[Paper Review] On the Energy Proportionality of Scale-Out Workloads

Balaji Subramaniam, Wu-chun Feng|arXiv (Cornell University)|Jan 12, 2015

Cloud Computing and Resource Management18 references3 citations

TL;DR

This paper investigates energy proportionality in scale-out workloads using data serving, web search, and data caching workloads on real clusters. It evaluates power provisioning (active/low-power modes) and resource provisioning, finding that active and idle low-power modes together reduce system power by up to 47% and processor power by up to 77%, while enabling power-performance trade-offs that meet strict latency SLOs.

ABSTRACT

Our increasing reliance on the cloud has led to the emergence of scale-out workloads. These scale-out workloads are latency-sensitive as they are user driven. In order to meet strict latency constraints, they require massive computing infrastructure, which consume significant amount of energy and contribute to operational costs. This cost is further aggravated by the lack of energy proportionality in servers. As Internet services become even more ubiquitous, scale-out workloads will need increasingly larger cluster installations. As such, we desire an investigation into the energy proportionality and the mechanisms to improve the power consumption of scale-out workloads. Therefore, in this paper, we study the energy proportionality and power consumption of clusters in the context of scale-out workloads. Towards this end, we evaluate the potential of power and resource provisioning to improve the energy proportionality for this class of workloads. Using data serving, web searching and data caching as our representative workloads, we first analyze the component-level power distribution on a cluster. Second, we characterize how these workloads utilize the cluster. Third, we analyze the potential of power provisioning techniques (i.e., active low-power, turbo and idle low-power modes) to improve the energy proportionality of scale-out workloads. We then describe the ability of active low-power modes to provide trade-offs in power and latency. Finally, we compare and contrast power provisioning and resource provisioning techniques. Our study reveals various insights which will help improve the energy proportionality and power consumption of scale-out workloads.

Motivation & Objective

• Investigate energy proportionality challenges in scale-out workloads, which are latency-sensitive and rely on large-scale clusters.
• Analyze component-level power distribution in clusters, identifying the CPU as the dominant power consumer (45–70% of system power).
• Characterize CPU utilization patterns at fine-grained time resolutions (sub-millisecond) to identify power-saving opportunities.
• Evaluate the effectiveness of power provisioning techniques (active low-power, turbo, idle low-power) in improving energy proportionality.
• Compare power provisioning and resource provisioning strategies, analyzing their trade-offs in power savings and performance impact.

Proposed method

• Conducted empirical power measurements on a four-node cluster using the RAPL interface to monitor component-level power consumption.
• Selected representative scale-out workloads: data serving (e.g., Netflix), web searching (e.g., Google), and data caching (e.g., Facebook).
• Evaluated power provisioning techniques: active low-power (C-states), idle low-power (C-states), and turbo modes, measuring their impact on latency and power.
• Used workload characterization to analyze CPU utilization at sub-millisecond resolution, revealing underutilization even at high load levels.
• Implemented and compared resource provisioning via server consolidation and dynamic offloading, assessing energy savings at varying load levels.
• Mapped the power-performance trade-off space using active low-power modes, enabling configurable power savings within SLO constraints.

Experimental results

Research questions

• RQ1 To what extent is the processor the dominant power consumer in scale-out workloads across varying load levels?
• RQ2 How does CPU utilization vary at fine-grained time resolutions, and what does this imply for power management opportunities?
• RQ3 What is the impact of active and idle low-power modes on energy proportionality and response time in latency-sensitive workloads?
• RQ4 How do power provisioning and resource provisioning compare in terms of energy savings and performance trade-offs at different load levels?
• RQ5 Can active low-power modes create a viable power-performance trade-off space that allows meeting strict SLOs while reducing energy consumption?

Key findings

• The processor accounts for 45–70% of total system power in scale-out workloads, making it the primary target for energy optimization.
• Even at sub-millisecond time resolutions, significant CPU utilization variability reveals substantial opportunities for power savings through dynamic management.
• Combining active and idle low-power modes achieves up to 47% reduction in system-level power and up to 77% reduction in processor-level power.
• Active low-power modes enable a tunable power-performance trade-off space, allowing workloads to operate in lower-power configurations while still meeting sub-second SLOs.
• Resource provisioning provides the best energy proportionality at low load levels, where idle power dominates; power provisioning becomes more effective at higher load levels.
• The study confirms that energy-proportional operation is achievable but not uniform across all load levels, with performance costs varying based on the chosen power management technique.