Energy-efficient deep reinforcement learning-based application scheduling for serverless computing environments

Satender Kumar

doi:10.1007/s11227-026-08349-0

Return

Energy-efficient deep reinforcement learning-based application scheduling for serverless computing environments

2026-03-07

PRE

Satender Kumar *

Sarika Jain

Ashutosh Kumar Singh

DOI：10.1007/s11227-026-08349-0

Original

Original request for help

Save

Abstract

En 中文

Serverless is an emerging cloud subset that enables users to develop and execute their code without the overhead of managing servers. Due to its ease of use, scalability, and resource effectiveness, it has gained a lot of attention. Serverless acts as Function-as-a-Service (FaaS) which includes only creation and deployment of functions based on which an app is executed. The key characteristics of the serverless paradigm are fine-grained billing, inherently scaling and which allows quick deployment of code. FaaS is replacing Infrastructure-as-a-Service (IaaS) with an exponential rate due to its widespread merits. Function scheduling plays a vital role in how efficiently a platform operates, yet it has received relatively limited attention in existing research. Balancing energy efficiency with reliable function performance is particularly challenging in this setting because serverless workloads are highly dynamic and the underlying infrastructure is shared by multiple tenants. A user may face various difficulties like selection of proper constraints, monitoring etc. Most of the existing literature review focus on heuristic function scheduling methods that fall short of capturing the real dynamic in these systems brought on by multi-tenancy and shifting user request patterns. On the other hand energy consumption is another critical domain that leads to a lot of carbon emission affecting our environment. To tackle these issues, the paper proposes a Deep Reinforcement Learning (DRL) based Energy Efficient Application Scheduler (DEESched) in serverless environments. The experimental results show that DRL-based scheduler outperforms the other baseline schedulers by attaining a consistent accuracy of 70-90% and minimum utilization rate.

Keywords:

Serverless computing

FaaS

Function instance

Function scheduling

Energy efficiency