Source-linked AI summary
Maximizing Cloud Providers Revenues via Energy Aware Allocation Policies
Michele Mazzucco, Dmytro Dyachuk, Ralph Deters
TL;DR
Cloud providers must reduce electricity costs without sacrificing user experience and revenue. The paper develops energy-aware allocation policies that dynamically power servers on and off using demand estimates and system models. Experiments show that the Adaptive heuristic nearly matches the Optimal algorithm across traffic conditions, while dynamic policies reduce power consumption and keep lost jobs low.
Problem
Cloud providers need to balance electricity costs with availability, performance, and provider revenue as data-center energy consumption grows.
Method
The paper develops energy-aware allocation policies that estimate demand and system behavior to dynamically power excess servers off.
Results
The Adaptive heuristic produces revenues almost as high as the more computationally expensive Optimal algorithm across varying loads, while dynamic policies reduce power consumption and keep lost jobs low.
Takeaways & Limitations
The Adaptive heuristic is identified as a suitable candidate for practical implementation under different traffic conditions.
Abstract
from arXiv · showhide
Cloud providers, like Amazon, offer their data centers' computational and storage capacities for lease to paying customers. High electricity consumption, associated with running a data center, not only reflects on its carbon footprint, but also increases the costs of running the data center itself. This paper addresses the problem of maximizing the revenues of Cloud providers by trimming down their electricity costs. As a solution allocation policies which are based on the dynamic powering servers on and off are introduced and evaluated. The policies aim at satisfying the conflicting goals of maximizing the users' experience while minimizing the amount of consumed electricity. The results of numerical experiments and simulations are described, showing that the proposed scheme performs well under different traffic conditions.
I. INTRODUCTION
Data centers consume substantial energy, creating ecological and operating-cost pressures. The paper proposes dynamic energy-aware server allocation to balance provider revenue, user requirements, and electricity use.
- Motivation: Data centers’ high energy consumption increases both their ecological footprint and operating costs.The paper notes substantial electricity and water use and identifies electricity costs as a major operating expense.
- Problem: Cloud providers must avoid over-provisioning while maintaining customers’ availability and performance requirements.Choosing how many servers to run is difficult because excess capacity wastes power while insufficient capacity harms service.
- Approach: Dynamic energy-aware allocation policies use demand estimates and system-behavior models to maximize average provider revenue per unit time.The policies improve server-farm utilization by powering excess servers off.
- Evaluation: The paper evaluates the proposed policies under different traffic conditions and reports that they perform well.The paper’s organization places policy analysis in Section IV and comparative experiments under varying traffic in Section VI.
II. RELATED WORK
Prior data-center energy research includes server-level, server-pool, and hybrid approaches. The paper’s related work also distinguishes reactive and proactive pool resizing and electricity-price-aware traffic routing.
- Approach categories: Data-center energy-efficiency approaches are categorized as intensive, extensive, or hybrid.Intensive methods optimize individual-server power, extensive methods switch servers on or off, and hybrid methods combine both.
- Intensive approaches: Intensive approaches reduce power at fixed server counts through mechanisms such as CPU voltage/frequency scaling.Servers may still consume substantial power while idle, motivating additional approaches.
- Extensive approaches: Extensive approaches control the number of running servers at runtime, using reactive or proactive pool resizing.Reactive methods respond to load changes, whereas proactive algorithms estimate how many servers are needed.
- Hybrid approaches: Hybrid approaches trade off server count against per-server CPU voltage/frequency because both too many and too few active servers can increase energy use.Running too many servers wastes idle power, while too few can require higher CPU frequencies.
- Alternative approach: A distinct CDN approach dynamically reroutes traffic toward locations with lower electricity prices.The approach exploits geographic and temporal variation in electricity costs rather than changing server capacity alone.
III. THE MODEL
The model represents a cloud provider’s finite server pool, leased jobs, revenue, and electricity costs. A dynamic allocation policy changes the active-server count in response to estimated demand while accounting for reconfiguration assumptions.
- System model: The provider operates a cluster of S identical servers, with n active and S − n switched off, and each server handles at most one job.A completed server becomes idle and continues consuming energy when no new job arrives.
- Contracts and costs: Customers pay charges proportional to job size, while providers pay per-server running costs and electricity charges; arrivals finding all servers busy are lost.The charge amount is outside the paper’s scope and may include storage or bandwidth charges.
- Allocation control: The provider’s resource-allocation policy chooses n, ranging from switching all S servers off to switching them all on.The objective is to optimize provider profit through the active-server count.
- Dynamic control: Static server counts underperform because low traffic leaves servers under-utilized while peak traffic can overload the active pool.The model therefore requires changing the number of running servers as demand changes.
- Policy inputs: Observation windows estimate arrival rate, service time, and their squared coefficients of variation before each policy invocation.The running-server count remains constant between consecutive invocations.
- Assumptions: The model treats server state changes as instantaneous relative to observation windows, but reconfiguration energy must be modeled when transitions approach window length.Jobs already in service may finish before a server is switched off.
- Objective: Average revenue per unit time depends on customer charge, throughput, and average power consumed by running servers.Switched-off servers consume no power in the model, and the most challenging case occurs when charges and costs are close.
IV. POLICIES
The paper models server allocation with loss-system queueing models and develops policies that choose how many servers to run. The resulting revenue is evaluated efficiently, with numerical evidence that it is unimodal in the number of active servers.
- Queueing model: The system is modeled as an M/GI/n/n loss queue, where jobs arriving when all n servers are busy are blocked and lost.For Poisson arrivals, the model reduces to an Erlang-B system; more general service-time distributions are incorporated through peakedness approximations.
- Queueing model: In the Erlang loss model, the number of jobs forms a finite-state birth-and-death process with stationary probabilities p_j for j=0,...,n.Arrivals occur at rate λ when fewer than n jobs are present, while departures from state j occur at rate jµ.
- Approximation: For non-Poisson arrivals, the exact G/GI/n/n solution is unavailable, so the blocking probability is approximated using asymptotic peakedness and service-time information.The paper distinguishes cases based on the squared coefficients of variation and uses a normal approximation with numerical integration when needed.
- Revenue optimization: The average revenue can be computed efficiently, and numerical experiments indicate that revenue R is unimodal in the number of running servers n.The proposed search evaluates consecutive n values until revenue decreases or its increase falls below ε, treating the resulting local maximum as global or near-global.
- Policy design: The resulting Optimal policy evaluates the revenue expressions directly, while simpler heuristics are motivated to make allocation decisions faster and require less information.The Optimal policy requires evaluating Equations (5) and (8).
A. Adaptive Heuristic
The Adaptive heuristic allocates servers from recent demand estimates while balancing utilization against service quality. It uses a QED-style safety adjustment to account for stochastic variability.
- A. Adaptive Heuristic: The QED framing balances server-farm utilization with availability, unlike heavy-load behavior where delays or losses are common.The paper focuses on heavily loaded farms with ρ ∼ n, where service quality is balanced against server efficiency.
- A. Adaptive Heuristic: The Adaptive heuristic estimates arrival rate λ and mean service time 1/µ from one observation window, then allocates servers for the next window.The policy uses statistics collected during the current window to determine the next allocation.
- A. Adaptive Heuristic: The allocation includes a β√ρ adjustment for stochastic variability, with −1 ≤ β ≤ 1.The parameter controls the safety margin around the estimated load.
B. Predictive Heuristic
The Predictive heuristic addresses the limitation of assuming that the next window will repeat the current load. It forecasts future arrival rates from historical data using double exponential smoothing.
- B. Predictive Heuristic: The predictive approach replaces the assumption ρ_k = ρ_k+1 with a forecast of the next window’s load using historical data.This is motivated by recurring daily or weekly traffic patterns that may not be captured by simple adaptation.
- B. Predictive Heuristic: A double exponential smoothing heuristic estimates the future arrival rate from a smoothed value and a trend.The smoothing parameters α and γ are selected using the least squared method.
- B. Predictive Heuristic: The forecast for the arrival rate at time k + 1 is λF_k+1 = S_k + b_k.S_k is the smoothed value and b_k is the trend used in the forecast.
V. SERVER POWER USAGE ESTIMATION
The paper estimates data-center power from server activity using workload measurements and a simple idle/busy-server model. The estimate stays within 10% of a performance-counter-based model.
- The experiment used a Wordpress application with increasing Poisson-arrival workloads and minute-level power measurements.The workload replayed client sessions involving browsing, tagging, and blog searches.
- The measured server power rose linearly with CPU utilization, from 140 W idle to 220 W above 375% utilization.Power fluctuations reflected request-traffic and cooling-fan variability.
- The proposed estimate models average power as idle-server energy plus busy-server energy multiplied by the average number of active servers.The model uses e1 for idle-server energy, e2 for busy-server energy, and m̄ for the average number of servers running jobs.
- The simple estimate was within 10% of a model using performance counters.This supports using the simpler model despite more detailed alternatives that incorporate CPU, disk, and other metrics.
VI. PERFORMANCE EVALUATION
Experiments evaluate energy-aware server-allocation policies across load levels, traffic variability, and nonstationary demand. Dynamic policies improve revenue and utilization while reducing electricity consumption relative to static allocation.
- Server-count analysis: Each load condition has an optimal number of running servers; too many erode revenue through idle-server costs, while too few miss potential revenue.The numerical experiment varies offered load from 30% to 90%.
- Markovian traffic: Adaptive allocation produces revenues that grow with offered load and approach the computationally expensive Optimal policy, outperforming Static policies.Static policies yield negative revenues under light load or lose too many jobs as load increases.
- Markovian traffic: The Adaptive heuristic stays close to a busy-to-running server ratio of 1, while proposed policies maintain very low job-loss percentages.The ratio indicates how effectively running servers are utilized and is linked to user experience.
- Power consumption: Dynamic policies reduce average power consumption by running servers only when needed, thereby lowering electricity costs and improving provider profits.The evaluation directly compares average power consumption across policies.
- Traffic variability: Under highly variable non-Markovian traffic, achieved revenues are lower than under Markovian traffic, reflecting the greater difficulty of selecting the best server count.The variability setting uses ca2 = 2 and cs2 = 20.
- Nonstationary demand: In nonstationary-load experiments, QED variants have nearly identical cumulative revenues, while using fewer servers markedly reduces power consumption.The comparison includes adaptive, predictive, static, and Oracle policies.
VII. CONCLUSIONS
The paper introduces and evaluates implementable policies for dynamically adapting cloud capacity, finding that server-power decisions materially affect provider revenue. Experiments support the proposed policies under varied traffic, while identifying additional energy factors for future work.
- The proposed policies work well under different traffic conditions, and the Adaptive heuristic is a candidate for practical implementation.
- Server-power decisions can significantly affect provider revenue, and contractual obligations between clients and providers influence those decisions.
- Future research should include reconfiguration time and energy, CPU-frequency trade-offs, and networking-equipment power consumption.