Need for speed: CORA scheduler for optimizing completion-times in the cloud

Huang, Zhe; Balasubramanian, Bharath; Wang, Michael; Lan, Tian; Chiang, Mung; Tsang, D.H.K.

doi:10.1109/infocom.2015.7218460

Cited by 47 publications

(25 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Lett i be the completion time slot of job i. Each job i has a non-negative utility f i (t i − a i ), non-increasing witht i − a i , specifying the job's value in terms of different completion times [23] [35]. The offline optimization problem to maximize overall utility is formulated as follows.…”

Section: Offline Optimization Problemmentioning

confidence: 99%

“…For different jobs, E i is set within [50,200], N i is in [5,100], M i is in [10,100], τ i is in [0.001, 0.1] time slots, and e i is within [30,575]MB [6]. We use a sigmoid utility function [23], f i (t − a i ) = γ1 1+e γ 2 (t−a i −γ 3 ) , where γ 1 is priority of job i in [1,100], γ 2 is a decay factor, and γ 3 is the target completion time of job i set in [1,15]. We set γ 2 = 0 for time-insensitive jobs (constant utility), γ 2 in [0.01, 1] to [13] and Mesos [16]; jobs are all admitted and numbers of workers/parameter servers are computed to achieve max-min fairness in dominant resources upon job arrival and job completion [15].…”

Section: A Simulation Studiesmentioning

confidence: 99%

See 1 more Smart Citation

Online Job Scheduling in Distributed Machine Learning Clusters

Bao

Peng

et al. 2018

IEEE INFOCOM 2018 - IEEE Conference on Computer Communications

102

View full text Add to dashboard Cite

Nowadays large-scale distributed machine learning systems have been deployed to support various analytics and intelligence services in IT firms. To train a large dataset and derive the prediction/inference model, e.g., a deep neural network, multiple workers are run in parallel to train partitions of the input dataset, and update shared model parameters. In a shared cluster handling multiple training jobs, a fundamental issue is how to efficiently schedule jobs and set the number of concurrent workers to run for each job, such that server resources are maximally utilized and model training can be completed in time. Targeting a distributed machine learning system using the parameter server framework, we design an online algorithm for scheduling the arriving jobs and deciding the adjusted numbers of concurrent workers and parameter servers for each job over its course, to maximize overall utility of all jobs, contingent on their completion times. Our online algorithm design utilizes a primal-dual framework coupled with efficient dual subroutines, achieving good long-term performance guarantees with polynomial time complexity. Practical effectiveness of the online algorithm is evaluated using trace-driven simulation and testbed experiments, which demonstrate its outperformance as compared to commonly adopted scheduling algorithms in today's cloud systems.

show abstract

Section: Offline Optimization Problemmentioning

confidence: 99%

Section: A Simulation Studiesmentioning

confidence: 99%

Online Job Scheduling in Distributed Machine Learning Clusters

Bao

Peng

et al. 2018

IEEE INFOCOM 2018 - IEEE Conference on Computer Communications

102

View full text Add to dashboard Cite

show abstract

“…By varying θ, the proposed joint optimization can generate a wide range of solutions for diverse application scenarios, ranging from PoCD-critical optimization with small tradeoff factor θ and cost-sensitive optimization with large θ. Finally, while our joint optimization framework applies to any concave, increasing utility function f , in this paper, we focus on logarithmic utility functions, f (R(r)−R min ) = lg(R(r)−R min ), which is known to achieve proportional fairness [60]. In the following, we will prove the concavity of the optimization objective U (r) for different strategies and propose an efficient algorithm to find the optimal solution to the proposed joint PoCD and cost optimization.…”

Section: Comparing Pocd Of Different Strategiesmentioning

confidence: 99%

Chronos: A Unifying Optimization Framework for Speculative Execution of Deadline-Critical MapReduce Jobs

Alamro

Lan

et al. 2018

2018 IEEE 38th International Conference on Distributed Computing Systems (ICDCS)

Self Cite

View full text Add to dashboard Cite

Meeting desired application deadlines in cloud processing systems such as MapReduce is crucial as the nature of cloud applications is becoming increasingly mission-critical and deadline-sensitive. It has been shown that the execution times of MapReduce jobs are often adversely impacted by a few slow tasks, known as stragglers, which result in high latency and deadline violations. While a number of strategies have been developed in existing work to mitigate stragglers by launching speculative or clone task attempts, none of them provides a quantitative framework that optimizes the speculative execution for offering guaranteed Service Level Agreements (SLAs) to meet application deadlines. In this paper, we bring several speculative scheduling strategies together under a unifying optimization framework, called Chronos, which defines a new metric, Probability of Completion before Deadlines (PoCD), to measure the probability that MapReduce jobs meet their desired deadlines. We systematically analyze PoCD for popular strategies including Clone, Speculative-Restart, and Speculative-Resume, and quantify their PoCD in closed-form. The result illuminates an important tradeoff between PoCD and the cost of speculative execution, measured by the total (virtual) machine time required under different strategies. We propose an optimization problem to jointly optimize PoCD and execution cost in different strategies, and develop an algorithmic solution that is guaranteed to be optimal. Chronos is prototyped on Hadoop MapReduce and evaluated against three baseline strategies using both experiments and trace-driven simulations, achieving 50% net utility increase with up to 80% PoCD and 88% cost improvements.

show abstract

“…The objective is to maximize the probability that tasks can be executed within their delay requirements. In [14], a job scheduling scheme in the cloud computing clusters considering job resource requirements and completion time sensitivity is proposed. The problem is formulated as a maximization of the minimum utility achieved across all the jobs in the cluster, where the job utilities are functions of their completion times.…”

Section: Related Workmentioning

confidence: 99%

A strategy for joint service offloading and scheduling in heterogeneous cloud radio access networks

Chabbouh¹,

Rejeb²,

Choukair³

et al. 2017

J Wireless Com Network

View full text Add to dashboard Cite

Cloud radio access network is one of the most promising cellular networks for the next generation of mobile networks. The basic idea of cloud RAN (radio access network) is virtualizing and centralizing the intelligent part of the base station, the base band unit, and keeping remote radio heads on cell site enabling a centralized processing and management. Offloading data computation to edge cloud was proposed as a solution to deal with resource limitation while keeping a good quality of service. In this paper, we propose a strategy to jointly handle offloading decision and offloading request scheduling in cloud RAN. We aim to improve network quality of service while reducing the scheduling cost expressed in terms of overload, network delay, and migration cost. Numerical results show that the proposed approach is able to reduce the response time of the applications, mobile terminal energy consumption, and total execution cost.

show abstract

Need for speed: CORA scheduler for optimizing completion-times in the cloud

Cited by 47 publications

References 18 publications

Online Job Scheduling in Distributed Machine Learning Clusters

Online Job Scheduling in Distributed Machine Learning Clusters

Chronos: A Unifying Optimization Framework for Speculative Execution of Deadline-Critical MapReduce Jobs

A strategy for joint service offloading and scheduling in heterogeneous cloud radio access networks

Contact Info

Product

Resources

About