Tecniche di ottimizzazione del software per sistemi su singolo chip per applicazioni di Nomadic Computing

I moderni sistemi embedded sono equipaggiati con risorse hardware che consentono l’esecuzione di applicazioni molto complesse come il decoding audio e video. La progettazione di simili sistemi deve soddisfare due esigenze opposte. Da un lato è necessario fornire un elevato potenziale computazionale, dall’altro bisogna rispettare dei vincoli stringenti riguardo il consumo di energia. Uno dei trend più diffusi per rispondere a queste esigenze opposte è quello di integrare su uno stesso chip un numero elevato di processori caratterizzati da un design semplificato e da bassi consumi. Tuttavia, per sfruttare effettivamente il potenziale computazionale offerto da una batteria di processoriè necessario rivisitare pesantemente le metodologie di sviluppo delle applicazioni. Con l’avvento dei sistemi multi-processore su singolo chip (MPSoC) il parallel programming si è diffuso largamente anche in ambito embedded. Tuttavia, i progressi nel campo della programmazione parallela non hanno mantenuto il passo con la capacità di integrare hardware parallelo su un singolo chip. Oltre all’introduzione di multipli processori, la necessità di ridurre i consumi degli MPSoC comporta altre soluzioni architetturali che hanno l’effetto diretto di complicare lo sviluppo delle applicazioni. Il design del sottosistema di memoria, in particolare, è un problema critico. Integrare sul chip dei banchi di memoria consente dei tempi d’accesso molto brevi e dei consumi molto contenuti. Sfortunatamente, la quantità di memoria on-chip che può essere integrata in un MPSoC è molto limitata. Per questo motivo è necessario aggiungere dei banchi di memoria off-chip, che hanno una capacità molto maggiore, come maggiori sono i consumi e i tempi d’accesso. La maggior parte degli MPSoC attualmente in commercio destina una parte del budget di area all’implementazione di memorie cache e/o scratchpad. Le scratchpad (SPM) sono spesso preferite alle cache nei sistemi MPSoC embedded, per motivi di maggiore predicibilità, minore occupazione d’area e – soprattutto – minori consumi. Per contro, mentre l’uso delle cache è completamente trasparente al programmatore, le SPM devono essere esplicitamente gestite dall’applicazione. Esporre l’organizzazione della gerarchia di memoria ll’applicazione consente di sfruttarne in maniera efficiente i vantaggi (ridotti tempi d’accesso e consumi). Per contro, per ottenere questi benefici è necessario scrivere le applicazioni in maniera tale che i dati vengano partizionati e allocati sulle varie memorie in maniera opportuna. L’onere di questo compito complesso ricade ovviamente sul programmatore. Questo scenario descrive bene l’esigenza di modelli di programmazione e strumenti di supporto che semplifichino lo sviluppo di applicazioni parallele. In questa tesi viene presentato un framework per lo sviluppo di software per MPSoC embedded basato su OpenMP. OpenMP è uno standard di fatto per la programmazione di multiprocessori con memoria shared, caratterizzato da un semplice approccio alla parallelizzazione tramite annotazioni (direttive per il compilatore). La sua interfaccia di programmazione consente di esprimere in maniera naturale e molto efficiente il parallelismo a livello di loop, molto diffuso tra le applicazioni embedded di tipo signal processing e multimedia. OpenMP costituisce un ottimo punto di partenza per la definizione di un modello di programmazione per MPSoC, soprattutto per la sua semplicità d’uso. D’altra parte, per sfruttare in maniera efficiente il potenziale computazionale di un MPSoC è necessario rivisitare profondamente l’implementazione del supporto OpenMP sia nel compilatore che nell’ambiente di supporto a runtime. Tutti i costrutti per gestire il parallelismo, la suddivisione del lavoro e la sincronizzazione inter-processore comportano un costo in termini di overhead che deve essere minimizzato per non comprometterre i vantaggi della parallelizzazione. Questo può essere ottenuto soltanto tramite una accurata analisi delle caratteristiche hardware e l’individuazione dei potenziali colli di bottiglia nell’architettura. Una implementazione del task management, della sincronizzazione a barriera e della condivisione dei dati che sfrutti efficientemente le risorse hardware consente di ottenere elevate performance e scalabilità. La condivisione dei dati, nel modello OpenMP, merita particolare attenzione. In un modello a memoria condivisa le strutture dati (array, matrici) accedute dal programma sono fisicamente allocate su una unica risorsa di memoria raggiungibile da tutti i processori. Al crescere del numero di processori in un sistema, l’accesso concorrente ad una singola risorsa di memoria costituisce un evidente collo di bottiglia. Per alleviare la pressione sulle memorie e sul sistema di connessione vengono da noi studiate e proposte delle tecniche di partizionamento delle strutture dati. Queste tecniche richiedono che una singola entità di tipo array venga trattata nel programma come l’insieme di tanti sotto-array, ciascuno dei quali può essere fisicamente allocato su una risorsa di memoria differente. Dal punto di vista del programma, indirizzare un array partizionato richiede che ad ogni accesso vengano eseguite delle istruzioni per ri-calcolare l’indirizzo fisico di destinazione. Questo è chiaramente un compito lungo, complesso e soggetto ad errori. Per questo motivo, le nostre tecniche di partizionamento sono state integrate nella l’interfaccia di programmazione di OpenMP, che è stata significativamente estesa. Specificamente, delle nuove direttive e clausole consentono al programmatore di annotare i dati di tipo array che si vuole partizionare e allocare in maniera distribuita sulla gerarchia di memoria. Sono stati inoltre sviluppati degli strumenti di supporto che consentono di raccogliere informazioni di profiling sul pattern di accesso agli array. Queste informazioni vengono sfruttate dal nostro compilatore per allocare le partizioni sulle varie risorse di memoria rispettando una relazione di affinità tra il task e i dati. Più precisamente, i passi di allocazione nel nostro compilatore assegnano una determinata partizione alla memoria scratchpad locale al processore che ospita il task che effettua il numero maggiore di accessi alla stessa.

Parallel Architectures for Many-Core Systems-On-Chip in Deep Sub-Micron Technology

Despite the several issues faced in the past, the evolutionary trend of silicon has kept its constant pace. Today an ever increasing number of cores is integrated onto the same die. Unfortunately, the extraordinary performance achievable by the many-core paradigm is limited by several factors. Memory bandwidth limitation, combined with inefficient synchronization mechanisms, can severely overcome the potential computation capabilities. Moreover, the huge HW/SW design space requires accurate and flexible tools to perform architectural explorations and validation of design choices. In this thesis we focus on the aforementioned aspects: a flexible and accurate Virtual Platform has been developed, targeting a reference many-core architecture. Such tool has been used to perform architectural explorations, focusing on instruction caching architecture and hybrid HW/SW synchronization mechanism. Beside architectural implications, another issue of embedded systems is considered: energy efficiency. Near Threshold Computing is a key research area in the Ultra-Low-Power domain, as it promises a tenfold improvement in energy efficiency compared to super-threshold operation and it mitigates thermal bottlenecks. The physical implications of modern deep sub-micron technology are severely limiting performance and reliability of modern designs. Reliability becomes a major obstacle when operating in NTC, especially memory operation becomes unreliable and can compromise system correctness. In the present work a novel hybrid memory architecture is devised to overcome reliability issues and at the same time improve energy efficiency by means of aggressive voltage scaling when allowed by workload requirements. Variability is another great drawback of near-threshold operation. The greatly increased sensitivity to threshold voltage variations in today a major concern for electronic devices. We introduce a variation-tolerant extension of the baseline many-core architecture. By means of micro-architectural knobs and a lightweight runtime control unit, the baseline architecture becomes dynamically tolerant to variations.

Heterogeneous Multi-core Architectures for High Performance Computing

This thesis deals with heterogeneous architectures in standard workstations. Heterogeneous architectures represent an appealing alternative to traditional supercomputers because they are based on commodity components fabricated in large quantities. Hence their price-performance ratio is unparalleled in the world of high performance computing (HPC). In particular, different aspects related to the performance and consumption of heterogeneous architectures have been explored. The thesis initially focuses on an efficient implementation of a parallel application, where the execution time is dominated by an high number of floating point instructions. Then the thesis touches the central problem of efficient management of power peaks in heterogeneous computing systems. Finally it discusses a memory-bounded problem, where the execution time is dominated by the memory latency. Specifically, the following main contributions have been carried out: A novel framework for the design and analysis of solar field for Central Receiver Systems (CRS) has been developed. The implementation based on desktop workstation equipped with multiple Graphics Processing Units (GPUs) is motivated by the need to have an accurate and fast simulation environment for studying mirror imperfection and non-planar geometries. Secondly, a power-aware scheduling algorithm on heterogeneous CPU-GPU architectures, based on an efficient distribution of the computing workload to the resources, has been realized. The scheduler manages the resources of several computing nodes with a view to reducing the peak power. The two main contributions of this work follow: the approach reduces the supply cost due to high peak power whilst having negligible impact on the parallelism of computational nodes. from another point of view the developed model allows designer to increase the number of cores without increasing the capacity of the power supply unit. Finally, an implementation for efficient graph exploration on reconfigurable architectures is presented. The purpose is to accelerate graph exploration, reducing the number of random memory accesses.

Parallel-recusive filter structures for the computation of discrete transforms

A general approach is presented for implementing discrete transforms as a set of first-order or second-order recursive digital filters. Clenshaw's recurrence formulae are used to formulate the second-order filters. The resulting structure is suitable for efficient implementation of discrete transforms in VLSI or FPGA circuits. The general approach is applied to the discrete Legendre transform as an illustration.

Lock-free Parallel Dynamic Programming

We show a method for parallelizing top down dynamic programs in a straightforward way by a careful choice of a lock-free shared hash table implementation and randomization of the order in which the dynamic program computes its subproblems. This generic approach is applied to dynamic programs for knapsack, shortest paths, and RNA structure alignment, as well as to a state-of-the-art solution for minimizing the máximum number of open stacks. Experimental results are provided on three different modern multicore architectures which show that this parallelization is effective and reasonably scalable. In particular, we obtain over 10 times speedup for 32 threads on the open stacks problem.

And-or parallel prolog: a recomputation based approach

We argüe that in order to exploit both Independent And- and Or-parallelism in Prolog programs there is advantage in recomputing some of the independent goals, as opposed to all their solutions being reused. We present an abstract model, called the Composition-Tree, for representing and-or parallelism in Prolog Programs. The Composition-tree closely mirrors sequential Prolog execution by recomputing some independent goals rather than fully re-using them. We also outline two environment representation techniques for And-Or parallel execution of full Prolog based on the Composition-tree model abstraction. We argüe that these techniques have advantages over earlier proposals for exploiting and-or parallelism in Prolog.

Some techniques for automated, resource-aware distributed and mobile computing in a multi-paradigm programming system

Distributed parallel execution systems speed up applications by splitting tasks into processes whose execution is assigned to different receiving nodes in a high-bandwidth network. On the distributing side, a fundamental problem is grouping and scheduling such tasks such that each one involves sufñcient computational cost when compared to the task creation and communication costs and other such practical overheads. On the receiving side, an important issue is to have some assurance of the correctness and characteristics of the code received and also of the kind of load the particular task is going to pose, which can be specified by means of certificates. In this paper we present in a tutorial way a number of general solutions to these problems, and illustrate them through their implementation in the Ciao multi-paradigm language and program development environment. This system includes facilities for parallel and distributed execution, an assertion language for specifying complex programs properties (including safety and resource-related properties), and compile-time and run-time tools for performing automated parallelization and resource control, as well as certification of programs with resource consumption assurances and efñcient checking of such certificates.

IDRA (IDeal Resource Allocation): A tool for computing ideal speedups

Performance studies of actual parallel systems usually tend to concéntrate on the effectiveness of a given implementation. This is often done in the absolute, without quantitave reference to the potential parallelism contained in the programs from the point of view of the execution paradigm. We feel that studying the parallelism inherent to the programs is interesting, as it gives information about the best possible behavior of any implementation and thus allows contrasting the results obtained. We propose a method for obtaining ideal speedups for programs through a combination of sequential or parallel execution and simulation, and the algorithms that allow implementing the method. Our approach is novel and, we argüe, more accurate than previously proposed methods, in that a crucial part of the data - the execution times of tasks - is obtained from actual executions, while speedup is computed by simulation. This allows obtaining speedup (and other) data under controlled and ideal assumptions regarding issues such as number of processor, scheduling algorithm and overheads, etc. The results obtained can be used for example to evalúate the ideal parallelism that a program contains for a given model of execution and to compare such "perfect" parallelism to that obtained by a given implementation of that model. We also present a tool, IDRA, which implements the proposed method, and results obtained with IDRA for benchmark programs, which are then compared with those obtained in actual executions on real parallel systems.

An abstract machine based execution model for computer architecture design and efficient implementation of logic programs in parallel.

The term "Logic Programming" refers to a variety of computer languages and execution models which are based on the traditional concept of Symbolic Logic. The expressive power of these languages offers promise to be of great assistance in facing the programming challenges of present and future symbolic processing applications in Artificial Intelligence, Knowledge-based systems, and many other areas of computing. The sequential execution speed of logic programs has been greatly improved since the advent of the first interpreters. However, higher inference speeds are still required in order to meet the demands of applications such as those contemplated for next generation computer systems. The execution of logic programs in parallel is currently considered a promising strategy for attaining such inference speeds. Logic Programming in turn appears as a suitable programming paradigm for parallel architectures because of the many opportunities for parallel execution present in the implementation of logic programs. This dissertation presents an efficient parallel execution model for logic programs. The model is described from the source language level down to an "Abstract Machine" level suitable for direct implementation on existing parallel systems or for the design of special purpose parallel architectures. Few assumptions are made at the source language level and therefore the techniques developed and the general Abstract Machine design are applicable to a variety of logic (and also functional) languages. These techniques offer efficient solutions to several areas of parallel Logic Programming implementation previously considered problematic or a source of considerable overhead, such as the detection and handling of variable binding conflicts in AND-Parallelism, the specification of control and management of the execution tree, the treatment of distributed backtracking, and goal scheduling and memory management issues, etc. A parallel Abstract Machine design is offered, specifying data areas, operation, and a suitable instruction set. This design is based on extending to a parallel environment the techniques introduced by the Warren Abstract Machine, which have already made very fast and space efficient sequential systems a reality. Therefore, the model herein presented is capable of retaining sequential execution speed similar to that of high performance sequential systems, while extracting additional gains in speed by efficiently implementing parallel execution. These claims are supported by simulations of the Abstract Machine on sample programs.

Parallel Algorithms for FE-BE Coupling

When non linear physical systems of infinite extent are modelled, such as tunnels and perforations, it is necessary to simulate suitably the solution in the infinite as well as the non linearity. The finite element method (FEM) is a well known procedure for simulating the non linear behavior. However, the treatment of the infinite field with domain truncations is often questionable. On the other hand, the boundary element method (BEM) is suitable to simulate the infinite behavior without truncations. Because of this, by the combination of both methods, suitable use of the advantages of each one may be obtained. Several possibilities of FEM-BEM coupling and their performance in some practical cases are discussed in this paper. Parallelizable coupling algorithms based on domain decomposition are developed and compared with the most traditional coupling methods.

Parallel workflows to personalize clinical guidelines recommendations: application to gestational diabetes mellitus

The MobiGuide system provides patients with personalized decision support tools, based on computerized clinical guidelines, in a mobile environment. The generic capabilities of the system will be demonstrated applied to the clinical domain of Gestational Diabetes (GD). This paper presents a methodology to identify personalized recommendations, obtained from the analysis of the GD guideline. We added a conceptual parallel part to the formalization of the GD guideline called "parallel workflow" that allows considering patient?s personal context and preferences. As a result of analysing the GD guideline and eliciting medical knowledge, we identified three different types of personalized advices (therapy, measurements and upcoming events) that will be implemented to perform patients? guiding at home, supported by the MobiGuide system. These results will be essential to determine the distribution of functionalities between mobile and server decision support capabilities.

ParAlign: a parallel sequence alignment algorithm for rapid and sensitive database searches

There is a need for faster and more sensitive algorithms for sequence similarity searching in view of the rapidly increasing amounts of genomic sequence data available. Parallel processing capabilities in the form of the single instruction, multiple data (SIMD) technology are now available in common microprocessors and enable a single microprocessor to perform many operations in parallel. The ParAlign algorithm has been specifically designed to take advantage of this technology. The new algorithm initially exploits parallelism to perform a very rapid computation of the exact optimal ungapped alignment score for all diagonals in the alignment matrix. Then, a novel heuristic is employed to compute an approximate score of a gapped alignment by combining the scores of several diagonals. This approximate score is used to select the most interesting database sequences for a subsequent Smith–Waterman alignment, which is also parallelised. The resulting method represents a substantial improvement compared to existing heuristics. The sensitivity and specificity of ParAlign was found to be as good as Smith–Waterman implementations when the same method for computing the statistical significance of the matches was used. In terms of speed, only the significantly less sensitive NCBI BLAST 2 program was found to outperform the new approach. Online searches are available at http://dna.uio.no/search/

Janus: an FPGA-based system for high-performance scientific computing

This paper describes JANUS, a modular massively parallel and reconfigurable FPGA-based computing system. Each JANUS module has a computational core and a host. The computational core is a 4x4 array of FPGA-based processing elements with nearest-neighbor data links. Processors are also directly connected to an I/O node attached to the JANUS host, a conventional PC. JANUS is tailored for, but not limited to, the requirements of a class of hard scientific applications characterized by regular code structure, unconventional data manipulation instructions and not too large data-base size. We discuss the architecture of this configurable machine, and focus on its use on Monte Carlo simulations of statistical mechanics. On this class of application JANUS achieves impressive performances: in some cases one JANUS processing element outperfoms high-end PCs by a factor ≈1000. We also discuss the role of JANUS on other classes of scientific applications.

Real-time 3D semi-local surface patch extraction using GPGPU

Feature vectors can be anything from simple surface normals to more complex feature descriptors. Feature extraction is important to solve various computer vision problems: e.g. registration, object recognition and scene understanding. Most of these techniques cannot be computed online due to their complexity and the context where they are applied. Therefore, computing these features in real-time for many points in the scene is impossible. In this work, a hardware-based implementation of 3D feature extraction and 3D object recognition is proposed to accelerate these methods and therefore the entire pipeline of RGBD based computer vision systems where such features are typically used. The use of a GPU as a general purpose processor can achieve considerable speed-ups compared with a CPU implementation. In this work, advantageous results are obtained using the GPU to accelerate the computation of a 3D descriptor based on the calculation of 3D semi-local surface patches of partial views. This allows descriptor computation at several points of a scene in real-time. Benefits of the accelerated descriptor have been demonstrated in object recognition tasks. Source code will be made publicly available as contribution to the Open Source Point Cloud Library.

Parallel 3D Simulation of a Fault Gouge using the Lattice Solid Model

Despite the insight gained from 2-D particle models, and given that the dynamics of crustal faults occur in 3-D space, the question remains, how do the 3-D fault gouge dynamics differ from those in 2-D? Traditionally, 2-D modeling has been preferred over 3-D simulations because of the computational cost of solving 3-D problems. However, modern high performance computing architectures, combined with a parallel implementation of the Lattice Solid Model (LSM), provide the opportunity to explore 3-D fault micro-mechanics and to advance understanding of effective constitutive relations of fault gouge layers. In this paper, macroscopic friction values from 2-D and 3-D LSM simulations, performed on an SGI Altix 3700 super-cluster, are compared. Two rectangular elastic blocks of bonded particles, with a rough fault plane and separated by a region of randomly sized non-bonded gouge particles, are sheared in opposite directions by normally-loaded driving plates. The results demonstrate that the gouge particles in the 3-D models undergo significant out-of-plane motion during shear. The 3-D models also exhibit a higher mean macroscopic friction than the 2-D models for varying values of interparticle friction. 2-D LSM gouge models have previously been shown to exhibit accelerating energy release in simulated earthquake cycles, supporting the Critical Point hypothesis. The 3-D models are shown to also display accelerating energy release, and good fits of power law time-to-failure functions to the cumulative energy release are obtained.