No description available.

The level of Thread-Level Parallelism (TLP), Instruction-Level Parallelism (ILP), and Memory-Level Parallelism (MLP) varies across programs and across program phases. Hence, every program requires different underlying core microarchitecture resources for high performance and/or energy efficiency. Current core microarchitectures are inefficient because they are fixed at design time and do not adapt to variable TLP, ILP, or MLP. I show that if a core microarchitecture can adapt to the variation in TLP, ILP, and MLP, significantly higher performance and/or energy efficiency can be achieved. I propose MorphCore, a low-overhead adaptive microarchitecture built from a traditional OOO core with small changes. MorphCore adapts to TLP by operating in two modes: (a) as a wide-width large-OOO-window core when TLP is low and ILP is high, and (b) as a high-performance low-energy highly-threaded in-order SMT core when TLP is high. MorphCore adapts to ILP and MLP by varying the superscalar width and the out-of-order (OOO) window size by operating in four modes: (1) as a wide-width large-OOO-window core, 2) as a wide-width medium-OOO-window core, 3) as a medium-width large-OOO-window core, and 4) as a medium-width medium-OOO-window core. My evaluation with single-thread and multi-thread benchmarks shows that when highest single-thread performance is desired, MorphCore achieves performance similar to a traditional out-of-order core. When energy efficiency is desired on single-thread programs, MorphCore reduces energy by up to 15% (on average 8%) over an out-of-order core. When high multi-thread performance is desired, MorphCore increases performance by 21% and reduces energy consumption by 20% over an out-of-order core. Thus, for multi-thread programs, MorphCore's energy efficiency is similar to highly-threaded throughput-optimized small and medium core architectures, and its performance is two-thirds of their potential. / text

The gap between processor speed and memory latency has led to the use of caches in the memory systems of modern computers. Programs must use the caches efficiently and exploit data locality for maximum performance. Multiprocessors, built from many processing units, are becoming commonplace not only in large servers but also in smaller systems such as personal computers. Multiprocessors require careful data locality optimizations since accesses from other processors can lead to invalidations and false sharing cache misses. This thesis explores hardware and software approaches for creating and exploiting temporal and spatial locality in multiprocessors. We propose the capacity prefetching technique, which efficiently reduces the number of cache misses but avoids false sharing by distinguishing between cache lines involved in communication from non-communicating cache lines at run-time. Prefetching techniques often lead to increased coherence and data traffic. The new bundling technique avoids one of these drawbacks and reduces the coherence traffic in multiprocessor prefetchers. This is especially important in snoop-based systems where the coherence bandwidth is a scarce resource. Most of the studies have been performed on advanced scientific algorithms. This thesis demonstrates that a cc-NUMA multiprocessor, with hardware data migration and replication optimizations, efficiently exploits the temporal locality in such codes. We further present a method of parallelizing a multigrid Gauss-Seidel partial differential equation solver, which creates temporal locality at the expense of increased communication. Our conclusion is that on modern chip multiprocessors, it is more important to optimize algorithms for data locality than to avoid communication, since communication can take place using a shared cache.

No description available.

<p>Plentiful research has addressed low-complexity software-based shared-memory systems since the idea was first introduced more than two decades ago. However, software-coherent systems have not been very successful in the commercial marketplace. We believe there are two main reasons for this: lack of performance and/or lack of binary compatibility.</p><p>This thesis studies multiple aspects of how to design future binary-compatible high-performance scalable shared-memory servers while keeping the hardware complexity at a minimum. It starts with a software-based distributed shared-memory system relying on no specific hardware support and gradually moves towards architectures with simple hardware support.</p><p>The evaluation is made in a modern chip-multiprocessor environment with both high-performance compute workloads and commercial applications. It shows that implementing the coherence-violation detection in hardware while solving the interchip coherence in software allows for high-performing binary-compatible systems with very low hardware complexity. Our second-generation hardware-software hybrid performs on par with, and often better than, traditional hardware-only designs.</p><p>Based on our results, we conclude that it is not only possible to design simple systems while maintaining performance and the binary-compatibility envelope, it is often possible to get better performance than in traditional and more complex designs.</p><p>We also explore two new techniques for evaluating a new shared-memory design throughout this work: adjustable simulation fidelity and statistical multiprocessor cache modeling.</p>

Plentiful research has addressed low-complexity software-based shared-memory systems since the idea was first introduced more than two decades ago. However, software-coherent systems have not been very successful in the commercial marketplace. We believe there are two main reasons for this: lack of performance and/or lack of binary compatibility. This thesis studies multiple aspects of how to design future binary-compatible high-performance scalable shared-memory servers while keeping the hardware complexity at a minimum. It starts with a software-based distributed shared-memory system relying on no specific hardware support and gradually moves towards architectures with simple hardware support. The evaluation is made in a modern chip-multiprocessor environment with both high-performance compute workloads and commercial applications. It shows that implementing the coherence-violation detection in hardware while solving the interchip coherence in software allows for high-performing binary-compatible systems with very low hardware complexity. Our second-generation hardware-software hybrid performs on par with, and often better than, traditional hardware-only designs. Based on our results, we conclude that it is not only possible to design simple systems while maintaining performance and the binary-compatibility envelope, it is often possible to get better performance than in traditional and more complex designs. We also explore two new techniques for evaluating a new shared-memory design throughout this work: adjustable simulation fidelity and statistical multiprocessor cache modeling.

The number of processor cores and on-chip cache size has been increasing on chip multiprocessors (CMPs). As a result, leakage power dissipated in the on-chip cache has become very significant. We explore various techniques to switch-off the over-allocated cache so as to reduce leakage power consumed by it. A large cache offers non-uniform access latency to different cores present on a CMP and such a cache is called “Non-Uniform Cache Architecture (NUCA)”. Past studies have explored techniques to reduce leakage power for uniform access latency caches and with a single application executing on a uniprocessor. Our ideas of power optimized caches are applicable to any memory technology and architecture for which the difference of leakage power in the on-state and off-state of on-chip cache bank is significant. Switching off the last level shared cache on a CMP is a challenging problem due to concurrently executing threads/processes and large dispersed NUCA cache. Hence, to determine cache requirement on a CMP, first we propose a new highly accurate method to estimate working set size of an application, which we call “tagged working set size estimation (TWSS)” method. This method has a negligible hardware storage overhead of 0.1% of the cache size. The use of TWSS is demonstrated by adaptively adjusting cache associativity. Our ideas of adaptable associative cache is scalable with respect to the number of cores present on a CMP. It uses information available locally in a tile on a tiled CMP and thus avoids network access unlike other commonly used heuristics such as average memory access latency and cache miss ratio. Our implementation gives 25% and 19% higher EDP savings than that obtained with average memory access latency and cache miss ratio heuristics on a static NUCA platform (SNUCA), respectively. Cache misses increase with reduced cache associativity. Hence, we also propose to map some of the L2 slices onto the rest L2 slices and switch-off mapped L2 slices. The L2 slice includes all L2 banks in a tile. We call this technique the “remap policy”. Some applications execute with lesser number of threads than available cores during their execution. In such applications L2 slices which are farther to those threads are switched-off and mapped on-to L2 slices which are located nearer to those threads. By using nearer L2 slices with the help of remapped technology, some applications show improved execution time apart from reduction in leakage power consumption in NUCA caches. To estimate the maximum possible gains that can be obtained using the remap policy, we statically determine the near-optimal remap configuration using the genetic algorithms. We formulate this problem as a energy-delay product minimization problem. Our dynamic remap policy implementation gives energy-delay savings within an average of 5% than that obtained with the near-optimal remap configuration. Energy-delay product can also be minimized by improving execution time, which depends mainly on the static and dynamic NUCA access policies (DNUCA). The suitability of cache access policy depends on data sharing properties of a multi-threaded application. Hence, we propose three indices to quantify data sharing properties of an application and use them to predict a more suitable cache access policy among SNUCA and DNUCA for an application.

Les réseaux-sur-puces (NoCs) utilisés dans les architectures multiprocesseurs-sur-puces posent des défis importants aux approches d'ordonnancement temps réel en ligne (dynamique) et hors-ligne (statique). Un NoC contient un grand nombre de points de contention potentiels, a une capacité de bufferisation limitée et le contrôle réseau fonctionne à l'échelle de petits paquets de données. Par conséquent, l'allocation efficace de ressources nécessite l'utilisation des algorithmes da faible complexité sur des modèles de matériel avec un niveau de détail sans précédent dans l'ordonnancement temps réel. Nous considérons dans cette thèse une approche d'ordonnancement statique sur des architectures massivement parallèles (Massively parallel processor arrays ou MPPAs) caractérisées par un grand nombre (quelques centaines) de c¿urs de calculs. Nous identifions les mécanismes matériels facilitant l'analyse temporelle et l'allocation efficace de ressources dans les MPPAs existants. Nous déterminons que le NoC devrait permettre l'ordonnancement hors-ligne de communications, d'une manière synchronisée avec l'ordonnancement de calculs sur les processeurs. Au niveau logiciel, nous proposons une nouvelle méthode d'allocation et d'ordonnancement capable de synthétiser des ordonnancements globaux de calculs et de communications couvrants toutes les ressources d'exécution, de communication et de la mémoire d'un MPPA. Afin de permettre une utilisation efficace de ressources du matériel, notre méthode prend en compte les spécificités architecturales d'un MPPA et implémente des techniques d'ordonnancement avancées comme la préemption pré-calculée de transmissions de données. Nous avons évalué n / On-chip networks (NoCs) used in multiprocessor systems-on-chips (MPSoCs) pose significant challenges to both on-line (dynamic) and off-line (static) real-time scheduling approaches. They have large numbers of potential contention points, have limited internal buffering capabilities, and network control operates at the scale of small data packets. Therefore, efficient resource allocation requires scalable algorithms working on hardware models with a level of detail that is unprecedented in real-time scheduling. We consider in this thesis a static scheduling approach, and we target massively parallel processor arrays (MPPAs), which are MPSoCs with large numbers (hundreds) of processing cores. We first identify and compare the hardware mechanisms supporting precise timing analysis and efficient resource allocation in existing MPPA platforms. We determine that the NoC should ideally provide the means of enforcing a global communications schedule that is computed off-line (before execution) and which is synchronized with the scheduling of computations on processors. On the software side, we propose a novel allocation and scheduling method capable of synthesizing such global computation and communication schedules covering all the execution, communication, and memory resources in an MPPA. To allow an efficient use of the hardware resources, our method takes into account the specificities of MPPA hardware and implements advanced scheduling techniques such as pre-computed preemption of data transmissions. We evaluate our technique by mapping two signal processing applications, for which we obtain good latency, throughput, and resource use figures.

Les systèmes multiprocesseurs sur puce électronique (On-Chip Multiprocessor [OCM]) sont considérés comme les meilleures structures pour occuper l'espace disponible sur les circuits intégrés actuels. Dans nos travaux, nous nous intéressons à un modèle architectural, appelé architecture isométrique de systèmes multiprocesseurs sur puce, qui permet d'évaluer, de prédire et d'optimiser les systèmes OCM en misant sur une organisation efficace des nœuds (processeurs et mémoires), et à des méthodologies qui permettent d'utiliser efficacement ces architectures. Dans la première partie de la thèse, nous nous intéressons à la topologie du modèle et nous proposons une architecture qui permet d'utiliser efficacement et massivement les mémoires sur la puce. Les processeurs et les mémoires sont organisés selon une approche isométrique qui consiste à rapprocher les données des processus plutôt que d'optimiser les transferts entre les processeurs et les mémoires disposés de manière conventionnelle. L'architecture est un modèle maillé en trois dimensions. La disposition des unités sur ce modèle est inspirée de la structure cristalline du chlorure de sodium (NaCl), où chaque processeur peut accéder à six mémoires à la fois et où chaque mémoire peut communiquer avec autant de processeurs à la fois. Dans la deuxième partie de notre travail, nous nous intéressons à une méthodologie de décomposition où le nombre de nœuds du modèle est idéal et peut être déterminé à partir d'une spécification matricielle de l'application qui est traitée par le modèle proposé. Sachant que la performance d'un modèle dépend de la quantité de flot de données échangées entre ses unités, en l'occurrence leur nombre, et notre but étant de garantir une bonne performance de calcul en fonction de l'application traitée, nous proposons de trouver le nombre idéal de processeurs et de mémoires du système à construire. Aussi, considérons-nous la décomposition de la spécification du modèle à construire ou de l'application à traiter en fonction de l'équilibre de charge des unités. Nous proposons ainsi une approche de décomposition sur trois points : la transformation de la spécification ou de l'application en une matrice d'incidence dont les éléments sont les flots de données entre les processus et les données, une nouvelle méthodologie basée sur le problème de la formation des cellules (Cell Formation Problem [CFP]), et un équilibre de charge de processus dans les processeurs et de données dans les mémoires. Dans la troisième partie, toujours dans le souci de concevoir un système efficace et performant, nous nous intéressons à l'affectation des processeurs et des mémoires par une méthodologie en deux étapes. Dans un premier temps, nous affectons des unités aux nœuds du système, considéré ici comme un graphe non orienté, et dans un deuxième temps, nous affectons des valeurs aux arcs de ce graphe. Pour l'affectation, nous proposons une modélisation des applications décomposées en utilisant une approche matricielle et l'utilisation du problème d'affectation quadratique (Quadratic Assignment Problem [QAP]). Pour l'affectation de valeurs aux arcs, nous proposons une approche de perturbation graduelle, afin de chercher la meilleure combinaison du coût de l'affectation, ceci en respectant certains paramètres comme la température, la dissipation de chaleur, la consommation d'énergie et la surface occupée par la puce. Le but ultime de ce travail est de proposer aux architectes de systèmes multiprocesseurs sur puce une méthodologie non traditionnelle et un outil systématique et efficace d'aide à la conception dès la phase de la spécification fonctionnelle du système. / On-Chip Multiprocessor (OCM) systems are considered to be the best structures to occupy the abundant space available on today integrated circuits (IC). In our thesis, we are interested on an architectural model, called Isometric on-Chip Multiprocessor Architecture (ICMA), that optimizes the OCM systems by focusing on an effective organization of cores (processors and memories) and on methodologies that optimize the use of these architectures. In the first part of this work, we study the topology of ICMA and propose an architecture that enables efficient and massive use of on-chip memories. ICMA organizes processors and memories in an isometric structure with the objective to get processed data close to the processors that use them rather than to optimize transfers between processors and memories, arranged in a conventional manner. ICMA is a mesh model in three dimensions. The organization of our architecture is inspired by the crystal structure of sodium chloride (NaCl), where each processor can access six different memories and where each memory can communicate with six processors at once. In the second part of our work, we focus on a methodology of decomposition. This methodology is used to find the optimal number of nodes for a given application or specification. The approach we use is to transform an application or a specification into an incidence matrix, where the entries of this matrix are the interactions between processors and memories as entries. In other words, knowing that the performance of a model depends on the intensity of the data flow exchanged between its units, namely their number, we aim to guarantee a good computing performance by finding the optimal number of processors and memories that are suitable for the application computation. We also consider the load balancing of the units of ICMA during the specification phase of the design. Our proposed decomposition is on three points: the transformation of the specification or application into an incidence matrix, a new methodology based on the Cell Formation Problem (CFP), and load balancing processes in the processors and data in memories. In the third part, we focus on the allocation of processor and memory by a two-step methodology. Initially, we allocate units to the nodes of the system structure, considered here as an undirected graph, and subsequently we assign values to the arcs of this graph. For the assignment, we propose modeling of the decomposed application using a matrix approach and the Quadratic Assignment Problem (QAP). For the assignment of the values to the arcs, we propose an approach of gradual changes of these values in order to seek the best combination of cost allocation, this under certain metric constraints such as temperature, heat dissipation, power consumption and surface occupied by the chip. The ultimate goal of this work is to propose a methodology for non-traditional, systematic and effective decision support design tools for multiprocessor system architects, from the phase of functional specification.

Les systèmes multiprocesseurs sur puce électronique (On-Chip Multiprocessor [OCM]) sont considérés comme les meilleures structures pour occuper l'espace disponible sur les circuits intégrés actuels. Dans nos travaux, nous nous intéressons à un modèle architectural, appelé architecture isométrique de systèmes multiprocesseurs sur puce, qui permet d'évaluer, de prédire et d'optimiser les systèmes OCM en misant sur une organisation efficace des nœuds (processeurs et mémoires), et à des méthodologies qui permettent d'utiliser efficacement ces architectures. Dans la première partie de la thèse, nous nous intéressons à la topologie du modèle et nous proposons une architecture qui permet d'utiliser efficacement et massivement les mémoires sur la puce. Les processeurs et les mémoires sont organisés selon une approche isométrique qui consiste à rapprocher les données des processus plutôt que d'optimiser les transferts entre les processeurs et les mémoires disposés de manière conventionnelle. L'architecture est un modèle maillé en trois dimensions. La disposition des unités sur ce modèle est inspirée de la structure cristalline du chlorure de sodium (NaCl), où chaque processeur peut accéder à six mémoires à la fois et où chaque mémoire peut communiquer avec autant de processeurs à la fois. Dans la deuxième partie de notre travail, nous nous intéressons à une méthodologie de décomposition où le nombre de nœuds du modèle est idéal et peut être déterminé à partir d'une spécification matricielle de l'application qui est traitée par le modèle proposé. Sachant que la performance d'un modèle dépend de la quantité de flot de données échangées entre ses unités, en l'occurrence leur nombre, et notre but étant de garantir une bonne performance de calcul en fonction de l'application traitée, nous proposons de trouver le nombre idéal de processeurs et de mémoires du système à construire. Aussi, considérons-nous la décomposition de la spécification du modèle à construire ou de l'application à traiter en fonction de l'équilibre de charge des unités. Nous proposons ainsi une approche de décomposition sur trois points : la transformation de la spécification ou de l'application en une matrice d'incidence dont les éléments sont les flots de données entre les processus et les données, une nouvelle méthodologie basée sur le problème de la formation des cellules (Cell Formation Problem [CFP]), et un équilibre de charge de processus dans les processeurs et de données dans les mémoires. Dans la troisième partie, toujours dans le souci de concevoir un système efficace et performant, nous nous intéressons à l'affectation des processeurs et des mémoires par une méthodologie en deux étapes. Dans un premier temps, nous affectons des unités aux nœuds du système, considéré ici comme un graphe non orienté, et dans un deuxième temps, nous affectons des valeurs aux arcs de ce graphe. Pour l'affectation, nous proposons une modélisation des applications décomposées en utilisant une approche matricielle et l'utilisation du problème d'affectation quadratique (Quadratic Assignment Problem [QAP]). Pour l'affectation de valeurs aux arcs, nous proposons une approche de perturbation graduelle, afin de chercher la meilleure combinaison du coût de l'affectation, ceci en respectant certains paramètres comme la température, la dissipation de chaleur, la consommation d'énergie et la surface occupée par la puce. Le but ultime de ce travail est de proposer aux architectes de systèmes multiprocesseurs sur puce une méthodologie non traditionnelle et un outil systématique et efficace d'aide à la conception dès la phase de la spécification fonctionnelle du système. / On-Chip Multiprocessor (OCM) systems are considered to be the best structures to occupy the abundant space available on today integrated circuits (IC). In our thesis, we are interested on an architectural model, called Isometric on-Chip Multiprocessor Architecture (ICMA), that optimizes the OCM systems by focusing on an effective organization of cores (processors and memories) and on methodologies that optimize the use of these architectures. In the first part of this work, we study the topology of ICMA and propose an architecture that enables efficient and massive use of on-chip memories. ICMA organizes processors and memories in an isometric structure with the objective to get processed data close to the processors that use them rather than to optimize transfers between processors and memories, arranged in a conventional manner. ICMA is a mesh model in three dimensions. The organization of our architecture is inspired by the crystal structure of sodium chloride (NaCl), where each processor can access six different memories and where each memory can communicate with six processors at once. In the second part of our work, we focus on a methodology of decomposition. This methodology is used to find the optimal number of nodes for a given application or specification. The approach we use is to transform an application or a specification into an incidence matrix, where the entries of this matrix are the interactions between processors and memories as entries. In other words, knowing that the performance of a model depends on the intensity of the data flow exchanged between its units, namely their number, we aim to guarantee a good computing performance by finding the optimal number of processors and memories that are suitable for the application computation. We also consider the load balancing of the units of ICMA during the specification phase of the design. Our proposed decomposition is on three points: the transformation of the specification or application into an incidence matrix, a new methodology based on the Cell Formation Problem (CFP), and load balancing processes in the processors and data in memories. In the third part, we focus on the allocation of processor and memory by a two-step methodology. Initially, we allocate units to the nodes of the system structure, considered here as an undirected graph, and subsequently we assign values to the arcs of this graph. For the assignment, we propose modeling of the decomposed application using a matrix approach and the Quadratic Assignment Problem (QAP). For the assignment of the values to the arcs, we propose an approach of gradual changes of these values in order to seek the best combination of cost allocation, this under certain metric constraints such as temperature, heat dissipation, power consumption and surface occupied by the chip. The ultimate goal of this work is to propose a methodology for non-traditional, systematic and effective decision support design tools for multiprocessor system architects, from the phase of functional specification.