PaperlayerView exact revision on arXiv

Source-linked AI summary

Power Consumption of Virtualization Technologies: an Empirical Investigation

Roberto Morabito

arXiv:1511.01232v1cs.DCcs.PF

TL;DR

Data-center energy use motivates measuring virtualization power, while prior related work generally omitted container platforms, especially Docker. This paper empirically compares four virtualization technologies across workloads and finds similar power consumption under heavy CPU and memory workloads, with differences in network tests and lower consumption for containers in most cases.

  • Problem

    Related work generally omitted container-based platforms, especially Docker, from power-related virtualization analysis.

    Source: §VI · p. 6

  • Method

    The paper measures power across Xen, KVM, Docker, and LXC using repeated testbed experiments and idle, CPU, memory, and network workloads.

    Source: §IV · p. 3, §III · p. 3, B. CPU performance · p. 4

  • Results

    Hypervisors and containers show similar power consumption under heavy CPU and memory workloads, while network tests differ and containers use less power in most cases.

    Source: §VII · p. 6

  • Takeaways & Limitations

    The observed container power advantage is limited to most network cases; CPU and memory workloads show similar consumption across the two virtualization types.

    Source: §VII · p. 6

Abstract

from arXiv · show

Virtualization is growing rapidly as a result of the increasing number of alternative solutions in this area, and of the wide range of application field. Until now, hypervisor-based virtualization has been the de facto solution to perform server virtualization. Recently, container-based virtualization - an alternative to hypervisors - has gained more attention because of lightweight characteristics, attracting cloud providers that have already made use of it to deliver their services. However, a gap in the existing research on containers exists in the area of power consumption. This paper presents the results of a performance comparison in terms of power consumption of four different virtualization technologies: KVM and Xen, which are based on hypervisor virtualization, Docker and LXC which are based on container virtualization. The aim of this empirical investigation, carried out by means of a testbed, is to understand how these technologies react to particular workloads. Our initial results show how, despite of the number of virtual entities running, both kinds of virtualization alternatives behave similarly in idle state and in CPU/Memory stress test. Contrarily, the results on network performance show differences between the two technologies.

I. INTRODUCTION

Growing data-center energy use makes virtualization power metering relevant to energy-efficiency efforts. The paper compares virtualization technologies, while the background notes that LXD could not be configured for a fair comparison.

I. INTRODUCTION

  • Container adoption has increased, with cloud providers already using container solutions to offer services.Docker is cited as a technology that contributed to this growth.Source: §I · p. 1
  • The paper frames virtualization power metering as relevant to reducing power consumption amid data-center growth.The passage reports an estimate of approximately 91 million MWh of U.S. data-center usage in 2013.Source: §I · p. 1
  • The study examines whether emerging solutions such as containers can reduce power consumption compared with hypervisors.Source: §I · p. 1

II. BACKGROUND

  • LXD was excluded because bugs in its API prevented configuration for a fair comparison with the other technologies.Source: §II · p. 1

A. Hypervisor-based virtualization

The paper contrasts hardware-level hypervisors with operating-system-level containers and compares KVM, Xen, Docker, and LXC. Xen additionally uses a privileged Domain 0 to manage other domains.

A. Hypervisor-based virtualization

B. Container-based virtualization

  • Containers isolate processes without emulating virtual hardware, allowing higher instance density and smaller disk images; the study compares Docker with LXC.Docker replaced LXC as its default execution environment starting with version 0.9, prompting a performance comparison.Source: B. Container-based virtualization · p. 2

III. EXPERIMENTAL SETUP AND METHODOLOGY

The methodology compares four virtualization technologies on a two-server testbed using CPU, memory, and network workloads. It standardizes virtualized configurations and measures active power with an external PDU.

III. EXPERIMENTAL SETUP AND METHODOLOGY

  • The testbed uses two identical Dell Precision T5500 servers with Xeon X5560 processors, 12 GB memory, and 10 Gbps NICs.The hosts run Ubuntu 14.04 LTS and are directly connected without a switch.Source: §III · p. 2, Figure 2 · p. 2
  • An external PDU measures each server’s active power, defined as mean instantaneous power over a time period.Measurements use a 15-second period and a 3-second sample rate.Source: §III · p. 2
  • The virtualized setup uses up to eight guests, each allocated 2 vCPUs and 1 GB memory, with Xen, QEMU/KVM, Docker, and LXC versions specified.Source: §III · p. 2
  • CPU pinning reserves two cores for Xen’s Domain 0, while KVM and containers use corresponding configurations without Domain 0.The selected CPU-affinity arrangements are shown for Xen, KVM, and containers.Source: §III · p. 3, Figure 3 · p. 3
  • All virtual entities share a network bridge mapped to the host’s physical Ethernet, while workload tools target CPU, memory, and NIC behavior.Source: §III · p. 3, Figure 4 · p. 3
  • Sleep generates idle conditions, sysbench runs intensive multithreaded tests, and iperf generates bidirectional TCP or UDP network traffic.Source: §III · p. 3

IV. POWER MEASUREMENTS RESULTS

The results section reports measurements for idle state, CPU, memory, and network I/O. Reported values are averages across repeated tests, with standard deviation shown in some graphs.

IV. POWER MEASUREMENTS RESULTS

  • The results are organized around idle state, CPU, memory, and network I/O performance.Source: §IV · p. 3
  • Each measurement test was repeated at least 20 times, and reported results are averages.Source: §IV · p. 3
  • Some graphs also report the standard deviation of measurements.Source: §IV · p. 3

A. Idle State

With eight virtual guests running, all four technologies show roughly the same average idle power consumption, around 4 watts. Xen consumes the most, but its difference from containers is small.

A. Idle State

  • All four technologies consume roughly the same average power, in the range of 4 watts, in idle state.The comparison covers eight running guests.Source: A. Idle State · p. 3, Table I · p. 3
  • The idle-state measurement uses eight guests running sleep and compares virtualization against the SUT without virtualization.Source: A. Idle State · p. 3
  • Xen consumes the most power, although its difference from containers is not high.Source: A. Idle State · p. 3

B. CPU performance

Under sysbench CPU stress, no virtualization technology outperforms the others. Power rises with additional active guests, while the reported averages for hypervisors and containers remain close at four and eight guests.

B. CPU performance

  • At one active virtual guest, all platforms consume around 158 Watts; activating another guest on another physical core adds 10 Watts.Source: B. CPU performance · p. 4
  • The CPU stress test uses sysbench to calculate prime numbers, and none of the four technologies outperforms the others.Source: B. CPU performance · p. 4
  • With four and eight guests balanced across two physical cores, hypervisors average around 185 Watts and containers around 190 Watts.Figure 5 presents CPU active power for one, two, four, and eight active virtual entities.Source: Figure 5 · p. 4, B. CPU performance · p. 4

C. Memory Performance

In the memory stress test, hypervisors and containers behave similarly on average, with no clear platform difference. With guests balanced across two physical cores, power rises from 212 Watts at two guests to 222 Watts at eight.

C. Memory Performance

D. Network Performance

The network tests compare power use across receive and transmit workloads, standard and jumbo frames, and multiple concurrently active virtual entities. Results show power differences across technologies, with the pattern varying by workload and direction.

D. Network Performance

  • Network tests used at least ten simultaneous iperf sessions per VM/container because one session could not saturate a 10 Gbps link even at low CPU workload.Except for the final case study, tests ran on a single virtual entity.Source: D. Network Performance · p. 4
  • Jumbo-frame tests set MTU to 9000 bytes on all physical and virtual NICs to compare frame handling at equal throughput.The test rationale was that standard 1500-byte MTU could not saturate a 10 Gbps link; bidirectional testing accounted for TCP's different send and receive code paths.Source: D. Network Performance · p. 4
  • Containers used about 8 W less than KVM when receiving TCP at 1500-byte MTU; Xen could not saturate the link with standard configuration.With 9000-byte frames, Xen saturated the link and used power at the same level as KVM; container power showed no valuable MTU-related difference.Source: D. Network Performance · p. 4, Figure 7 · p. 5
  • For TCP transmission at an average 9.40 Gbps, KVM, Docker, and LXC used the least power, while Xen used 10 W more on average.For UDP transmission, Xen's power use exceeded the other technologies by 22–25 W.Source: D. Network Performance · p. 5, Figure 8 · p. 5
  • With eight virtual entities receiving TCP traffic, hypervisors used about 15 W more than containers; Docker used the least power, with LXC close behind.The reported average power was 199 W for Xen and 197 W for KVM.Source: D. Network Performance · p. 5
  • With virtual entities generating TCP traffic, KVM used less power than Xen, while Docker and LXC drew the same amount in this case.This is the multiple-entity client case shown in Figure 10.Source: D. Network Performance · p. 5

V. ANALYSIS

Idle power and CPU/memory stress results were similar across virtualization designs, while network tests showed differences. The analysis attributes network differences mainly to extra packet-processing layers in hypervisor environments.

V. ANALYSIS

  • CPU-heavy workloads showed no noticeable power-consumption difference across virtualization designs, regardless of the number of running instances.The idle-state results also indicate optimized use of Linux power saving by both hypervisors and containers, with no significant overhead from Xen's Domain-0.Source: §V · p. 5
  • Network power results differed across case studies, mainly because hypervisors process packets through extra layers compared with containers.KVM appeared to perform slightly better than Xen; Xen's Domain-0 first receives packets from the physical NIC and forwards them to Domain-N, which can add overhead.Source: §V · p. 5

VI. RELATED WORK

The related work distinguishes empirical measurement studies from analytical power-estimation models and reviews investigations most similar to this paper. It emphasizes that container platforms, especially Docker, were largely absent from prior analyses.

VI. RELATED WORK

  • Prior power-modeling research is divided into empirical studies based on direct measurement and analytical power-estimation models.This review focuses on investigations similar to the paper's empirical approach.Source: §VI · p. 5
  • Earlier empirical studies compared KVM and Xen, included OpenVZ and found adaptive KVM packet buffering could reduce network-transaction energy, and examined hypervisor-dependent energy overhead.Other cited work introduced runtime power models from resource usage and application-agnostic process-level estimation.Source: §VI · pp. 5–6
  • Except for Shea et al., related work did not include container platforms, particularly Docker; this study also examines a wider range of network parameters than..Source: §VI · p. 6

VII. CONCLUSION AND FUTURE WORK

The study empirically compares server power use for KVM, Xen, Docker, and LXC under different workloads. CPU and memory stress produced similar power use, while containers used less power in most network cases; future work would broaden technologies and test conditions.

VII. CONCLUSION AND FUTURE WORK

  • Across heavy CPU and memory workloads, hypervisors and containers showed similar power consumption; in most network cases, both container solutions used less power.The comparison used measurements across workloads and virtualization solutions.Source: §VII · p. 6
  • Future measurements could include LXD once stable enough for fair comparison, live migration, and alternative CPU-pinning and network configurations..Source: §VII · p. 6
Loading 1511.01232v1…