Abstract

Modern data centers face increasing pressure to improve energy 

efficiency while guaranteeing Service Level Objectives (SLOs) for 

Latency-Critical (LC) applications. Resource management in public 

cloud environments, typically operating at the node or instance 

level, often results in underutilized idle cores and unnecessary en- 

ergy consumption, highlighting the need for fine-grained core-level 

management. Existing studies on core management primarily focus 

on allocating cores for application threads, aiming to maximize 

idle cores without violating SLOs, but often neglect the impact of 

network packet processing on core idleness and energy consump- 

tion. In this work, we demonstrate that separate core allocation 

for network packet processing is essential to optimize both perfor- 

mance and energy efficiency in LC applications. Additionally, we 

show that co-managing packet processing intervals alongside core 

allocation further enhances core idleness and reduces energy con- 

sumption without compromising SLO compliance. Based on these 

insights, we propose EcoCore, a dynamic core management tech- 

nique that jointly manages core allocation for application threads 

and network packet processing while adaptively adjusting packet 

processing intervals. EcoCore employs a lightweight predictive 

model to estimate energy consumption and tail latency, enabling 

it to select energy-efficient configurations without violating SLOs. 

Through comprehensive evaluations, including deployment on the 

AWS cloud platform, EcoCore demonstrates its practicality, reduc- 

ing energy consumption by 11.7% on average and by up to 20.3% 

without SLO violations. In the AWS cloud environment, EcoCore 

extends core sleep states by up to 45.9%, resulting in additional 

energy savings of up to 35.8%. These results highlight EcoCore’s 

potential for bridging the gap between coarse-grained resource 

management and fine-grained core management in data centers.