Academic literature on the topic 'Streaming Dataflows'

Les applications de gestion de flux sont responsables de la majorité des calculs des systèmes embarqués (vidéo conférence, vision par ordinateur). Leurs exigences de haute performance rendent leur mise en œuvre parallèle nécessaire. Par conséquent, il est de plus en plus courant que les systèmes embarqués modernes incluent des processeurs multi-cœurs qui permettent un parallélisme massif. La mise en œuvre des applications de gestion de flux sur des multi-cœurs est difficile à cause de leur complexité, qui tend à augmenter, et de leurs exigences strictes à la fois qualitatives (robustesse, fiabilité) et quantitatives (débit, consommation d'énergie). Ceci est observé dans l'évolution de codecs vidéo qui ne cessent d'augmenter en complexité, tandis que leurs exigences de performance demeurent les mêmes. Les modèles de calcul (MdC) flot de données ont été développés pour faciliter la conception de ces applications qui sont typiquement composées de filtres qui échangent des flux de données via des liens de communication. Ces modèles fournissent une représentation intuitive des applications de gestion de flux, tout en exposant le parallélisme de tâches de l'application. En outre, ils fournissent des analyses statiques pour la vivacité et l'exécution en mémoire bornée. Cependant, les applications de gestion de flux modernes comportent des filtres qui échangent des quantités de données variables, et des liens de communication qui peuvent être activés / désactivés. Dans cette thèse, nous présentons un nouveau MdC flot de données, le Boolean Parametric Data Flow (BPDF), qui permet le paramétrage de la quantité de données échangées entre les filtres en utilisant des paramètres entiers et l'activation et la désactivation de liens de communication en utilisant des paramètres booléens. De cette manière, BPDF est capable de exprimer des applications plus complexes, comme les décodeurs vidéo modernes. Malgré l'augmentation de l'expressivité, les applications BPDF restent statiquement analysables pour la vivacité et l'exécution en mémoire bornée. Cependant, l'expressivité accrue complique grandement la mise en œuvre. Les paramètres entiers entraînent des dépendances de données de type paramétrique et les paramètres booléens peuvent désactiver des liens de communication et ainsi éliminer des dépendances de données. Pour cette raison, nous proposons un cadre d'ordonnancement qui produit des ordonnancements de type ``aussi tôt que possible'' (ASAP) pour un placement statique donné. Il utilise des contraintes d'ordonnancement, soit issues de l'application (dépendance de données) ou de l'utilisateur (optimisations d'ordonnancement). Les contraintes sont analysées pour la vivacité et, si possible, simplifiées. De cette façon, notre cadre permet une grande variété de politiques d'ordonnancement, tout en garantissant la vivacité de l'application. Enfin, le calcul du débit d'une application est important tant avant que pendant l'exécution. Il permet de vérifier que l'application satisfait ses exigences de performance et il permet de prendre des décisions d'ordonnancement à l'exécution qui peuvent améliorer la performance ou la consommation d'énergie. Nous traitons ce problème en trouvant des expressions paramétriques pour le débit maximum d'un sous-ensemble de BPDF. Enfin, nous proposons un algorithme qui calcule une taille des buffers suffisante pour que l'application BPDF ait un débit maximum<br>Streaming applications are responsible for the majority of the computation load in many embedded systems (video conferencing, computer vision etc). Their high performance requirements make parallel implementations a necessity. Hence, more and more modern embedded systems include many-core processors that allow massive parallelism. Parallel implementation of streaming applications on many-core platforms is challenging because of their complexity, which tends to increase, and their strict requirements both qualitative (e.g., robustness, reliability) and quantitative (e.g., throughput, power consumption). This is observed in the evolution of video codecs that keep increasing in complexity, while their performance requirements remain the same or even increase. Data flow models of computation (MoCs) have been developed to facilitate the design process of such applications, which are typically composed of filters exchanging streams of data via communication links. Data flow MoCs provide an intuitive representation of streaming applications, while exposing the available parallelism of the application. Moreover, they provide static analyses for liveness and boundedness. However, modern streaming applications feature filters that exchange variable amounts of data, and communication links that are not always active. In this thesis, we present a new data flow MoC, the Boolean Parametric Data Flow (BPDF), that allows parametrization of the amount of data exchanged between the filters using integer parameters and the enabling and disabling of communication links using boolean parameters. In this way, BPDF is able to capture more complex streaming applications, like video decoders. Despite the increase in expressiveness, BPDF applications remain statically analyzable for liveness and boundedness. However, increased expressiveness greatly complicates implementation. Integer parameters result in parametric data dependencies and the boolean parameters disable communication links, effectively removing data dependencies. We propose a scheduling framework that facilitates the scheduling of BPDF applications. Our scheduling framework produces as soon as possible schedules for a given static mapping. It takes us input scheduling constraints that derive either from the application (data dependencies) or from the user (schedule optimizations). The constraints are analyzed for liveness and, if possible, simplified. In this way, our framework provides flexibility, while guaranteeing the liveness of the application. Finally, calculation of the throughput of an application is important both at compile-time and at run-time. It allows to verify at compile-time that the application meets its performance requirements and it allows to take scheduling decisions at run-time that can improve performance or power consumption. We approach this problem by finding parametric throughput expressions for the maximum throughput of a subset of BPDF graphs. Finally, we provide an algorithm that calculates sufficient buffer sizes for the BPDF graph to operate at maximum throughput

The velocity dimension of Big Data refers to the need to rapidly process data that arrives continuously as streams of messages or events. Distributed Stream Processing Systems (DSPS) refer to distributed programming and runtime platforms that allow users to define a composition of dataflow logic that are executed on distributed resources over streams of incoming messages. A DSPS uses commodity clusters and Cloud Virtual Machines (VMs) for its execution. In order to meet the required performance for these applications, the DSPS needs to schedule these dataßows efficiently over the resources. Despite their growing use, resource scheduling for DSPSÕs tends to be done in an ad hoc manner, favoring empirical and reactive approaches, rather than a model-driven and analytical approach. Such empirical strategies may arrive at an approximate schedule for the dataflow that needs further tuning to meet the quality of service. We propose a model-based scheduling approach that makes use of performance profiles and benchmarks developed for tasks in the dataßow to plan both the resource allocation and the resource mapping that together form the schedule planning process. We propose the Model Based Allocation (MBA) and the Slot Aware Mapping (SAM) approaches that efectively utilize knowledge of the performance model of logic tasks to provide an efficient and predictable scheduling behavior. We implemented and validate these algorithms using the popular open source Apache Storm DSPS for several micro and application dataflows. The results show that our model-driven approach is able to reduce the amount of required resources (VMs) by 30% − 50% relative to existing techniques. Also we see that our strategies o↵er a predictable behavior that ensures that the expected and actual rates supported and resources used match closely. This can enable deterministic schedule planning even under dynamic conditions. Besides this static scheduling, we also examine the ability to dynamically consolidate tasks onto fewer VMs when the load on the dataßow decreases or the VMs get fragmented. We propose reliable task migration models for Apache Storm dataßows that are able to rapidly move the task assignment in the cluster, and resume the dataflow execution without any message loss.

The velocity dimension of Big Data refers to the need to rapidly process data that arrives continuously as streams of messages or events. Distributed Stream Processing Systems (DSPS) refer to distributed programming and runtime platforms that allow users to define a composition of dataflow logic that are executed on distributed resources over streams of incoming messages. A DSPS uses commodity clusters and Cloud Virtual Machines (VMs) for its execution. In order to meet the required performance for these applications, the DSPS needs to schedule these dataßows efficiently over the resources. Despite their growing use, resource scheduling for DSPSÕs tends to be done in an ad hoc manner, favoring empirical and reactive approaches, rather than a model-driven and analytical approach. Such empirical strategies may arrive at an approximate schedule for the dataflow that needs further tuning to meet the quality of service. We propose a model-based scheduling approach that makes use of performance profiles and benchmarks developed for tasks in the dataßow to plan both the resource allocation and the resource mapping that together form the schedule planning process. We propose the Model Based Allocation (MBA) and the Slot Aware Mapping (SAM) approaches that efectively utilize knowledge of the performance model of logic tasks to provide an efficient and predictable scheduling behavior. We implemented and validate these algorithms using the popular open source Apache Storm DSPS for several micro and application dataflows. The results show that our model-driven approach is able to reduce the amount of required resources (VMs) by 30% − 50% relative to existing techniques. Also we see that our strategies o↵er a predictable behavior that ensures that the expected and actual rates supported and resources used match closely. This can enable deterministic schedule planning even under dynamic conditions. Besides this static scheduling, we also examine the ability to dynamically consolidate tasks onto fewer VMs when the load on the dataßow decreases or the VMs get fragmented. We propose reliable task migration models for Apache Storm dataßows that are able to rapidly move the task assignment in the cluster, and resume the dataflow execution without any message loss.

abstract: Stream processing has emerged as an important model of computation especially in the context of multimedia and communication sub-systems of embedded System-on-Chip (SoC) architectures. The dataflow nature of streaming applications allows them to be most naturally expressed as a set of kernels iteratively operating on continuous streams of data. The kernels are computationally intensive and are mainly characterized by real-time constraints that demand high throughput and data bandwidth with limited global data reuse. Conventional architectures fail to meet these demands due to their poorly matched execution models and the overheads associated with instruction and data movements. This work presents StreamWorks, a multi-core embedded architecture for energy-efficient stream computing. The basic processing element in the StreamWorks architecture is the StreamEngine (SE) which is responsible for iteratively executing a stream kernel. SE introduces an instruction locking mechanism that exploits the iterative nature of the kernels and enables fine-grain instruction reuse. Each instruction in a SE is locked to a Reservation Station (RS) and revitalizes itself after execution; thus never retiring from the RS. The entire kernel is hosted in RS Banks (RSBs) close to functional units for energy-efficient instruction delivery. The dataflow semantics of stream kernels are captured by a context-aware dataflow execution mode that efficiently exploits the Instruction Level Parallelism (ILP) and Data-level parallelism (DLP) within stream kernels. Multiple SEs are grouped together to form a StreamCluster (SC) that communicate via a local interconnect. A novel software FIFO virtualization technique with split-join functionality is proposed for efficient and scalable stream communication across SEs. The proposed communication mechanism exploits the Task-level parallelism (TLP) of the stream application. The performance and scalability of the communication mechanism is evaluated against the existing data movement schemes for scratchpad based multi-core architectures. Further, overlay schemes and architectural support are proposed that allow hosting any number of kernels on the StreamWorks architecture. The proposed oevrlay schemes for code management supports kernel(context) switching for the most common use cases and can be adapted for any multi-core architecture that use software managed local memories. The performance and energy-efficiency of the StreamWorks architecture is evaluated for stream kernel and application benchmarks by implementing the architecture in 45nm TSMC and comparison with a low power RISC core and a contemporary accelerator.<br>Dissertation/Thesis<br>Doctoral Dissertation Computer Science 2014