this paper introduces the Aster distributed configuration-based programming system that is aimed at easing the development of emerging distributed applications having quality of service requirements. Our approach is b...
详细信息
ISBN:
(纸本)0818673982
this paper introduces the Aster distributed configuration-based programming system that is aimed at easing the development of emerging distributed applications having quality of service requirements. Our approach is based on high-level customization: given the specification of application requirements using the Aster interconnection language, a distributed runtime system, customerized for meeting these requirements is built. So as to make the customization process sound, we propose a formal method that allows to reason about specification matching of a customized distributed runtime system withthe application's requirements.
We claim in this paper that both remote process creation and process migration are efficient mechanisms to be used in the improvement or development of high performance computer systems. In particular, we demonstrate ...
详细信息
ISBN:
(纸本)0818673982
We claim in this paper that both remote process creation and process migration are efficient mechanisms to be used in the improvement or development of high performance computer systems. In particular, we demonstrate that the claims made by some researchers that process migration is too heavy to be used to support dynamic load balancing are unsubstantiated. We support our claim by presenting these two mechanisms available in the RHODOS distributed operating system, comparing and contrasting these mechanisms and reporting on their performance.
Neural enhancement through super-resolution (SR) deep neural networks (DNNs) opens up new possibilities for ultra-high-definition (UHD) live streaming. Yet, the heavy SR DNN inference overhead leads to severe deployme...
详细信息
ISBN:
(纸本)9798400714672
Neural enhancement through super-resolution (SR) deep neural networks (DNNs) opens up new possibilities for ultra-high-definition (UHD) live streaming. Yet, the heavy SR DNN inference overhead leads to severe deployment challenges. To reduce the overhead, existing systems propose to apply DNN-based SR only on carefully selected anchor frames while upscaling non-anchor frames via the lightweight reusing-based SR approach. However, frame-level scheduling is coarse-grained and fails to deliver optimal efficiency. In this work, we propose Palantír, the first neural-enhanced UHD live streaming system with fine-grained patch-level *** the core of Palantír is its SR video quality estimation strategy which guides the low-delay selection of the most beneficial anchor patches. Although existing systems propose estimation strategies for anchor frame selection, these strategies heavily rely on empirical insights that cannot be transferred to our context, making fine-grained scheduling a challenging problem that requires a fundamentally new solution. Facing the challenge, we follow the first-principles approach and derive a directed acyclic graph (DAG) model to address the problem. the model can also be generalized to various settings due to its first-principles nature. Compared to the state-of-the-art real-time frame-level scheduling strategy for live streaming, Palantír reduces the anchor size by 80.1% at most and 38.4% on average without compromising the quality gain. Furthermore, Palantír incurs a scheduling latency accounting for only 0.6-3.9% of the end-to-end latency requirement for UHD live streaming.
In this research, we explore the technical and computational merits of a machine learning algorithm on a large data set, employing distributedsystems. Using 167 million (10 GB) energy consumption observations collect...
详细信息
ISBN:
(纸本)9781538666142
In this research, we explore the technical and computational merits of a machine learning algorithm on a large data set, employing distributedsystems. Using 167 million (10 GB) energy consumption observations collected by smart meters from residential consumers in London, England, we predict future residential energy consumption using a Random Forest machine learning algorithm. distributedsystems such as AWS S3 and EMR, MongoDB and Apache Spark are used. Computational times and predictive accuracy are evaluated. We conclude that there are significant computational advantages to using distributedsystems when applying machine learning algorithms on large-scale data. We also observe that distributedsystems can be computationally burdensome when the amount of data being processed is below a threshold at which it can leverage the computational efficiencies provided by distributedsystems.
In many distributed real-time systems, the workload can be modeled as a set of periodic tasks, each of which consists of a chain of subtasks executing on different processors. Synchronization protocols are used to gov...
详细信息
ISBN:
(纸本)0818673982
In many distributed real-time systems, the workload can be modeled as a set of periodic tasks, each of which consists of a chain of subtasks executing on different processors. Synchronization protocols are used to govern the release of subtasks so that the precedence constraints among subtasks are satisfied and the schedulability of the resultant system is analyzable. Tasks have different worst-case and average end-to-end response times when different protocols are used. In this paper, we consider distributed real-time systems with independent, periodic tasks and fixed-priority scheduling algorithms. We propose three synchronization protocols and conduct simulation to compare their performance with respect to the two timing aspects.
We consider a shared store based on distributed shared memory (DSM), supporting persistence by reachability (PBR), a very simple data sharing model for a distributed system. this DSM+PBR model is based on distributed ...
详细信息
ISBN:
(纸本)0818673982
We consider a shared store based on distributed shared memory (DSM), supporting persistence by reachability (PBR), a very simple data sharing model for a distributed system. this DSM+PBR model is based on distributed garbage collection (GC). Within a general model for DSM+PBR, we specify a distributed GC algorithm that is efficient and scalable. Its main features are: (i) independent collection of memory subsets (even when replicated), (ii) orthogonal from coherence, (iii) asynchrony, and (iv) a simple heuristic to collect cycles avoiding extra I/O costs. We briefly describe our implementation and show some performance results.
High Performance computingsystems are moving heavily towards many-core processors with a deep hierarchy of memory. Accelerators like GPUs are widely being used for general purpose computing and processor architecture...
详细信息
ISBN:
(纸本)9781538666142
High Performance computingsystems are moving heavily towards many-core processors with a deep hierarchy of memory. Accelerators like GPUs are widely being used for general purpose computing and processor architectures are becoming increasingly complex to accommodate performance boost. this trend towards complex heterogeneous architecture makes the job of scientific application developers difficult in terms of performance, portability and productivity. With memory being distributed, this challenge becomes even more complex. Programming many-core shared memory systems are most widely accomplished using OpenMP, while MPI is used to manage the communications in a distributed system. Even though MPI provides a rich set of features, it is too explicit making users responsible for overlapping communication and computation. Task-based runtime systems have emerged as a solution to this challenge of programming these modern complex systems. this study surveys the landscape of task-based runtime systemsthat support distributed memory and presents a set of benchmark for evaluating and understanding runtime-performance and overheads of these systems.
We address the problem of how to handle exceptions in distributed object-oriented systems. In a distributedcomputing environment exceptions may be raised simultaneously and thus need to be treated in a coordinated ma...
详细信息
ISBN:
(纸本)0818673982
We address the problem of how to handle exceptions in distributed object-oriented systems. In a distributedcomputing environment exceptions may be raised simultaneously and thus need to be treated in a coordinated manner. We take two kinds of concurrency into account: 1) several objects are designed collectively and invoked concurrently to achieve a global goal, and 2) concurrent objects or object groups that are designed independently compete for the same system resources. We propose a new distributed algorithm for resolving concurrent exceptions and show that the algorithm works correctly even in complex nested situations, and is an improvement over previous proposals in that it requires only O(N2) messages, and is fully object-oriented.
暂无评论