One of the most important issues in current high-performance packet switches is the availability of efficient algorithms to maximize instantaneous throughput. (D)PHM and iSLIP are well-known algorithms for virtual out...
详细信息
ISBN:
(纸本)9781424407545
One of the most important issues in current high-performance packet switches is the availability of efficient algorithms to maximize instantaneous throughput. (D)PHM and iSLIP are well-known algorithms for virtual output-queued switches. In this paper, a comparative study of the implementation of both types of schedulers in different families of FPGAs is presented. Experimental results show that, in addition to the well known advantages of using field-programmable logic, the proposed implementations provide a performance-complexity trade-off which makes them a suitable practical alternative for high-performance scheduling tasks.
Parallel applications that present tasks with imbalanced loads or complex communication behavior usually do not exploit the underlying resources of parallel platforms to their full potential. In order to mitigate this...
详细信息
Parallel applications that present tasks with imbalanced loads or complex communication behavior usually do not exploit the underlying resources of parallel platforms to their full potential. In order to mitigate this issue, global scheduling algorithms are employed. As finding the optimal task distribution is an NP-Hard problem, identifying the most suitable algorithm for a specific scenario and comparing algorithms are not trivial tasks. In this context, this paper presents COMPREHENSIVEBENCH, a benchmark for global scheduling algorithms that enables the variation of a vast range of parameters that affect performance. COMPREHENSIVEBENCH can be used to assist in the development and evaluation of new scheduling algorithms, to help choose a specific algorithm for an arbitrary application, to emulate other applications, and to enable statistical tests. We illustrate its use in this paper with an evaluation of Charm++ periodic load balancers that stresses their characteristics.
A significant amount of energy is consumed to render high-level computation tasks in large scale cloud computing applications. The state-of-the-art energy saving techniques based on centralized job placement approache...
详细信息
ISBN:
(纸本)9780769550886
A significant amount of energy is consumed to render high-level computation tasks in large scale cloud computing applications. The state-of-the-art energy saving techniques based on centralized job placement approaches reduce the reliability of operation due to a single point of failure. Moreover, the existing works do not consider energy consumption cost for communication devices and network appliances which contribute a lot. In this paper, we have proposed a mechanism for cluster formation based on network vicinity among the data servers. We have developed two distributed and localized intra-cluster and inter-cluster VM scheduling algorithms based on energy calculation, resource requirement and availability. Our proposed scheduling algorithms manage VMs to reduce the energy consumption of both the servers and networking devices. Simulation results show that our proposed distributed VM scheduling algorithms can conserve significant amount of energy compared to state-of-the-art works.
An increasing demand for high-performance systems has been observed in the domain of both general purpose and real-time systems, pushing the industry towards a pervasive transition to multi-core platforms. Unfortunate...
详细信息
ISBN:
(纸本)9781479939541
An increasing demand for high-performance systems has been observed in the domain of both general purpose and real-time systems, pushing the industry towards a pervasive transition to multi-core platforms. Unfortunately, well-known and efficient scheduling results for single-core systems do not scale well to the multi-core domain. This justifies the adoption of more computationally intensive algorithms, but the complexity and computational overhead of these algorithms impact their applicability to real OSes. We propose an architecture to migrate the burden of multicore scheduling to a dedicated hardware component. We show that it is possible to mitigate the overhead of complex algorithms, while achieving power efficiency and optimizing processors utilization. We develop the idea of "active monitoring" to continuously track the evolution of scheduling parameters as tasks execute on processors. This allows reducing the gap between implementable scheduling techniques and the ideal fluid scheduling model, under the constraints of realistic hardware.
In this paper, we consider a multi-user optical grid environment where the optical grid network is shared among multiple users who have a high demand for data processing. In such multi-user grid environment, since the...
详细信息
ISBN:
(纸本)9781424436880
In this paper, we consider a multi-user optical grid environment where the optical grid network is shared among multiple users who have a high demand for data processing. In such multi-user grid environment, since the supercomputer can start the data processing only after the corresponding data file has reached it, one critical problem is how to effectively schedule large data file transfers on the optical grid network. In this paper, we propose four dynamic scheduling algorithms to address the issue of large file transfers on multi-user optical grid network. Two objectives, namely the efficiency and fairness, are considered to evaluate the performance of each scheduling algorithm. Efficiency aims at how to achieve the best overall quality of grid service, and fairness regards the service quality deviation among grid users. We study the performance of each algorithm, and a tradeoff between efficiency and fairness is shown by our results.
When considering issues of fairness of scheduling algorithms for packet data in networks, there is no well defined quantitative value for the 'fairness" of the system. Rather, fairness is usually expressed in...
详细信息
ISBN:
(纸本)0780375149
When considering issues of fairness of scheduling algorithms for packet data in networks, there is no well defined quantitative value for the 'fairness" of the system. Rather, fairness is usually expressed in broader terms. Generally, a system is deemed to be fair if it meets certain criteria on throughput or delay, and unfair if the criteria are not met. For example, a system in which some user receives a throughput less than X bits/sec or experiences a delay T with probability greater than p% may be deemed to be unfair. These definitions of fairness say if the system is or is not fair, but not how fair or unfair. This paper attempts to make a quantitative definition for a value of 'fairness" that makes sense from both a mathematical and semantic standpoint. Such a value could be used to quickly compare the fairness of differing systems or algorithms. Definitions for the self-fairness and self-unfairness of the users of the system, and the average fairness and average unfairness of the overall system are proposed. The case is first considered when all users are weighted equally. That is, every user is considered to have the same importance, and hence each user should receive an equal proportion of the allocated resources for the system to be fair. The definitions are then extended to the case when the users have different weightings (for example, to achieve different levels of Quality of Service). The definitions take the viewpoint that the users themselves control what resources they are allocated. However, the definitions can just as easily be applied to the case when a central scheduling algorithm determines the allocation of resources to the users.
This work studies online scheduling algorithms for buffer management, develops new algorithms, and analyzes their performances. Packets arrive at a release time r, with a non-negative weight w and an integer deadline ...
详细信息
ISBN:
(数字)9783319388519
ISBN:
(纸本)9783319388502;9783319388519
This work studies online scheduling algorithms for buffer management, develops new algorithms, and analyzes their performances. Packets arrive at a release time r, with a non-negative weight w and an integer deadline d. At each time step, at most one packet is scheduled. The modified greedy (MG) algorithm is 1.618-competitive for the objective of maximizing the sum of weights of packets sent, assuming agreeable deadlines. We analyze the empirical behavior of MG in a situation with arbitrary deadlines and demonstrate that it is at a disadvantage when frequently preferring maximum weight packets over early deadline ones. We develop the MLP algorithm, which remedies this problem whilst mimicking the behavior of the offline algorithm. Our comparative analysis shows that, although the competitive ratio of MLP is not as good as that of MG, it performs better in practice. We validate this by simulating the behavior of both algorithms under a spectrum of parameter settings. Finally, we propose the design of three additional algorithms, which may help in improving performance in practice.
Most existing wireless networking solutions are best-effort and do not provide any delay guarantee required by important applications such as the control traffic of cyber-physical systems. Recently, Hou and Kumar prov...
详细信息
ISBN:
(纸本)9781467399531
Most existing wireless networking solutions are best-effort and do not provide any delay guarantee required by important applications such as the control traffic of cyber-physical systems. Recently, Hou and Kumar provided the first framework for analyzing and designing delay-guaranteed network solutions. While inspiring, their idle-time-based analysis appears to apply only to flows with a special traffic (arrival and expiration) pattern, and the problem remains largely open for general traffic patterns. This paper addresses this challenge by proposing a new framework that characterizes and achieves the complete delay-constrained capacity region with general traffic patterns in singlehop downlink access-point wireless networks. We first formulate the timely capacity problem as an infinite-horizon Markov Decision Process (MDP) and then judiciously combine different simplification methods to convert it to an equivalent finite-size linear program (LP). This allows us to characterize the timely capacity region of flows with general traffic patterns for the first time in the literature. We then design three timely-flow scheduling algorithms for general traffic patterns. The first algorithm achieves the optimal utility but suffers from the curse of dimensionality. The second and third algorithms are inspired by our MDP framework and are of polynomial-time complexity. Simulation results show that both achieve near-optimal performance and outperform other existing alternatives.
The paper presents comparison of the two scheduling algorithms developed for program structurization for execution in dynamic SMP clusters implemented in Systems on Chip (SoC) technology SoC modules are built of a set...
详细信息
ISBN:
(纸本)9783642144028
The paper presents comparison of the two scheduling algorithms developed for program structurization for execution in dynamic SMP clusters implemented in Systems on Chip (SoC) technology SoC modules are built of a set of processors, memory modules and a multi- bus interconnection network A set of such SoCs is interconnected by a global communication network Inter-processor communication inside SoC modules uses a novel technique of data transfers on the fly The algorithms present two different scheduling approaches The first uses ETF-based genetically supported list scheduling heuristics to map nodes of a program to processors The second is a clustering-based algorithm using Moldable Tasks (MT) to structure the graph Both algorithms structure computations and local data transfers to introduce processor switching and data transfers on the fly The algorithms were tested using a set of automatically generated parameterized program graphs The results were compared to results obtained using a classic ETF-based list scheduling without data transmissions on the fly
This paper compares the performance of different scheduling algorithms, with and without void filling, considering a complete Optical Burst Switched (OBS) network scenario, given by edge and core routers, under real t...
详细信息
ISBN:
(纸本)9780780397460
This paper compares the performance of different scheduling algorithms, with and without void filling, considering a complete Optical Burst Switched (OBS) network scenario, given by edge and core routers, under real traffic conditions. Performance is investigated taking into account both burst loss probability and the complexity of the algorithms, evaluated in terms of scheduling time. For an effective comparison both aspects have to be considered in order to understand if an algorithm which provides low burst blocking probabilities is actually feasible or not. In this paper the authors propose a scheduling algorithm which represents a good trade-off between burst blocking performance and scheduling time.
暂无评论