Accelerating Serverless Computing by Harvesting Idle Resources

Yu, Hanfei; Wang, Hao; Li, Jian; Xu, Yangsheng; Park, Seung-Jong

doi:10.1145/3485447.3511979

Cited by 7 publications

(4 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Instead, serverless computing executes tasks with lightweight containers, thus allowing fine-grained resource provisioning with instant function launch/release, which charges users by the amount of resources (e.g., CPU/GPU and memory) only in actual execution (e.g., second). Due to the unique features, serverless computing is particularly appealing for tasks that require elasticity and high concurrency, such as scientific computing (Chard et al 2020;Roy et al 2022) and distributed training (Wang, Niu, and Li 2019; Guo et al 2022;Thorpe et al 2021;Yu et al 2021Yu et al , 2022.…”

Section: Serverless Drl Trainingmentioning

confidence: 99%

“…Unlike physical clusters and traditional cloud computing that require tedious configuration, serverless computing packages and executes tasks (e.g., DRL actors and learner) as functions with instant toggling (i.e., sub-second level) and auto-scaling. Thus, serverless computing has been widely deployed to serve computation-intensive applications, such as deep learning (Ali et al 2020;Carreira et al 2019;Wang, Niu, and Li 2019;Yu et al 2021Yu et al , 2022 and scientific computing (Chard et al 2020;Roy et al 2022). Fig.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Cheaper and Faster: Distributed Deep Reinforcement Learning with Serverless Computing

Yu,

Li,

Hua

et al. 2024

AAAI

View full text Add to dashboard Cite

Deep reinforcement learning (DRL) has gained immense success in many applications, including gaming AI, robotics, and system scheduling. Distributed algorithms and architectures have been vastly proposed (e.g., actor-learner architecture) to accelerate DRL training with large-scale server-based clusters. However, training on-policy algorithms with the actor-learner architecture unavoidably induces resource wasting due to synchronization between learners and actors, thus resulting in significantly extra billing. As a promising alternative, serverless computing naturally fits on-policy synchronization and alleviates resource wasting in distributed DRL training with pay-as-you-go pricing. Yet, none has leveraged serverless computing to facilitate DRL training. This paper proposes MinionsRL, the first serverless distributed DRL training framework that aims to accelerate DRL training- and cost-efficiency with dynamic actor scaling. We prototype MinionsRL on top of Microsoft Azure Container Instances and evaluate it with popular DRL tasks from OpenAI Gym. Extensive experiments show that MinionsRL reduces total training time by up to 52% and training cost by 86% compared to latest solutions.

show abstract

Section: Serverless Drl Trainingmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Cheaper and Faster: Distributed Deep Reinforcement Learning with Serverless Computing

Yu,

Li,

Hua

et al. 2024

AAAI

View full text Add to dashboard Cite

show abstract

“…This is seen as an alternative or used in conjunction with horizontal scaling in order to meet intended targets, in the face of changing traffic levels. An actor critic architecture with Proximal Policy Optimization (PPO) is used in [25] to harvest idle resources from functions and direct them to under-provisioned instances. A Q-Learning based solution is given in [26] to identify the level of concurrency, i.e the number of concurrent requests served per instance, to optimize function latency and system throughput.…”

Section: A Serverless Resource Scalingmentioning

confidence: 99%

Deep reinforcement learning for application scheduling in resource-constrained, multi-tenant serverless computing environments

Mampage

Karunasekera

Buyya

2023

Future Generation Computer Systems

View full text Add to dashboard Cite

“…For example, Zafeiropoulos et al [14] proposed an approach in which autoscaling is assisted with a DQL agent that trains in an environment with continuous state space and discrete action space. Meanwhile, the PPO agent in [42] learns to make resource adjustments per invocation based on the realistic serverless environment. Furthermore, Qiu et al [43] customize the implementation of PPO to fit the multi-agent training approach, while the TD A2C agent in [44] predicts the future idle container window by learning the past invocation patterns of serverless functions to mitigate the cold start problem.…”

Section: Ai-based Techniquesmentioning

confidence: 99%

Autoscaling in Serverless Computing : Taxonomy and OpenChallenges

Jawaddi

Ismail

2023

Preprint

View full text Add to dashboard Cite

The popularity of serverless computing has been fueled by its operational simplicity, pay-per-use pricing model, and the ability to autoscale. However, there is a lack of comprehensive reviews that focus on the autoscaling context in serverless computing. In this paper, we address this gap by proposing a taxonomy of autoscaling properties for serverless computing. To gather relevant information, we review recent contributions on autoscaling in serverless computing from 2018 to 2022. Using the proposed taxonomy, we analyze the existing autoscaling solutions. Our analysis reveals that the scaling objectives explored by researchers are limited to certain elements, and the existing serverless autoscaling approaches do not provide guarantees that the scaling policies or strategies can meet the Service Level Agreement (SLA) requirements. We conclude by recommending open challenges from three perspectives: verification of autoscaling, energy-driven autoscaling, and anomaly-aware autoscaling. These challenges highlight the need for future research to address the limitations of existing autoscaling approaches and provide more robust and reliable autoscaling mechanisms for serverless computing.

show abstract

Accelerating Serverless Computing by Harvesting Idle Resources

Cited by 7 publications

References 21 publications

Cheaper and Faster: Distributed Deep Reinforcement Learning with Serverless Computing

Cheaper and Faster: Distributed Deep Reinforcement Learning with Serverless Computing

Deep reinforcement learning for application scheduling in resource-constrained, multi-tenant serverless computing environments

Autoscaling in Serverless Computing : Taxonomy and OpenChallenges

Contact Info

Product

Resources

About