Taming Control Divergence in GPUs through Control Flow Linearization

Branch divergence is a very commonly occurring performance problem in GPGPU in which the execution of diverging branches is serialized to execute only one control flow path at a time. Existing hardware mechanism to reconverge threads using a stack causes duplicate execution of code for unstructured control flow graphs. Also the stack mechanism cannot effectively utilize the available parallelism among diverging branches. Further, the amount of nested divergence allowed is also limited by depth of the branch divergence stack. In this paper we propose a simple and elegant transformation to handle all of the above mentioned problems. The transformation converts an unstructured CFG to a structured CFG without duplicating user code. It incurs only a linear increase in the number of basic blocks and also the number of instructions. Our solution linearizes the CFG using a predicate variable. This mechanism reconverges the divergent threads as early as possible. It also reduces the depth of the reconvergence stack. The available parallelism in nested branches can be effectively extracted by scheduling the basic blocks to reduce the effect of stalls due to memory accesses. It can also increase execution efficiency of nested loops with different trip counts for different threads. We implemented the proposed transformation at PTX level using the Ocelot compiler infrastructure. We evaluated the technique using various benchmarks to show that it can be effective in handling the performance problem due to divergence in unstructured CFGs.

Automated wordlength optimization framework for multi-source statistical interval-based analysis of nonlinear systems with control-flow structures

El uso de aritmética de punto fijo es una opción de diseño muy extendida en sistemas con fuertes restricciones de área, consumo o rendimiento. Para producir implementaciones donde los costes se minimicen sin impactar negativamente en la precisión de los resultados debemos llevar a cabo una asignación cuidadosa de anchuras de palabra. Encontrar la combinación óptima de anchuras de palabra en coma fija para un sistema dado es un problema combinatorio NP-hard al que los diseñadores dedican entre el 25 y el 50 % del ciclo de diseño. Las plataformas hardware reconfigurables, como son las FPGAs, también se benefician de las ventajas que ofrece la aritmética de coma fija, ya que éstas compensan las frecuencias de reloj más bajas y el uso más ineficiente del hardware que hacen estas plataformas respecto a los ASICs. A medida que las FPGAs se popularizan para su uso en computación científica los diseños aumentan de tamaño y complejidad hasta llegar al punto en que no pueden ser manejados eficientemente por las técnicas actuales de modelado de señal y ruido de cuantificación y de optimización de anchura de palabra. En esta Tesis Doctoral exploramos distintos aspectos del problema de la cuantificación y presentamos nuevas metodologías para cada uno de ellos: Las técnicas basadas en extensiones de intervalos han permitido obtener modelos de propagación de señal y ruido de cuantificación muy precisos en sistemas con operaciones no lineales. Nosotros llevamos esta aproximación un paso más allá introduciendo elementos de Multi-Element Generalized Polynomial Chaos (ME-gPC) y combinándolos con una técnica moderna basada en Modified Affine Arithmetic (MAA) estadístico para así modelar sistemas que contienen estructuras de control de flujo. Nuestra metodología genera los distintos caminos de ejecución automáticamente, determina las regiones del dominio de entrada que ejercitarán cada uno de ellos y extrae los momentos estadísticos del sistema a partir de dichas soluciones parciales. Utilizamos esta técnica para estimar tanto el rango dinámico como el ruido de redondeo en sistemas con las ya mencionadas estructuras de control de flujo y mostramos la precisión de nuestra aproximación, que en determinados casos de uso con operadores no lineales llega a tener tan solo una desviación del 0.04% con respecto a los valores de referencia obtenidos mediante simulación. Un inconveniente conocido de las técnicas basadas en extensiones de intervalos es la explosión combinacional de términos a medida que el tamaño de los sistemas a estudiar crece, lo cual conlleva problemas de escalabilidad. Para afrontar este problema presen tamos una técnica de inyección de ruidos agrupados que hace grupos con las señales del sistema, introduce las fuentes de ruido para cada uno de los grupos por separado y finalmente combina los resultados de cada uno de ellos. De esta forma, el número de fuentes de ruido queda controlado en cada momento y, debido a ello, la explosión combinatoria se minimiza. También presentamos un algoritmo de particionado multi-vía destinado a minimizar la desviación de los resultados a causa de la pérdida de correlación entre términos de ruido con el objetivo de mantener los resultados tan precisos como sea posible. La presente Tesis Doctoral también aborda el desarrollo de metodologías de optimización de anchura de palabra basadas en simulaciones de Monte-Cario que se ejecuten en tiempos razonables. Para ello presentamos dos nuevas técnicas que exploran la reducción del tiempo de ejecución desde distintos ángulos: En primer lugar, el método interpolativo aplica un interpolador sencillo pero preciso para estimar la sensibilidad de cada señal, y que es usado después durante la etapa de optimización. En segundo lugar, el método incremental gira en torno al hecho de que, aunque es estrictamente necesario mantener un intervalo de confianza dado para los resultados finales de nuestra búsqueda, podemos emplear niveles de confianza más relajados, lo cual deriva en un menor número de pruebas por simulación, en las etapas iniciales de la búsqueda, cuando todavía estamos lejos de las soluciones optimizadas. Mediante estas dos aproximaciones demostramos que podemos acelerar el tiempo de ejecución de los algoritmos clásicos de búsqueda voraz en factores de hasta x240 para problemas de tamaño pequeño/mediano. Finalmente, este libro presenta HOPLITE, una infraestructura de cuantificación automatizada, flexible y modular que incluye la implementación de las técnicas anteriores y se proporciona de forma pública. Su objetivo es ofrecer a desabolladores e investigadores un entorno común para prototipar y verificar nuevas metodologías de cuantificación de forma sencilla. Describimos el flujo de trabajo, justificamos las decisiones de diseño tomadas, explicamos su API pública y hacemos una demostración paso a paso de su funcionamiento. Además mostramos, a través de un ejemplo sencillo, la forma en que conectar nuevas extensiones a la herramienta con las interfaces ya existentes para poder así expandir y mejorar las capacidades de HOPLITE. ABSTRACT Using fixed-point arithmetic is one of the most common design choices for systems where area, power or throughput are heavily constrained. In order to produce implementations where the cost is minimized without negatively impacting the accuracy of the results, a careful assignment of word-lengths is required. The problem of finding the optimal combination of fixed-point word-lengths for a given system is a combinatorial NP-hard problem to which developers devote between 25 and 50% of the design-cycle time. Reconfigurable hardware platforms such as FPGAs also benefit of the advantages of fixed-point arithmetic, as it compensates for the slower clock frequencies and less efficient area utilization of the hardware platform with respect to ASICs. As FPGAs become commonly used for scientific computation, designs constantly grow larger and more complex, up to the point where they cannot be handled efficiently by current signal and quantization noise modelling and word-length optimization methodologies. In this Ph.D. Thesis we explore different aspects of the quantization problem and we present new methodologies for each of them: The techniques based on extensions of intervals have allowed to obtain accurate models of the signal and quantization noise propagation in systems with non-linear operations. We take this approach a step further by introducing elements of MultiElement Generalized Polynomial Chaos (ME-gPC) and combining them with an stateof- the-art Statistical Modified Affine Arithmetic (MAA) based methodology in order to model systems that contain control-flow structures. Our methodology produces the different execution paths automatically, determines the regions of the input domain that will exercise them, and extracts the system statistical moments from the partial results. We use this technique to estimate both the dynamic range and the round-off noise in systems with the aforementioned control-flow structures. We show the good accuracy of our approach, which in some case studies with non-linear operators shows a 0.04 % deviation respect to the simulation-based reference values. A known drawback of the techniques based on extensions of intervals is the combinatorial explosion of terms as the size of the targeted systems grows, which leads to scalability problems. To address this issue we present a clustered noise injection technique that groups the signals in the system, introduces the noise terms in each group independently and then combines the results at the end. In this way, the number of noise sources in the system at a given time is controlled and, because of this, the combinato rial explosion is minimized. We also present a multi-way partitioning algorithm aimed at minimizing the deviation of the results due to the loss of correlation between noise terms, in order to keep the results as accurate as possible. This Ph.D. Thesis also covers the development of methodologies for word-length optimization based on Monte-Carlo simulations in reasonable times. We do so by presenting two novel techniques that explore the reduction of the execution times approaching the problem in two different ways: First, the interpolative method applies a simple but precise interpolator to estimate the sensitivity of each signal, which is later used to guide the optimization effort. Second, the incremental method revolves on the fact that, although we strictly need to guarantee a certain confidence level in the simulations for the final results of the optimization process, we can do it with more relaxed levels, which in turn implies using a considerably smaller amount of samples, in the initial stages of the process, when we are still far from the optimized solution. Through these two approaches we demonstrate that the execution time of classical greedy techniques can be accelerated by factors of up to ×240 for small/medium sized problems. Finally, this book introduces HOPLITE, an automated, flexible and modular framework for quantization that includes the implementation of the previous techniques and is provided for public access. The aim is to offer a common ground for developers and researches for prototyping and verifying new techniques for system modelling and word-length optimization easily. We describe its work flow, justifying the taken design decisions, explain its public API and we do a step-by-step demonstration of its execution. We also show, through an example, the way new extensions to the flow should be connected to the existing interfaces in order to expand and improve the capabilities of HOPLITE.

Efficient Mitigation of Data and Control Flow Errors in Microprocessors

The use of microprocessor-based systems is gaining importance in application domains where safety is a must. For this reason, there is a growing concern about the mitigation of SEU and SET effects. This paper presents a new hybrid technique aimed to protect both the data and the control-flow of embedded applications running on microprocessors. On one hand, the approach is based on software redundancy techniques for correcting errors produced in the data. On the other hand, control-flow errors can be detected by reusing the on-chip debug interface, existing in most modern microprocessors. Experimental results show an important increase in the system reliability even superior to two orders of magnitude, in terms of mitigation of both SEUs and SETs. Furthermore, the overheads incurred by our technique can be perfectly assumable in low-cost systems.

Timing analysis of assembler code control-flow paths

Timinganalysis of assembler code is essential to achieve the strongest possible guarantee of correctness for safety-critical, real-time software. Previous work has shown how timingconstrain ts on controlflow paths through high-level language programs can be formalised using the semantics of the statements comprisingthe path. We extend these results to assembler-level code where it becomes possible to not only determine timingconstrain ts, but also to verify them against the known execution times for each instruction. A minimal formal model is developed with both a weakest liberal precondition and a strongest postcondition semantics. However, despite the formalism’s simplicity, it is shown that complex timingb ehaviour associated with instruction pipeliningand iterative code can be modelled accurately.

Blown Jet Vortex Generator Control of a Separated Diffuser Flow

The application of blown jet vortex generators to control flow separation in a diffuser with an opening angle of 10° has been studied using the computational fluid dynamics (CFD) code Fluent 6™. Experimental data is available for the uncontrolled flow in the diffuser. The section of the duct upstream of the diffuser has a height H equal to 15 mm; its length and breadth are 101H and 41H respectively; the diffuser has an expansion ratio of 4.7:1. Fully developed flow is achieved upstream of the diffuser. Pipes of diameters equal to 1.5%, 2.5% and 5% of H were considered; pitch angle was constant at 45° and yaw angle was fixed at 60°; velocity ratio was varied from 1.7 to 8.0; both co-rotating and counter-rotating arrays were studied. The best results were obtained with a counter-rotating array of generators with a hole diameter of 5% of H and a velocity ratio of 3.7.

Comprehensive Path-sensitive Data-flow Analysis

Data-flow analysis is an integral part of any aggressive optimizing compiler. We propose a framework for improving the precision of data-flow analysis in the presence of complex control-flow. W initially perform data-flow analysis to determine those control-flow merges which cause the loss in data-flow analysis precision. The control-flow graph of the program is then restructured such that performing data-flow analysis on the resulting restructured graph gives more precise results. The proposed framework is both simple, involving the familiar notion of product automata, and also general, since it is applicable to any forward data-flow analysis. Apart from proving that our restructuring process is correct, we also show that restructuring is effective in that it necessarily leads to more optimization opportunities. Furthermore, the framework handles the trade-off between the increase in data-flow precision and the code size increase inherent in the restructuring. We show that determining an optimal restructuring is NP-hard, and propose and evaluate a greedy strategy. The framework has been implemented in the Scale research compiler, and instantiated for the specific problem of Constant Propagation. On the SPECINT 2000 benchmark suite we observe an average speedup of 4% in the running times over Wegman-Zadeck conditional constant propagation algorithm and 2% over a purely path profile guided approach.

Partial Flow Sensitivity

Compiler optimizations need precise and scalable analyses to discover program properties. We propose a partially flow-sensitive framework that tries to draw on the scalability of flow-insensitive algorithms while providing more precision at some specific program points. Provided with a set of critical nodes — basic blocks at which more precise information is desired — our partially flow-sensitive algorithm computes a reduced control-flow graph by collapsing some sets of non-critical nodes. The algorithm is more scalable than a fully flow-sensitive one as, assuming that the number of critical nodes is small, the reduced flow-graph is much smaller than the original flow-graph. At the same time, a much more precise information is obtained at certain program points than would had been obtained from a flow-insensitive algorithm.

Taking an Object-Centric View on Dynamic Information with Object Flow Analysis

A large body of research analyzes the runtime execution of a system to extract abstract behavioral views. Those approaches primarily analyze control flow by tracing method execution events or they analyze object graphs of heap snapshots. However, they do not capture how objects are passed through the system at runtime. We refer to the exchange of objects as the object flow, and we claim that object flow is necessary to analyze if we are to understand the runtime of an object-oriented application. We propose and detail Object Flow Analysis, a novel dynamic analysis technique that takes this new information into account. To evaluate its usefulness, we present a visual approach that allows a developer to study classes and components in terms of how they exchange objects at runtime. We illustrate our approach on three case studies.

Flow-Centric, Back-In-Time Debugging

Conventional debugging tools present developers with means to explore the run-time context in which an error has occurred. In many cases this is enough to help the developer discover the faulty source code and correct it. However, rather often errors occur due to code that has executed in the past, leaving certain objects in an inconsistent state. The actual run-time error only occurs when these inconsistent objects are used later in the program. So-called back-in-time debuggers help developers step back through earlier states of the program and explore execution contexts not available to conventional debuggers. Nevertheless, even back-in-time debuggers do not help answer the question, ``Where did this object come from?'' The Object-Flow Virtual Machine, which we have proposed in previous work, tracks the flow of objects to answer precisely such questions, but this VM does not provide dedicated debugging support to explore faulty programs. In this paper we present a novel debugger, called Compass, to navigate between conventional run-time stack-oriented control flow views and object flows. Compass enables a developer to effectively navigate from an object contributing to an error back-in-time through all the code that has touched the object. We present the design and implementation of Compass, and we demonstrate how flow-centric, back-in-time debugging can be used to effectively locate the source of hard-to-find bugs.

The Order Fulfillment Process Model

This appendix describes the Order Fulfillment process followed by a fictitious company named Genko Oil. The process is freely inspired by the VICS (Voluntary Inter-industry Commerce Solutions) reference model1 and provides a demonstration of YAWL’s capabilities in modelling complex control-flow, data and resourcing requirements.

Configurable multi-perspective business process models

A configurable process model provides a consolidated view of a family of business processes. It promotes the reuse of proven practices by providing analysts with a generic modelling artifact from which to derive individual process models. Unfortunately, the scope of existing notations for configurable process modelling is restricted, thus hindering their applicability. Specifically, these notations focus on capturing tasks and control-flow dependencies, neglecting equally important ingredients of business processes such as data and resources. This research fills this gap by proposing a configurable process modelling notation incorporating features for capturing resources, data and physical objects involved in the performance of tasks. The proposal has been implemented in a toolset that assists analysts during the configuration phase and guarantees the correctness of the resulting process models. The approach has been validated by means of a case study from the film industry.

YAWL : power through patterns

Workflow Management Systems (WfMSs) enable the development and maintenance of workflow specifications at design time and their execution and monitoring at runtime. The open source WfMS YAWL supports the YAWL language – a formally defined language based on Petri nets which offers comprehensive support for control-flow and resource patterns. In addition, the YAWL system provides extensive support for process flexibility, in particular for process configuration, exception handling, dynamic workflow and declarative workflow. Due to its formal foundation, sophisticated verification support can also be achieved. This paper presents the YAWL system and its main applications.