that Reconfigurable Computing has enabled is the known as Evolvable Hardware (EHW), in which this dissertation is framed. The main idea behind the concept is turning hardware that is adaptable through reconfiguration into an evolvable entity subject to the forces of an evolutionary process, inspired by that of natural, biological species, that guides the direction of change. It is yet another application of the field of Evolutionary Computation (EC), which comprises a set of global optimisation algorithms known as Evolutionary Algorithms (EAs), considered as universal problem solvers. In analogy to the biological process of evolution, in EHW the subject of evolution is a population of circuits that tries to get adapted to its surrounding environment by progressively getting better fitted to it generation after generation. Individuals become circuit configurations representing bitstreams that feature reconfigurable circuit descriptions. By selecting those that behave better, i.e., with a higher fitness value after being evaluated, and using them as parents of the following generation, the EA creates a new offspring population by using so called genetic operators like mutation and recombination. As generations succeed one another, the whole population is expected to approach to the optimum solution to the problem of finding an adequate circuit configuration that fulfils system objectives. The state of reconfiguration technology after Xilinx XC6200 FPGA family was discontinued and replaced by Virtex families in the late 90s, was a major obstacle for advancements in EHW; closed (non publicly known) bitstream formats; dependence on manufacturer tools with highly limiting support of DPR; slow speed of reconfiguration; and random bitstream modifications being potentially hazardous for device integrity, are some of these reasons. However, a proposal in the first 2000s allowed to keep investigating in this field while DPR technology kept maturing, the Virtual Reconfigurable Circuit (VRC). In essence, a VRC in an FPGA is a virtual layer acting as an application specific reconfigurable circuit on top of an FPGA fabric that reduces the complexity of the reconfiguration process and increases its speed (compared to native reconfiguration). It is an array of computational nodes specified using standard HDL descriptions that define ad-hoc reconfigurable resources; routing multiplexers and a set of configurable processing elements, each one containing all the required functions, which are selectable through functionality multiplexers as in microprocessor ALUs. A large register acts as configuration memory, so VRC reconfiguration is very fast given it only involves writing this register, which drives the selection signals of the set of multiplexers. However, large overheads are introduced by this virtual layer; an area overhead due to the simultaneous implementation of every function in every node of the array plus the multiplexers, and a delay overhead due to the multiplexers, which also reduces maximum frequency of operation. The very nature of Evolvable Hardware, able to optimise its own computational behaviour, makes it a good candidate to advance research in self-adaptive systems. Combining a selfreconfigurable computing substrate able to be dynamically changed at run-time with an embedded algorithm that provides a direction for change, can help fulfilling requirements for autonomous lifetime adaptation of FPGA-based embedded systems. The main proposal of this thesis is hence directed to contribute to autonomous self-adaptation of the underlying computational hardware of FPGA-based embedded systems by means of Evolvable Hardware. This is tackled by considering that the computational behaviour of a system can be modified by changing any of its two constituent parts: an underlying hard structure and a set of soft parameters. Two main lines of work derive from this distinction. On one side, parametric self-adaptation and, on the other side, structural self-adaptation. The goal pursued in the case of parametric self-adaptation is the implementation of complex evolutionary optimisation techniques in resource constrained embedded systems for online parameter adaptation of signal processing circuits. The application selected as proof of concept is the optimisation of Discrete Wavelet Transforms (DWT) filters coefficients for very specific types of images, oriented to image compression. Hence, adaptive and improved compression efficiency, as compared to standard techniques, is the required goal of evolution. The main quest lies in reducing the supercomputing resources reported in previous works for the optimisation process in order to make it suitable for embedded systems. Regarding structural self-adaptation, the thesis goal is the implementation of self-adaptive circuits in FPGA-based evolvable systems through an efficient use of native reconfiguration capabilities. In this case, evolution of image processing tasks such as filtering of unknown and changing types of noise and edge detection are the selected proofs of concept. In general, evolving unknown image processing behaviours (within a certain complexity range) at design time is the required goal. In this case, the mission of the proposal is the incorporation of DPR in EHW to evolve a systolic array architecture adaptable through reconfiguration whose evolvability had not been previously checked. In order to achieve the two stated goals, this thesis originally proposes an evolvable platform that integrates an Adaptation Engine (AE), a Reconfiguration Engine (RE) and an adaptable Computing Engine (CE). In the case of parametric adaptation, the proposed platform is characterised by: • a CE featuring a DWT hardware processing core adaptable through reconfigurable registers that holds wavelet filters coefficients • an evolutionary algorithm as AE that searches for candidate wavelet filters through a parametric optimisation process specifically developed for systems featured by scarce computing resources • a new, simplified mutation operator for the selected EA, that together with a fast evaluation mechanism of candidate wavelet filters derived from existing literature, assures the feasibility of the evolutionary search involved in wavelets adaptation In the case of structural adaptation, the platform proposal takes the form of: • a CE based on a reconfigurable 2D systolic array template composed of reconfigurable processing nodes • an evolutionary algorithm as AE that searches for candidate configurations of the array using a set of computational functionalities for the nodes available in a run time accessible library • a hardware RE that exploits native DPR capabilities of FPGAs and makes an efficient use of the available reconfigurable resources of the device to change the behaviour of the CE at run time • a library of reconfigurable processing elements featured by position-independent partial bitstreams used as the set of available configurations for the processing nodes of the array Main contributions of this thesis can be summarised in the following list. • An FPGA-based evolvable platform for parametric and structural self-adaptation of embedded systems composed of a Computing Engine, an evolutionary Adaptation Engine and a Reconfiguration Engine. This platform is further developed and tailored for both parametric and structural self-adaptation. • Regarding parametric self-adaptation, main contributions are: – A CE adaptable through reconfigurable registers that enables parametric adaptation of the coefficients of an adaptive hardware implementation of a DWT core. – An AE based on an Evolutionary Algorithm specifically developed for numerical optimisation applied to wavelet filter coefficients in resource constrained embedded systems. – A run-time self-adaptive DWT IP core for embedded systems that allows for online optimisation of transform performance for image compression for specific deployment environments characterised by different types of input signals. – A software model and hardware implementation of a tool for the automatic, evolutionary construction of custom wavelet transforms. • Lastly, regarding structural self-adaptation, main contributions are: – A CE adaptable through native FPGA fabric reconfiguration featured by a two dimensional systolic array template of reconfigurable processing nodes. Different processing behaviours can be automatically mapped in the array by using a library of simple reconfigurable processing elements. – Definition of a library of such processing elements suited for autonomous runtime synthesis of different image processing tasks. – Efficient incorporation of DPR in EHW systems, overcoming main drawbacks from the previous approach of virtual reconfigurable circuits. Implementation details for both approaches are also originally compared in this work. – A fault tolerant, self-healing platform that enables online functional recovery in hazardous environments. The platform has been characterised from a fault tolerance perspective: fault models at FPGA CLB level and processing elements level are proposed, and using the RE, a systematic fault analysis for one fault in every processing element and for two accumulated faults is done. – A dynamic filtering quality platform that permits on-line adaptation to different types of noise and different computing behaviours considering the available computing resources. On one side, non-destructive filters are evolved, enabling scalable cascaded filtering schemes; and on the other, size-scalable filters are also evolved considering dynamically changing computational filtering requirements. This dissertation is organized in four parts and nine chapters. First part contains chapter 1, the introduction to and motivation of this PhD work. Following, the reference framework in which this dissertation is framed is analysed in the second part: chapter 2 features an introduction to the notions of self-adaptation and autonomic computing as a more general research field to the very specific one of this work; chapter 3 introduces evolutionary computation as the technique to drive adaptation; chapter 4 analyses platforms for reconfigurable computing as the technology to hold self-adaptive hardware; and finally chapter 5 defines, classifies and surveys the field of Evolvable Hardware. Third part of the work follows, which contains the proposal, development and results obtained: while chapter 6 contains an statement of the thesis goals and the description of the proposal as a whole, chapters 7 and 8 address parametric and structural self-adaptation, respectively. Finally, chapter 9 in part 4 concludes the work and describes future research paths.

An optimising compiler for generated tiny virtual machines

VMSCRIPT is a scripting language designed to allow small programs to be compiled for a range of generated tiny virtual machines, suitable for sensor network devices. The VMSCRIPT compiler is an optimising compiler designed to allow quick re-targeting, based on a template, code rewriting model. A compiler backend can be specified at the same time as a virtual machine, with the compiler reading the specification and using it as a code generator.

A random key genetic algorithm for live migration of multiple virtual machines in data centers

Live migration of multiple Virtual Machines (VMs) has become an integral management activity in data centers for power saving, load balancing and system maintenance. While state-of-the-art live migration techniques focus on the improvement of migration performance of an independent single VM, only a little has been investigated to the case of live migration of multiple interacting VMs. Live migration is mostly influenced by the network bandwidth and arbitrarily migrating a VM which has data inter-dependencies with other VMs may increase the bandwidth consumption and adversely affect the performances of subsequent migrations. In this paper, we propose a Random Key Genetic Algorithm (RKGA) that efficiently schedules the migration of a given set of VMs accounting both inter-VM dependency and data center communication network. The experimental results show that the RKGA can schedule the migration of multiple VMs with significantly shorter total migration time and total downtime compared to a heuristic algorithm.

Evaluation of Interconnect Fabrics for an Embedded MPSoC in 28 nm FD-SOI

Embedded many-core architectures contain dozens to hundreds of CPU cores that are connected via a highly scalable NoC interconnect. Our Multiprocessor-System-on-Chip CoreVAMPSoC combines the advantages of tightly coupled bus-based communication with the scalability of NoC approaches by adding a CPU cluster as an additional level of hierarchy. In this work, we analyze different cluster interconnect implementations with 8 to 32 CPUs and compare them in terms of resource requirements and performance to hierarchical NoCs approaches. Using 28nm FD-SOI technology the area requirement for 32 CPUs and AXI crossbar is 5.59mm2 including 23.61% for the interconnect at a clock frequency of 830 MHz. In comparison, a hierarchical MPSoC with 4 CPU cluster and 8 CPUs in each cluster requires only 4.83mm2 including 11.61% for the interconnect. To evaluate the performance, we use a compiler for streaming applications to map programs to the different MPSoC configurations. We use this approach for a design-space exploration to find the most efficient architecture and partitioning for an application.

Performance estimation of streaming applications for hierarchical MPSoCs

Parallel programming and effective partitioning of applications for embedded many-core architectures requires optimization algorithms. However, these algorithms have to quickly evaluate thousands of different partitions. We present a fast performance estimator embedded in a parallelizing compiler for streaming applications. The estimator combines a single execution-based simulation and an analytic approach. Experimental results demonstrate that the estimator has a mean error of 2.6% and computes its estimation 2848 times faster compared to a cycle accurate simulator.

Co-design of Distributed Systems Using Skeleton and Autonomic Management Abstractions

We discuss how common problems arising with multi/many core distributed architectures can he effectively handled through co-design of parallel/distributed programming abstractions and of autonomic management of non-functional concerns. In particular, we demonstrate how restricted patterns (or skeletons) may be efficiently managed by rule-based autonomic managers. We discuss the basic principles underlying pattern+manager co-design, current implementations inspired by this approach and some result achieved with proof-or-concept, prototype.

An abstract annotation model for skeletons

Multi-core and many-core platforms are becoming increasingly heterogeneous and asymmetric. This significantly increases the porting and tuning effort required for parallel codes, which in turn often leads to a growing gap between peak machine power and actual application performance. In this work a first step toward the automated optimization of high level skeleton-based parallel code is discussed. The paper presents an abstract annotation model for skeleton programs aimed at formally describing suitable mapping of parallel activities on a high-level platform representation. The derived mapping and scheduling strategies are used to generate optimized run-time code. © 2013 Springer-Verlag Berlin Heidelberg.

Analysis of Dependence Tracking Algorithms for Task Dataflow Execution

Processor architectures has taken a turn towards many-core processors, which integrate multiple processing cores on a single chip to increase overall performance, and there are no signs that this trend will stop in the near future. Many-core processors are harder to program than multi-core and single-core processors due to the need of writing parallel or concurrent programs with high degrees of parallelism. Moreover, many-cores have to operate in a mode of strong scaling because of memory bandwidth constraints. In strong scaling increasingly finer-grain parallelism must be extracted in order to keep all processing cores busy.

Task dataflow programming models have a high potential to simplify parallel program- ming because they alleviate the programmer from identifying precisely all inter-task de- pendences when writing programs. Instead, the task dataflow runtime system detects and enforces inter-task dependences during execution based on the description of memory each task accesses. The runtime constructs a task dataflow graph that captures all tasks and their dependences. Tasks are scheduled to execute in parallel taking into account dependences specified in the task graph.

Several papers report important overheads for task dataflow systems, which severely limits the scalability and usability of such systems. In this paper we study efficient schemes to manage task graphs and analyze their scalability. We assume a programming model that supports input, output and in/out annotations on task arguments, as well as commutative in/out and reductions. We analyze the structure of task graphs and identify versions and generations as key concepts for efficient management of task graphs. Then, we present three schemes to manage task graphs building on graph representations, hypergraphs and lists. We also consider a fourth edge-less scheme that synchronizes tasks using integers. Analysis using micro-benchmarks shows that the graph representation is not always scalable and that the edge-less scheme introduces least overhead in nearly all situations.

Pool Evolution: A Parallel Pattern for Evolutionary and Symbolic Computing

We introduce a new parallel pattern derived from a specific application domain and show how it turns out to have application beyond its domain of origin. The pool evolution pattern models the parallel evolution of a population subject to mutations and evolving in such a way that a given fitness function is optimized. The pattern has been demonstrated to be suitable for capturing and modeling the parallel patterns underpinning various evolutionary algorithms, as well as other parallel patterns typical of symbolic computation. In this paper we introduce the pattern, we discuss its implementation on modern multi/many core architectures and finally present experimental results obtained with FastFlow and Erlang implementations to assess its feasibility and scalability.

A Framework supporting collaboration on the distributed design of integrated systems

The work described in this thesis aims to support the distributed design of integrated systems and considers specifically the need for collaborative interaction among designers. Particular emphasis was given to issues which were only marginally considered in previous approaches, such as the abstraction of the distribution of design automation resources over the network, the possibility of both synchronous and asynchronous interaction among designers and the support for extensible design data models. Such issues demand a rather complex software infrastructure, as possible solutions must encompass a wide range of software modules: from user interfaces to middleware to databases. To build such structure, several engineering techniques were employed and some original solutions were devised. The core of the proposed solution is based in the joint application of two homonymic technologies: CAD Frameworks and object-oriented frameworks. The former concept was coined in the late 80's within the electronic design automation community and comprehends a layered software environment which aims to support CAD tool developers, CAD administrators/integrators and designers. The latter, developed during the last decade by the software engineering community, is a software architecture model to build extensible and reusable object-oriented software subsystems. In this work, we proposed to create an object-oriented framework which includes extensible sets of design data primitives and design tool building blocks. Such object-oriented framework is included within a CAD Framework, where it plays important roles on typical CAD Framework services such as design data representation and management, versioning, user interfaces, design management and tool integration. The implemented CAD Framework - named Cave2 - followed the classical layered architecture presented by Barnes, Harrison, Newton and Spickelmier, but the possibilities granted by the use of the object-oriented framework foundations allowed a series of improvements which were not available in previous approaches: - object-oriented frameworks are extensible by design, thus this should be also true regarding the implemented sets of design data primitives and design tool building blocks. This means that both the design representation model and the software modules dealing with it can be upgraded or adapted to a particular design methodology, and that such extensions and adaptations will still inherit the architectural and functional aspects implemented in the object-oriented framework foundation; - the design semantics and the design visualization are both part of the object-oriented framework, but in clearly separated models. This allows for different visualization strategies for a given design data set, which gives collaborating parties the flexibility to choose individual visualization settings; - the control of the consistency between semantics and visualization - a particularly important issue in a design environment with multiple views of a single design - is also included in the foundations of the object-oriented framework. Such mechanism is generic enough to be also used by further extensions of the design data model, as it is based on the inversion of control between view and semantics. The view receives the user input and propagates such event to the semantic model, which evaluates if a state change is possible. If positive, it triggers the change of state of both semantics and view. Our approach took advantage of such inversion of control and included an layer between semantics and view to take into account the possibility of multi-view consistency; - to optimize the consistency control mechanism between views and semantics, we propose an event-based approach that captures each discrete interaction of a designer with his/her respective design views. The information about each interaction is encapsulated inside an event object, which may be propagated to the design semantics - and thus to other possible views - according to the consistency policy which is being used. Furthermore, the use of event pools allows for a late synchronization between view and semantics in case of unavailability of a network connection between them; - the use of proxy objects raised significantly the abstraction of the integration of design automation resources, as either remote or local tools and services are accessed through method calls in a local object. The connection to remote tools and services using a look-up protocol also abstracted completely the network location of such resources, allowing for resource addition and removal during runtime; - the implemented CAD Framework is completely based on Java technology, so it relies on the Java Virtual Machine as the layer which grants the independence between the CAD Framework and the operating system. All such improvements contributed to a higher abstraction on the distribution of design automation resources and also introduced a new paradigm for the remote interaction between designers. The resulting CAD Framework is able to support fine-grained collaboration based on events, so every single design update performed by a designer can be propagated to the rest of the design team regardless of their location in the distributed environment. This can increase the group awareness and allow a richer transfer of experiences among them, improving significantly the collaboration potential when compared to previously proposed file-based or record-based approaches. Three different case studies were conducted to validate the proposed approach, each one focusing one a subset of the contributions of this thesis. The first one uses the proxy-based resource distribution architecture to implement a prototyping platform using reconfigurable hardware modules. The second one extends the foundations of the implemented object-oriented framework to support interface-based design. Such extensions - design representation primitives and tool blocks - are used to implement a design entry tool named IBlaDe, which allows the collaborative creation of functional and structural models of integrated systems. The third case study regards the possibility of integration of multimedia metadata to the design data model. Such possibility is explored in the frame of an online educational and training platform.