SHEPARD: Scheduling on heterogeneous platforms using application resource demands

Heterogeneous computing technologies, such as multi-core CPUs, GPUs and FPGAs can provide significant performance improvements. However, developing applications for these technologies often results in coupling applications to specific devices, typically through the use of proprietary tools. This paper presents SHEPARD, a compile time and run-time framework that decouples application development from the target platform and enables run-time allocation of tasks to heterogeneous computing devices. Through the use of special annotated functions, called managed tasks, SHEPARD approximates a task's performance on available devices, and coupled with the approximation of current device demand, decides which device can satisfy the task with the lowest overall execution time. Experiments using a task parallel application, based on an in-memory database, demonstrate the opportunity for automatic run-time task allocation to achieve speed-up over a static allocation to a single specific device. © 2014 IEEE.

Enhancing design space exploration by extending CPU/GPU specifications onto FPGAs

The design cycle for complex special-purpose computing systems is extremely costly and time-consuming. It involves a multiparametric design space exploration for optimization, followed by design verification. Designers of special purpose VLSI implementations often need to explore parameters, such as optimal bitwidth and data representation, through time-consuming Monte Carlo simulations. A prominent example of this simulation-based exploration process is the design of decoders for error correcting systems, such as the Low-Density Parity-Check (LDPC) codes adopted by modern communication standards, which involves thousands of Monte Carlo runs for each design point. Currently, high-performance computing offers a wide set of acceleration options that range from multicore CPUs to Graphics Processing Units (GPUs) and Field Programmable Gate Arrays (FPGAs). The exploitation of diverse target architectures is typically associated with developing multiple code versions, often using distinct programming paradigms. In this context, we evaluate the concept of retargeting a single OpenCL program to multiple platforms, thereby significantly reducing design time. A single OpenCL-based parallel kernel is used without modifications or code tuning on multicore CPUs, GPUs, and FPGAs. We use SOpenCL (Silicon to OpenCL), a tool that automatically converts OpenCL kernels to RTL in order to introduce FPGAs as a potential platform to efficiently execute simulations coded in OpenCL. We use LDPC decoding simulations as a case study. Experimental results were obtained by testing a variety of regular and irregular LDPC codes that range from short/medium (e.g., 8,000 bit) to long length (e.g., 64,800 bit) DVB-S2 codes. We observe that, depending on the design parameters to be simulated, on the dimension and phase of the design, the GPU or FPGA may suit different purposes more conveniently, thus providing different acceleration factors over conventional multicore CPUs.

Runtime Support for Adaptive Power Capping on Heterogeneous SoCs

Power capping is a fundamental method for reducing the energy consumption of a wide range of modern computing environments, ranging from mobile embedded systems to datacentres. Unfortunately, maximising performance and system efficiency under static power caps remains challenging, while maximising performance under dynamic power caps has been largely unexplored. We present an adaptive power capping method that reduces the power consumption and maximizes the performance of heterogeneous SoCs for mobile and server platforms. Our technique combines power capping with coordinated DVFS, data partitioning and core allocations on a heterogeneous SoC with ARM processors and FPGA resources. We design our framework as a run-time system based on OpenMP and OpenCL to utilise the heterogeneous resources. We evaluate it through five data-parallel benchmarks on the Xilinx SoC which allows fully voltage and frequency control. Our experiments show a significant performance boost of 30% under dynamic power caps with concurrent execution on ARM and FPGA, compared to a naive separate approach.

FAMOUS, faster: using parallel computing techniques to accelerate the FAMOUS/HadCM3 climate model with a focus on the radiative transfer algorithm

We have optimised the atmospheric radiation algorithm of the FAMOUS climate model on several hardware platforms. The optimisation involved translating the Fortran code to C and restructuring the algorithm around the computation of a single air column. Instead of the existing MPI-based domain decomposition, we used a task queue and a thread pool to schedule the computation of individual columns on the available processors. Finally, four air columns are packed together in a single data structure and computed simultaneously using Single Instruction Multiple Data operations. The modified algorithm runs more than 50 times faster on the CELL’s Synergistic Processing Elements than on its main PowerPC processing element. On Intel-compatible processors, the new radiation code runs 4 times faster. On the tested graphics processor, using OpenCL, we find a speed-up of more than 2.5 times as compared to the original code on the main CPU. Because the radiation code takes more than 60% of the total CPU time, FAMOUS executes more than twice as fast. Our version of the algorithm returns bit-wise identical results, which demonstrates the robustness of our approach. We estimate that this project required around two and a half man-years of work.

Image Re-Ranking Acceleration on GPUs

Huge image collections are becoming available lately. In this scenario, the use of Content-Based Image Retrieval (CBIR) systems has emerged as a promising approach to support image searches. The objective of CBIR systems is to retrieve the most similar images in a collection, given a query image, by taking into account image visual properties such as texture, color, and shape. In these systems, the effectiveness of the retrieval process depends heavily on the accuracy of ranking approaches. Recently, re-ranking approaches have been proposed to improve the effectiveness of CBIR systems by taking into account the relationships among images. The re-ranking approaches consider the relationships among all images in a given dataset. These approaches typically demands a huge amount of computational power, which hampers its use in practical situations. On the other hand, these methods can be massively parallelized. In this paper, we propose to speedup the computation of the RL-Sim algorithm, a recently proposed image re-ranking approach, by using the computational power of Graphics Processing Units (GPU). GPUs are emerging as relatively inexpensive parallel processors that are becoming available on a wide range of computer systems. We address the image re-ranking performance challenges by proposing a parallel solution designed to fit the computational model of GPUs. We conducted an experimental evaluation considering different implementations and devices. Experimental results demonstrate that significant performance gains can be obtained. Our approach achieves speedups of 7x from serial implementation considering the overall algorithm and up to 36x on its core steps.

General-Purpose Computing on Graphic Processing Unit (GPGPU): metodi e tecnologie per lo sviluppo di applicazioni

Microprocessori basati su singolo processore (CPU), hanno visto una rapida crescita di performances ed un abbattimento dei costi per circa venti anni. Questi microprocessori hanno portato una potenza di calcolo nell’ordine del GFLOPS (Giga Floating Point Operation per Second) sui PC Desktop e centinaia di GFLOPS su clusters di server. Questa ascesa ha portato nuove funzionalità nei programmi, migliori interfacce utente e tanti altri vantaggi. Tuttavia questa crescita ha subito un brusco rallentamento nel 2003 a causa di consumi energetici sempre più elevati e problemi di dissipazione termica, che hanno impedito incrementi di frequenza di clock. I limiti fisici del silicio erano sempre più vicini. Per ovviare al problema i produttori di CPU (Central Processing Unit) hanno iniziato a progettare microprocessori multicore, scelta che ha avuto un impatto notevole sulla comunità degli sviluppatori, abituati a considerare il software come una serie di comandi sequenziali. Quindi i programmi che avevano sempre giovato di miglioramenti di prestazioni ad ogni nuova generazione di CPU, non hanno avuto incrementi di performance, in quanto essendo eseguiti su un solo core, non beneficiavano dell’intera potenza della CPU. Per sfruttare appieno la potenza delle nuove CPU la programmazione concorrente, precedentemente utilizzata solo su sistemi costosi o supercomputers, è diventata una pratica sempre più utilizzata dagli sviluppatori. Allo stesso tempo, l’industria videoludica ha conquistato una fetta di mercato notevole: solo nel 2013 verranno spesi quasi 100 miliardi di dollari fra hardware e software dedicati al gaming. Le software houses impegnate nello sviluppo di videogames, per rendere i loro titoli più accattivanti, puntano su motori grafici sempre più potenti e spesso scarsamente ottimizzati, rendendoli estremamente esosi in termini di performance. Per questo motivo i produttori di GPU (Graphic Processing Unit), specialmente nell’ultimo decennio, hanno dato vita ad una vera e propria rincorsa alle performances che li ha portati ad ottenere dei prodotti con capacità di calcolo vertiginose. Ma al contrario delle CPU che agli inizi del 2000 intrapresero la strada del multicore per continuare a favorire programmi sequenziali, le GPU sono diventate manycore, ovvero con centinaia e centinaia di piccoli cores che eseguono calcoli in parallelo. Questa immensa capacità di calcolo può essere utilizzata in altri campi applicativi? La risposta è si e l’obiettivo di questa tesi è proprio quello di constatare allo stato attuale, in che modo e con quale efficienza pùo un software generico, avvalersi dell’utilizzo della GPU invece della CPU.

Gpu-based many-core architecture emulation: A double level approach

The efficient emulation of a many-core architecture is a challenging task, each core could be emulated through a dedicated thread and such threads would be interleaved on an either single-core or a multi-core processor. The high number of context switches will results in an unacceptable performance. To support this kind of application, the GPU computational power is exploited in order to schedule the emulation threads on the GPU cores. This presents a non trivial divergence issue, since GPU computational power is offered through SIMD processing elements, that are forced to synchronously execute the same instruction on different memory portions. Thus, a new emulation technique is introduced in order to overcome this limitation: instead of providing a routine for each ISA opcode, the emulator mimics the behavior of the Micro Architecture level, here instructions are date that a unique routine takes as input. Our new technique has been implemented and compared with the classic emulation approach, in order to investigate the chance of a hybrid solution.

Run-Time Dynamically-Adaptable FPGA-Based Architecture for High-Performance Autonomous Distributed Systems

Esta tesis doctoral se enmarca dentro del campo de los sistemas embebidos reconfigurables, redes de sensores inalámbricas para aplicaciones de altas prestaciones, y computación distribuida. El documento se centra en el estudio de alternativas de procesamiento para sistemas embebidos autónomos distribuidos de altas prestaciones (por sus siglas en inglés, High-Performance Autonomous Distributed Systems (HPADS)), así como su evolución hacia el procesamiento de alta resolución. El estudio se ha llevado a cabo tanto a nivel de plataforma como a nivel de las arquitecturas de procesamiento dentro de la plataforma con el objetivo de optimizar aspectos tan relevantes como la eficiencia energética, la capacidad de cómputo y la tolerancia a fallos del sistema. Los HPADS son sistemas realimentados, normalmente formados por elementos distribuidos conectados o no en red, con cierta capacidad de adaptación, y con inteligencia suficiente para llevar a cabo labores de prognosis y/o autoevaluación. Esta clase de sistemas suele formar parte de sistemas más complejos llamados sistemas ciber-físicos (por sus siglas en inglés, Cyber-Physical Systems (CPSs)). Los CPSs cubren un espectro enorme de aplicaciones, yendo desde aplicaciones médicas, fabricación, o aplicaciones aeroespaciales, entre otras muchas. Para el diseño de este tipo de sistemas, aspectos tales como la confiabilidad, la definición de modelos de computación, o el uso de metodologías y/o herramientas que faciliten el incremento de la escalabilidad y de la gestión de la complejidad, son fundamentales. La primera parte de esta tesis doctoral se centra en el estudio de aquellas plataformas existentes en el estado del arte que por sus características pueden ser aplicables en el campo de los CPSs, así como en la propuesta de un nuevo diseño de plataforma de altas prestaciones que se ajuste mejor a los nuevos y más exigentes requisitos de las nuevas aplicaciones. Esta primera parte incluye descripción, implementación y validación de la plataforma propuesta, así como conclusiones sobre su usabilidad y sus limitaciones. Los principales objetivos para el diseño de la plataforma propuesta se enumeran a continuación: • Estudiar la viabilidad del uso de una FPGA basada en RAM como principal procesador de la plataforma en cuanto a consumo energético y capacidad de cómputo. • Propuesta de técnicas de gestión del consumo de energía en cada etapa del perfil de trabajo de la plataforma. •Propuestas para la inclusión de reconfiguración dinámica y parcial de la FPGA (por sus siglas en inglés, Dynamic Partial Reconfiguration (DPR)) de forma que sea posible cambiar ciertas partes del sistema en tiempo de ejecución y sin necesidad de interrumpir al resto de las partes. Evaluar su aplicabilidad en el caso de HPADS. Las nuevas aplicaciones y nuevos escenarios a los que se enfrentan los CPSs, imponen nuevos requisitos en cuanto al ancho de banda necesario para el procesamiento de los datos, así como en la adquisición y comunicación de los mismos, además de un claro incremento en la complejidad de los algoritmos empleados. Para poder cumplir con estos nuevos requisitos, las plataformas están migrando desde sistemas tradicionales uni-procesador de 8 bits, a sistemas híbridos hardware-software que incluyen varios procesadores, o varios procesadores y lógica programable. Entre estas nuevas arquitecturas, las FPGAs y los sistemas en chip (por sus siglas en inglés, System on Chip (SoC)) que incluyen procesadores embebidos y lógica programable, proporcionan soluciones con muy buenos resultados en cuanto a consumo energético, precio, capacidad de cómputo y flexibilidad. Estos buenos resultados son aún mejores cuando las aplicaciones tienen altos requisitos de cómputo y cuando las condiciones de trabajo son muy susceptibles de cambiar en tiempo real. La plataforma propuesta en esta tesis doctoral se ha denominado HiReCookie. La arquitectura incluye una FPGA basada en RAM como único procesador, así como un diseño compatible con la plataforma para redes de sensores inalámbricas desarrollada en el Centro de Electrónica Industrial de la Universidad Politécnica de Madrid (CEI-UPM) conocida como Cookies. Esta FPGA, modelo Spartan-6 LX150, era, en el momento de inicio de este trabajo, la mejor opción en cuanto a consumo y cantidad de recursos integrados, cuando además, permite el uso de reconfiguración dinámica y parcial. Es importante resaltar que aunque los valores de consumo son los mínimos para esta familia de componentes, la potencia instantánea consumida sigue siendo muy alta para aquellos sistemas que han de trabajar distribuidos, de forma autónoma, y en la mayoría de los casos alimentados por baterías. Por esta razón, es necesario incluir en el diseño estrategias de ahorro energético para incrementar la usabilidad y el tiempo de vida de la plataforma. La primera estrategia implementada consiste en dividir la plataforma en distintas islas de alimentación de forma que sólo aquellos elementos que sean estrictamente necesarios permanecerán alimentados, cuando el resto puede estar completamente apagado. De esta forma es posible combinar distintos modos de operación y así optimizar enormemente el consumo de energía. El hecho de apagar la FPGA para ahora energía durante los periodos de inactividad, supone la pérdida de la configuración, puesto que la memoria de configuración es una memoria volátil. Para reducir el impacto en el consumo y en el tiempo que supone la reconfiguración total de la plataforma una vez encendida, en este trabajo, se incluye una técnica para la compresión del archivo de configuración de la FPGA, de forma que se consiga una reducción del tiempo de configuración y por ende de la energía consumida. Aunque varios de los requisitos de diseño pueden satisfacerse con el diseño de la plataforma HiReCookie, es necesario seguir optimizando diversos parámetros tales como el consumo energético, la tolerancia a fallos y la capacidad de procesamiento. Esto sólo es posible explotando todas las posibilidades ofrecidas por la arquitectura de procesamiento en la FPGA. Por lo tanto, la segunda parte de esta tesis doctoral está centrada en el diseño de una arquitectura reconfigurable denominada ARTICo3 (Arquitectura Reconfigurable para el Tratamiento Inteligente de Cómputo, Confiabilidad y Consumo de energía) para la mejora de estos parámetros por medio de un uso dinámico de recursos. ARTICo3 es una arquitectura de procesamiento para FPGAs basadas en RAM, con comunicación tipo bus, preparada para dar soporte para la gestión dinámica de los recursos internos de la FPGA en tiempo de ejecución gracias a la inclusión de reconfiguración dinámica y parcial. Gracias a esta capacidad de reconfiguración parcial, es posible adaptar los niveles de capacidad de procesamiento, energía consumida o tolerancia a fallos para responder a las demandas de la aplicación, entorno, o métricas internas del dispositivo mediante la adaptación del número de recursos asignados para cada tarea. Durante esta segunda parte de la tesis se detallan el diseño de la arquitectura, su implementación en la plataforma HiReCookie, así como en otra familia de FPGAs, y su validación por medio de diferentes pruebas y demostraciones. Los principales objetivos que se plantean la arquitectura son los siguientes: • Proponer una metodología basada en un enfoque multi-hilo, como las propuestas por CUDA (por sus siglas en inglés, Compute Unified Device Architecture) u Open CL, en la cual distintos kernels, o unidades de ejecución, se ejecuten en un numero variable de aceleradores hardware sin necesidad de cambios en el código de aplicación. • Proponer un diseño y proporcionar una arquitectura en la que las condiciones de trabajo cambien de forma dinámica dependiendo bien de parámetros externos o bien de parámetros que indiquen el estado de la plataforma. Estos cambios en el punto de trabajo de la arquitectura serán posibles gracias a la reconfiguración dinámica y parcial de aceleradores hardware en tiempo real. • Explotar las posibilidades de procesamiento concurrente, incluso en una arquitectura basada en bus, por medio de la optimización de las transacciones en ráfaga de datos hacia los aceleradores. •Aprovechar las ventajas ofrecidas por la aceleración lograda por módulos puramente hardware para conseguir una mejor eficiencia energética. • Ser capaces de cambiar los niveles de redundancia de hardware de forma dinámica según las necesidades del sistema en tiempo real y sin cambios para el código de aplicación. • Proponer una capa de abstracción entre el código de aplicación y el uso dinámico de los recursos de la FPGA. El diseño en FPGAs permite la utilización de módulos hardware específicamente creados para una aplicación concreta. De esta forma es posible obtener rendimientos mucho mayores que en el caso de las arquitecturas de propósito general. Además, algunas FPGAs permiten la reconfiguración dinámica y parcial de ciertas partes de su lógica en tiempo de ejecución, lo cual dota al diseño de una gran flexibilidad. Los fabricantes de FPGAs ofrecen arquitecturas predefinidas con la posibilidad de añadir bloques prediseñados y poder formar sistemas en chip de una forma más o menos directa. Sin embargo, la forma en la que estos módulos hardware están organizados dentro de la arquitectura interna ya sea estática o dinámicamente, o la forma en la que la información se intercambia entre ellos, influye enormemente en la capacidad de cómputo y eficiencia energética del sistema. De la misma forma, la capacidad de cargar módulos hardware bajo demanda, permite añadir bloques redundantes que permitan aumentar el nivel de tolerancia a fallos de los sistemas. Sin embargo, la complejidad ligada al diseño de bloques hardware dedicados no debe ser subestimada. Es necesario tener en cuenta que el diseño de un bloque hardware no es sólo su propio diseño, sino también el diseño de sus interfaces, y en algunos casos de los drivers software para su manejo. Además, al añadir más bloques, el espacio de diseño se hace más complejo, y su programación más difícil. Aunque la mayoría de los fabricantes ofrecen interfaces predefinidas, IPs (por sus siglas en inglés, Intelectual Property) comerciales y plantillas para ayudar al diseño de los sistemas, para ser capaces de explotar las posibilidades reales del sistema, es necesario construir arquitecturas sobre las ya establecidas para facilitar el uso del paralelismo, la redundancia, y proporcionar un entorno que soporte la gestión dinámica de los recursos. Para proporcionar este tipo de soporte, ARTICo3 trabaja con un espacio de soluciones formado por tres ejes fundamentales: computación, consumo energético y confiabilidad. De esta forma, cada punto de trabajo se obtiene como una solución de compromiso entre estos tres parámetros. Mediante el uso de la reconfiguración dinámica y parcial y una mejora en la transmisión de los datos entre la memoria principal y los aceleradores, es posible dedicar un número variable de recursos en el tiempo para cada tarea, lo que hace que los recursos internos de la FPGA sean virtualmente ilimitados. Este variación en el tiempo del número de recursos por tarea se puede usar bien para incrementar el nivel de paralelismo, y por ende de aceleración, o bien para aumentar la redundancia, y por lo tanto el nivel de tolerancia a fallos. Al mismo tiempo, usar un numero óptimo de recursos para una tarea mejora el consumo energético ya que bien es posible disminuir la potencia instantánea consumida, o bien el tiempo de procesamiento. Con el objetivo de mantener los niveles de complejidad dentro de unos límites lógicos, es importante que los cambios realizados en el hardware sean totalmente transparentes para el código de aplicación. A este respecto, se incluyen distintos niveles de transparencia: • Transparencia a la escalabilidad: los recursos usados por una misma tarea pueden ser modificados sin que el código de aplicación sufra ningún cambio. • Transparencia al rendimiento: el sistema aumentara su rendimiento cuando la carga de trabajo aumente, sin cambios en el código de aplicación. • Transparencia a la replicación: es posible usar múltiples instancias de un mismo módulo bien para añadir redundancia o bien para incrementar la capacidad de procesamiento. Todo ello sin que el código de aplicación cambie. • Transparencia a la posición: la posición física de los módulos hardware es arbitraria para su direccionamiento desde el código de aplicación. • Transparencia a los fallos: si existe un fallo en un módulo hardware, gracias a la redundancia, el código de aplicación tomará directamente el resultado correcto. • Transparencia a la concurrencia: el hecho de que una tarea sea realizada por más o menos bloques es transparente para el código que la invoca. Por lo tanto, esta tesis doctoral contribuye en dos líneas diferentes. En primer lugar, con el diseño de la plataforma HiReCookie y en segundo lugar con el diseño de la arquitectura ARTICo3. Las principales contribuciones de esta tesis se resumen a continuación. • Arquitectura de la HiReCookie incluyendo: o Compatibilidad con la plataforma Cookies para incrementar las capacidades de esta. o División de la arquitectura en distintas islas de alimentación. o Implementación de los diversos modos de bajo consumo y políticas de despertado del nodo. o Creación de un archivo de configuración de la FPGA comprimido para reducir el tiempo y el consumo de la configuración inicial. • Diseño de la arquitectura reconfigurable para FPGAs basadas en RAM ARTICo3: o Modelo de computación y modos de ejecución inspirados en el modelo de CUDA pero basados en hardware reconfigurable con un número variable de bloques de hilos por cada unidad de ejecución. o Estructura para optimizar las transacciones de datos en ráfaga proporcionando datos en cascada o en paralelo a los distinto módulos incluyendo un proceso de votado por mayoría y operaciones de reducción. o Capa de abstracción entre el procesador principal que incluye el código de aplicación y los recursos asignados para las diferentes tareas. o Arquitectura de los módulos hardware reconfigurables para mantener la escalabilidad añadiendo una la interfaz para las nuevas funcionalidades con un simple acceso a una memoria RAM interna. o Caracterización online de las tareas para proporcionar información a un módulo de gestión de recursos para mejorar la operación en términos de energía y procesamiento cuando además se opera entre distintos nieles de tolerancia a fallos. El documento está dividido en dos partes principales formando un total de cinco capítulos. En primer lugar, después de motivar la necesidad de nuevas plataformas para cubrir las nuevas aplicaciones, se detalla el diseño de la plataforma HiReCookie, sus partes, las posibilidades para bajar el consumo energético y se muestran casos de uso de la plataforma así como pruebas de validación del diseño. La segunda parte del documento describe la arquitectura reconfigurable, su implementación en varias FPGAs, y pruebas de validación en términos de capacidad de procesamiento y consumo energético, incluyendo cómo estos aspectos se ven afectados por el nivel de tolerancia a fallos elegido. Los capítulos a lo largo del documento son los siguientes: El capítulo 1 analiza los principales objetivos, motivación y aspectos teóricos necesarios para seguir el resto del documento. El capítulo 2 está centrado en el diseño de la plataforma HiReCookie y sus posibilidades para disminuir el consumo de energía. El capítulo 3 describe la arquitectura reconfigurable ARTICo3. El capítulo 4 se centra en las pruebas de validación de la arquitectura usando la plataforma HiReCookie para la mayoría de los tests. Un ejemplo de aplicación es mostrado para analizar el funcionamiento de la arquitectura. El capítulo 5 concluye esta tesis doctoral comentando las conclusiones obtenidas, las contribuciones originales del trabajo y resultados y líneas futuras. ABSTRACT This PhD Thesis is framed within the field of dynamically reconfigurable embedded systems, advanced sensor networks and distributed computing. The document is centred on the study of processing solutions for high-performance autonomous distributed systems (HPADS) as well as their evolution towards High performance Computing (HPC) systems. The approach of the study is focused on both platform and processor levels to optimise critical aspects such as computing performance, energy efficiency and fault tolerance. HPADS are considered feedback systems, normally networked and/or distributed, with real-time adaptive and predictive functionality. These systems, as part of more complex systems known as Cyber-Physical Systems (CPSs), can be applied in a wide range of fields such as military, health care, manufacturing, aerospace, etc. For the design of HPADS, high levels of dependability, the definition of suitable models of computation, and the use of methodologies and tools to support scalability and complexity management, are required. The first part of the document studies the different possibilities at platform design level in the state of the art, together with description, development and validation tests of the platform proposed in this work to cope with the previously mentioned requirements. The main objectives targeted by this platform design are the following: • Study the feasibility of using SRAM-based FPGAs as the main processor of the platform in terms of energy consumption and performance for high demanding applications. • Analyse and propose energy management techniques to reduce energy consumption in every stage of the working profile of the platform. • Provide a solution with dynamic partial and wireless remote HW reconfiguration (DPR) to be able to change certain parts of the FPGA design at run time and on demand without interrupting the rest of the system. • Demonstrate the applicability of the platform in different test-bench applications. In order to select the best approach for the platform design in terms of processing alternatives, a study of the evolution of the state-of-the-art platforms is required to analyse how different architectures cope with new more demanding applications and scenarios: security, mixed-critical systems for aerospace, multimedia applications, or military environments, among others. In all these scenarios, important changes in the required processing bandwidth or the complexity of the algorithms used are provoking the migration of the platforms from single microprocessor architectures to multiprocessing and heterogeneous solutions with more instant power consumption but higher energy efficiency. Within these solutions, FPGAs and Systems on Chip including FPGA fabric and dedicated hard processors, offer a good trade of among flexibility, processing performance, energy consumption and price, when they are used in demanding applications where working conditions are very likely to vary over time and high complex algorithms are required. The platform architecture proposed in this PhD Thesis is called HiReCookie. It includes an SRAM-based FPGA as the main and only processing unit. The FPGA selected, the Xilinx Spartan-6 LX150, was at the beginning of this work the best choice in terms of amount of resources and power. Although, the power levels are the lowest of these kind of devices, they can be still very high for distributed systems that normally work powered by batteries. For that reason, it is necessary to include different energy saving possibilities to increase the usability of the platform. In order to reduce energy consumption, the platform architecture is divided into different power islands so that only those parts of the systems that are strictly needed are powered on, while the rest of the islands can be completely switched off. This allows a combination of different low power modes to decrease energy. In addition, one of the most important handicaps of SRAM-based FPGAs is that they are not alive at power up. Therefore, recovering the system from a switch-off state requires to reload the FPGA configuration from a non-volatile memory device. For that reason, this PhD Thesis also proposes a methodology to compress the FPGA configuration file in order to reduce time and energy during the initial configuration process. Although some of the requirements for the design of HPADS are already covered by the design of the HiReCookie platform, it is necessary to continue improving energy efficiency, computing performance and fault tolerance. This is only possible by exploiting all the opportunities provided by the processing architectures configured inside the FPGA. Therefore, the second part of the thesis details the design of the so called ARTICo3 FPGA architecture to enhance the already intrinsic capabilities of the FPGA. ARTICo3 is a DPR-capable bus-based virtual architecture for multiple HW acceleration in SRAM-based FPGAs. The architecture provides support for dynamic resource management in real time. In this way, by using DPR, it will be possible to change the levels of computing performance, energy consumption and fault tolerance on demand by increasing or decreasing the amount of resources used by the different tasks. Apart from the detailed design of the architecture and its implementation in different FPGA devices, different validation tests and comparisons are also shown. The main objectives targeted by this FPGA architecture are listed as follows: • Provide a method based on a multithread approach such as those offered by CUDA (Compute Unified Device Architecture) or OpenCL kernel executions, where kernels are executed in a variable number of HW accelerators without requiring application code changes. • Provide an architecture to dynamically adapt working points according to either self-measured or external parameters in terms of energy consumption, fault tolerance and computing performance. Taking advantage of DPR capabilities, the architecture must provide support for a dynamic use of resources in real time. • Exploit concurrent processing capabilities in a standard bus-based system by optimizing data transactions to and from HW accelerators. • Measure the advantage of HW acceleration as a technique to boost performance to improve processing times and save energy by reducing active times for distributed embedded systems. • Dynamically change the levels of HW redundancy to adapt fault tolerance in real time. • Provide HW abstraction from SW application design. FPGAs give the possibility of designing specific HW blocks for every required task to optimise performance while some of them include the possibility of including DPR. Apart from the possibilities provided by manufacturers, the way these HW modules are organised, addressed and multiplexed in area and time can improve computing performance and energy consumption. At the same time, fault tolerance and security techniques can also be dynamically included using DPR. However, the inherent complexity of designing new HW modules for every application is not negligible. It does not only consist of the HW description, but also the design of drivers and interfaces with the rest of the system, while the design space is widened and more complex to define and program. Even though the tools provided by the majority of manufacturers already include predefined bus interfaces, commercial IPs, and templates to ease application prototyping, it is necessary to improve these capabilities. By adding new architectures on top of them, it is possible to take advantage of parallelization and HW redundancy while providing a framework to ease the use of dynamic resource management. ARTICo3 works within a solution space where working points change at run time in a 3D space defined by three different axes: Computation, Consumption, and Fault Tolerance. Therefore, every working point is found as a trade-off solution among these three axes. By means of DPR, different accelerators can be multiplexed so that the amount of available resources for any application is virtually unlimited. Taking advantage of DPR capabilities and a novel way of transmitting data to the reconfigurable HW accelerators, it is possible to dedicate a dynamically-changing number of resources for a given task in order to either boost computing speed or adding HW redundancy and a voting process to increase fault-tolerance levels. At the same time, using an optimised amount of resources for a given task reduces energy consumption by reducing instant power or computing time. In order to keep level complexity under certain limits, it is important that HW changes are transparent for the application code. Therefore, different levels of transparency are targeted by the system: • Scalability transparency: a task must be able to expand its resources without changing the system structure or application algorithms. • Performance transparency: the system must reconfigure itself as load changes. • Replication transparency: multiple instances of the same task are loaded to increase reliability and performance. • Location transparency: resources are accessed with no knowledge of their location by the application code. • Failure transparency: task must be completed despite a failure in some components. • Concurrency transparency: different tasks will work in a concurrent way transparent to the application code. Therefore, as it can be seen, the Thesis is contributing in two different ways. First with the design of the HiReCookie platform and, second with the design of the ARTICo3 architecture. The main contributions of this PhD Thesis are then listed below: • Architecture of the HiReCookie platform including: o Compatibility of the processing layer for high performance applications with the Cookies Wireless Sensor Network platform for fast prototyping and implementation. o A division of the architecture in power islands. o All the different low-power modes. o The creation of the partial-initial bitstream together with the wake-up policies of the node. • The design of the reconfigurable architecture for SRAM FPGAs: ARTICo3: o A model of computation and execution modes inspired in CUDA but based on reconfigurable HW with a dynamic number of thread blocks per kernel. o A structure to optimise burst data transactions providing coalesced or parallel data to HW accelerators, parallel voting process and reduction operation. o The abstraction provided to the host processor with respect to the operation of the kernels in terms of the number of replicas, modes of operation, location in the reconfigurable area and addressing. o The architecture of the modules representing the thread blocks to make the system scalable by adding functional units only adding an access to a BRAM port. o The online characterization of the kernels to provide information to a scheduler or resource manager in terms of energy consumption and processing time when changing among different fault-tolerance levels, as well as if a kernel is expected to work in the memory-bounded or computing-bounded areas. The document of the Thesis is divided into two main parts with a total of five chapters. First, after motivating the need for new platforms to cover new more demanding applications, the design of the HiReCookie platform, its parts and several partial tests are detailed. The design of the platform alone does not cover all the needs of these applications. Therefore, the second part describes the architecture inside the FPGA, called ARTICo3, proposed in this PhD Thesis. The architecture and its implementation are tested in terms of energy consumption and computing performance showing different possibilities to improve fault tolerance and how this impact in energy and time of processing. Chapter 1 shows the main goals of this PhD Thesis and the technology background required to follow the rest of the document. Chapter 2 shows all the details about the design of the FPGA-based platform HiReCookie. Chapter 3 describes the ARTICo3 architecture. Chapter 4 is focused on the validation tests of the ARTICo3 architecture. An application for proof of concept is explained where typical kernels related to image processing and encryption algorithms are used. Further experimental analyses are performed using these kernels. Chapter 5 concludes the document analysing conclusions, comments about the contributions of the work, and some possible future lines for the work.

A Kinetic Vlasov Model for Plasma Simulation Using Discontinuous Galerkin Method on Many-Core Architectures

Thesis (Ph.D.)--University of Washington, 2016-03

An Environment for Rapid Derivatives Design and Experimentation

In the highly competitive world of modern finance, new derivatives are continually required to take advantage of changes in financial markets, and to hedge businesses against new risks. The research described in this paper aims to accelerate the development and pricing of new derivatives in two different ways. Firstly, new derivatives can be specified mathematically within a general framework, enabling new mathematical formulae to be specified rather than just new parameter settings. This Generic Pricing Engine (GPE) is expressively powerful enough to specify a wide range of stand¬ard pricing engines. Secondly, the associated price simulation using the Monte Carlo method is accelerated using GPU or multicore hardware. The parallel implementation (in OpenCL) is automatically derived from the mathematical description of the derivative. As a test, for a Basket Option Pricing Engine (BOPE) generated using the GPE, on the largest problem size, an NVidia GPU runs the generated pricing engine at 45 times the speed of a sequential, specific hand-coded implementation of the same BOPE. Thus a user can more rapidly devise, simulate and experiment with new derivatives without actual programming.

The Loop-of-Stencil-Reduce Paradigm

In this paper we advocate the Loop-of-stencil-reduce pattern as a way to simplify the parallel programming of heterogeneous platforms (multicore+GPUs). Loop-of-Stencil-reduce is general enough to subsume map, reduce, map-reduce, stencil, stencil-reduce, and, crucially, their usage in a loop. It transparently targets (by using OpenCL) combinations of CPU cores and GPUs, and it makes it possible to simplify the deployment of a single stencil computation kernel on different GPUs. The paper discusses the implementation of Loop-of-stencil-reduce within the FastFlow parallel framework, considering a simple iterative data-parallel application as running example (Game of Life) and a highly effective parallel filter for visual data restoration to assess performance. Thanks to the high-level design of the Loop-of-stencil-reduce, it was possible to run the filter seamlessly on a multicore machine, on multi-GPUs, and on both.

Assemblage d’ADN avec graphes de de Bruijn sur FPGA

Ce mémoire est consacré à la parallélisation d’un algorithme d’assemblage d’ADN de type de novo sur différentes plateformes matérielles, soit les processeurs multicoeurs et les accélérateurs de type FPGA. Plus précisément, le langage OpenCL est utilisé pour accélérer l’algorithme dont il est question, et de permettre un comparatif direct entre les les plateformes. Cet algorithme est d’abord introduit, puis son implémentation originale, développée pour une exécution sur une grappe de noeuds, est discutée. Les modifications apportées à l’algorithme dans le but de faciliter la parallélisation sont ensuite divulgées. Ensuite, le coeur du travail est présenté, soit la programmation utilisant OpenCL. Finalement, les résultats sont présentés et discutés.

Supporting scientists in re-engineering sequential programs to parallel using model-driven engineering

Developing complex computational-intensiveand data-intensive scientific applications requires effectiveutilization of the computational power of the availablecomputing platforms including grids, clouds, clusters, multicoreand many-core processors, and graphical processingunits (GPUs). However, scientists who need to leverage suchplatforms are usually not parallel or distributed programmingexperts. Thus, they face numerous challenges whenimplementing and porting their software-based experimentaltools to such platforms. In this paper, we introduce asequential-to-parallel engineering approach to help scientistsin engineering their scientific applications. Our approach isbased on capturing sequential program details, plannedparallelization aspects, and program deployment details usinga set of domain-specific visual languages (DSVLs). Then, usingcode generation, we generate the corresponding parallelprogram using necessary parallel and distributedprogramming models (MPI, OpenCL, or OpenMP). Wesummarize three case studies (matrix multiplication, N-Bodysimulation, and signal processing) to evaluate our approach.

Optimization of Pattern Matching Algorithms for Multi- and Many-Core Platforms

Image and video compression play a major role in the world today, allowing the storage and transmission of large multimedia content volumes. However, the processing of this information requires high computational resources, hence the improvement of the computational performance of these compression algorithms is very important. The Multidimensional Multiscale Parser (MMP) is a pattern-matching-based compression algorithm for multimedia contents, namely images, achieving high compression ratios, maintaining good image quality, Rodrigues et al. [2008]. However, in comparison with other existing algorithms, this algorithm takes some time to execute. Therefore, two parallel implementations for GPUs were proposed by Ribeiro [2016] and Silva [2015] in CUDA and OpenCL-GPU, respectively. In this dissertation, to complement the referred work, we propose two parallel versions that run the MMP algorithm in CPU: one resorting to OpenMP and another that converts the existing OpenCL-GPU into OpenCL-CPU. The proposed solutions are able to improve the computational performance of MMP by 3 and 2:7 , respectively. The High Efficiency Video Coding (HEVC/H.265) is the most recent standard for compression of image and video. Its impressive compression performance, makes it a target for many adaptations, particularly for holoscopic image/video processing (or light field). Some of the proposed modifications to encode this new multimedia content are based on geometry-based disparity compensations (SS), developed by Conti et al. [2014], and a Geometric Transformations (GT) module, proposed by Monteiro et al. [2015]. These compression algorithms for holoscopic images based on HEVC present an implementation of specific search for similar micro-images that is more efficient than the one performed by HEVC, but its implementation is considerably slower than HEVC. In order to enable better execution times, we choose to use the OpenCL API as the GPU enabling language in order to increase the module performance. With its most costly setting, we are able to reduce the GT module execution time from 6.9 days to less then 4 hours, effectively attaining a speedup of 45 .