Lynceus: Cost-efficient Tuning and Provisioning of Data Analytic Jobs

Casimiro, Maria Helena; Didona, Diego; Romano, Paolo; Rodrigues, L.; Zwaenepoel, Willy; Garlan, David

doi:10.1109/icdcs47774.2020.00047

Cited by 23 publications

(16 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Several research works target automatic cloud configuration optimization. Cherrypick [6], Arrow [27], Scout [28], Micky [26], Vanir [12] and Lynceus [16] perform cloud configuration optimization for distributed data analytics frameworks such as Spark and Hadoop. Ernest [41] creates an analytical model for Spark applications and uses that to optimize cloud configurations.…”

Section: Related Workmentioning

confidence: 99%

With Great Freedom Comes Great Opportunity: Rethinking Resource Allocation for Serverless Functions

Bilal¹,

Canini²,

Fonseca³

et al. 2021

Preprint

View full text Add to dashboard Cite

Current serverless offerings give users a limited degree of flexibility for configuring the resources allocated to their function invocations by either coupling memory and CPU resources together or providing no knobs at all. These configuration choices simplify resource allocation decisions on behalf of users, but at the same time, create deployments that are resource inefficient.In this paper, we take a principled approach to the problem of resource allocation for serverless functions, allowing this choice to be made in an automatic way that leads to the best combination of performance and cost. In particular, we systematically explore the opportunities that come with decoupling memory and CPU resource allocations and also enabling the use of different VM types. We find a rich trade-off space between performance and cost. The provider can use this in a number of ways: from exposing all these parameters to the user, to eliciting preferences for performance and cost from users, or by simply offering the same performance with lower cost. This flexibility can also enable the provider to optimize its resource utilization and enable a cost-effective service with predictable performance.Our results show that, by decoupling memory and CPU allocation, there is potential to have up to 40% lower execution cost than the preset coupled configurations that are the norm in current serverless offerings. Similarly, making the correct choice of VM instance type can provide up to 50% better execution time. Furthermore, we demonstrate that providers can utilize different instance types for the same functions to maximize resource utilization while providing performance within 10-20% of the best resource configuration for each respective function.

show abstract

Section: Related Workmentioning

confidence: 99%

With Great Freedom Comes Great Opportunity: Rethinking Resource Allocation for Serverless Functions

Bilal¹,

Canini²,

Fonseca³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…load balancing, increase of output quality). We should note here the importance of cost models [3,9] for the evaluation of alternative plans, which play the role of a Knowledge Base and are continuously augmented with historical data from monitoring and reaction to past events.…”

Section: Aiops For Elastic Cloud To Edge Intelligencementioning

confidence: 99%

MEDAL: An AI-driven Data Fabric Concept for Elastic Cloud-to-Edge Intelligence

Theodorou¹,

Gerostathopoulos²,

Alshabani³

et al. 2021

Preprint

View full text Add to dashboard Cite

Current Cloud solutions for Edge Computing are inefficient for datacentric applications, as they focus on the IaaS/PaaS level and they miss the data modeling and operations perspective. Consequently, Edge Computing opportunities are lost due to cumbersome and data assets-agnostic processes for end-to-end deployment over the Cloud-to-Edge continuum. In this paper, we introduce MEDALan intelligent Cloud-to-Edge Data Fabric to support Data Operations (DataOps) across the continuum and to automate management and orchestration operations over a combined view of the data and the resource layer. MEDAL facilitates building and managing data workflows on top of existing flexible and composable data services, seamlessly exploiting and federating IaaS/PaaS/SaaS resources across different Cloud and Edge environments. We describe the MEDAL Platform as a usable tool for Data Scientists and Engineers, encompassing our concept and we illustrate its application though a connected cars use case

show abstract

“…The problem is further exacerbated by the fact that the (distributed) training process of ML jobs exposes several hyper-parameters -such as the batch size considered in each training iteration or the frequency of synchronization among workers. The optimal configuration for these parameters can be substantially affected by the choice of the type and number of provisioned cloud resources [3].…”

Section: Introductionmentioning

confidence: 99%

“…Given the complexity of modelling the dynamics of modern ML and cloud platforms via white-box methods, a common approach in the literature is to rely on black-box modelling and Bayesian Optimization (BO) techniques [3]- [5]. These techniques have the key advantage of requiring no prior knowledge of the target ML model to be optimized, and as such require the target ML model to be deployed and trained in several (cloud/hyper-parameter) configurations.…”

Section: Introductionmentioning

confidence: 99%

“…From a practical perspective, when compared to state of the art BO-based systems for the (constrained) optimization of ML jobs, such as Lynceus [3] or Cherrypick [4], TrimTuner reduces the cost of the exploration process by up to 50× thanks to the use of sub-sampling techniques. Another practical contribution of this work is making available to the community the data-sets obtained for the evaluation of TrimTuner, which consider the training of three neural networks via TensorFlow on AWS EC2 over a search space (encompassing both model hyper-parameters and cloud-related parameters) composed of approximately 1400 configurations, whose collection incurred a cost of approximately $1200 and took about two months.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

TrimTuner: Efficient Optimization of Machine Learning Jobs in the Cloud via Sub-Sampling

Mendes¹,

Casimiro²,

Romano³

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

This work introduces TrimTuner, the first system for optimizing machine learning jobs in the cloud to exploit sub-sampling techniques to reduce the cost of the optimization process, while keeping into account user-specified constraints. TrimTuner jointly optimizes the cloud and application-specific parameters and, unlike state of the art works for cloud optimization, eschews the need to train the model with the full training set every time a new configuration is sampled. Indeed, by leveraging sub-sampling techniques and data-sets that are up to 60× smaller than the original one, we show that TrimTuner can reduce the cost of the optimization process by up to 50×.Further, TrimTuner speeds-up the recommendation process by 65× with respect to state of the art techniques for hyperparameter optimization that use sub-sampling techniques. The reasons for this improvement are twofold: i) a novel domain specific heuristic that reduces the number of configurations for which the acquisition function has to be evaluated; ii) the adoption of an ensemble of decision trees that enables boosting the speed of the recommendation process by one additional order of magnitude.

show abstract

Lynceus: Cost-efficient Tuning and Provisioning of Data Analytic Jobs

Cited by 23 publications

References 27 publications

With Great Freedom Comes Great Opportunity: Rethinking Resource Allocation for Serverless Functions

With Great Freedom Comes Great Opportunity: Rethinking Resource Allocation for Serverless Functions

MEDAL: An AI-driven Data Fabric Concept for Elastic Cloud-to-Edge Intelligence

TrimTuner: Efficient Optimization of Machine Learning Jobs in the Cloud via Sub-Sampling

Contact Info

Product

Resources

About