Process-level Power Estimation in VM-based Systems

HAL Id: hal-01132495

https://hal.inria.fr/hal-01132495

Submitted on 20 Apr 2015

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Process-level Power Estimation in VM-based Systems

Maxime Colmant, Mascha Kurpicz, Pascal Felber, Loïc Huertas, Romain Rouvoy, Anita Sobe

To cite this version:

Maxime Colmant, Mascha Kurpicz, Pascal Felber, Loïc Huertas, Romain Rouvoy, et al.. Process-level Power Estimation in VM-based Systems. European Conference on Computer Systems (EuroSys) - Poster session, Apr 2015, Bordeaux, France. �hal-01132495�

Process-level Power Estimation in VM-based Systems

Maxime Colmant, Mascha Kurpicz, Pascal Felber, Loïc Huertas, Romain Rouvoy, Anita Sobe

Problem

• Massive power consumption by data centers

• Hard to identify the largest power consumers

• Cannot attach a power meter to a VM

• Current approaches not fine-grained enough Vision

• Process-level power estimation in VM-based systems

• Support for all CPU features

• Support for multi-tenant virtualization

Motivation

Hardware (HW) Performance Counters

• Representative and accurate metrics

• Mostly available on modern processors Criteria selection

• Counter availability per CPU

• Monitoring overhead

• Best fit under several workloads Selected HW Performance Counters

• CPU_CLK_UNHALTED:THREAD_P (uc)

• CPU_CLK_UNHALTED:REF_P

Metrics

1

2

3

Title, maybe is not too short like this one

Methods

Experiments and Results

Discussion

References and Acknowlegments

authors names (institution, mail)

Introduction

Machine Stress CPU

utility

PowerSpy Power meter

Regression Power model Power

consumption

CPU metrics Workloads

2 4

1 3

Unhalted cycles Reference cycles HW Counters

3

Architecture Learn the CPU power model

Host

BitWatts

File System

A P P

ZeroMQ

Formula

Clock Monitor Sensor Reporter

virtio libpfm libpfm

A P P

Guest

BitWatts

A P P Formula virtio

Sensor virtio

Monitor Reporter

Clock

File System A P P

VirtioSerial

Power Model

P_host(f) =P_idle(f) + X

pid2P IDs

P_cpu(f, uc¹_pid...uc^N_pid)

P_cpu(f, uc¹_pid...uc^N_pid) = XN n=1

P_f(ucⁿ_pid)

Host_C

Host_A VM1

BitWatts

App BitWatts

VM2 BitWatts

App

Monitoring console Publish/subscribe broker publish

publish

Host_B VM₁

BitWatts

App BitWatts

publish

subscribe

BitWatts

80 90 100 110 120 130 140

0 50 100 150 200 250 300 350 400

Power (W)

Time (ms) x264

freqmine BitWatts

Idle power PowerSpy

20 30 40 50 60 70

0 200 400 600 800 1000 1200 1400 1600 1800

Power consumption (W)

Time (s)

BitWatts PowerSpy

0 50 100 150 200 250 300

1 2 4 1 2 4 1 2 4 1 2 4 1 2 4 1 2 4 1 2 4

Median power (W)

BitWatts

-0.35% -4.91% 1.90%

PowerSpy

-0.80% -7.91% -5.15% 3.82% 1.40% -0.92% 3.63% -1.30% -9.92% 3.59% 0.65% -8.64% 3.02% 0.21% -10.22% -10.76% 7.27% 8.17%

vips swaptions

freqmine fluidanimate

facesim bodytrack

blackscholes

Validation

Implementation

Process-level power

consumption of BitWatts, x264, and freqmine on a Xeon W3520 processor

Power consumption during the execution of SPECjbb on a i3 2120 with two threads

Power consumption of the host when scaling PARSEC on multiple VMs

1

2

3

1

2

3

Title, maybe is not too short like this one

Methods

Experiments and Results

Discussion

References and Acknowlegments

authors names (institution, mail)

Introduction

Title, maybe is not too short like this one

Methods

Experiments and Results

Discussion

References and Acknowlegments

authors names (institution, mail)

Introduction

Research session 5: Virtualization, 23 April, 10:50 am.