Energy-aware VM Allocation on An Opportunistic Cloud Infrastructure

Energy-aware VM Allocation on An Opportunistic Cloud Infrastructure
ExtremeGreen: Extreme Green & Energy Efficiency in Large Scale
Distributed Systems (CCGrid 2013) – Delft, The Netherlands
Harold Castro, Mario Villamizar
(@mariocloud)
Department of Systems and Computing
Engineering
Universidad de los Andes
Colombia
Cesar Díaz, Johnatan Pecero,
Pascal Bouvry
Computer Science and Communications
Research Unit
University of Luxembourg
Luxembourg

THE PROBLEM
ENERGY IN AN OPPORTUNISTIC CLOUD ENVIRONMENT
ENERGY-AWARE TECHNIQUES
ALGORITHM TO CALCULATE THE ECR
EXPERIMENTAL RESULTS
CONCLUSIONS AND FUTURE WORK

THE PROBLEM
More than 2000 CPU cores

HOW UNACLOUD WORKS

Ubuntu
with OGE
HOW UNACLOUD WORKS

Ubuntu
with OGE
VMs begin to process
jobs of the
bioinformatics cluster
HOW UNACLOUD WORKS

Debian
with PBS
HOW UNACLOUD WORKS

Debian
with PBS
Ubuntu
with OGE
Both clusters are being executed
on the same physical/shared
commodity infrastructure.
HOW UNACLOUD WORKS

THE PROBLEM
VM1 VM2 VM3 VM4
8 CPU Cores 8 CPU Cores 8 CPU Cores 8 CPU Cores
How to assign the Set of Virtual Machines (SVM) to the Set of Physical
Machines (SPM) trying to reduce the Energy Consumption?
Set of Virtual Machines (SVM)
Set of Physical Machines (SPM)
4. TURNED OFF2. BUSY1. IDLE
0% 10% 0% 0%
3. HIBERNATED

T = Execution Time of a Task
f(x) = Function that return the ECR of a PM given its CPU used
ECRmon = ECR consumed by the monitor of the PM
ECRuser = ECR consumed by the user using a PM
ECRhib = ECR consumed by the PM whet it is hibernated

f(x) = Function that return the ECR of a PM given its CPU used

The best desktop machines to deploy a VM are those being used
(state 2), followed by the machines that are on idle (state 1),
hibernation (state 3) and turned off (state 4) state.
UnaCloud

A. Minimize the ECR using a
minimum set of Physical
Machines preferably in busy
state; increasing the
Execution Time of the
metatask (SVM) and executing
several VMs on a PM.
B. Assign to each PM
(preferably in idle, hibernated
or turned off state) only one
VM, avoiding the penalty
caused by the execution of
multiple VMs on the same PM,
and increasing the ECR due to
more PMs are in execution.
REDUCE ECR RESULTS FASTER

We propose three resource allocation algorithms that consider the
following assumptions and goals:
 Reduce the ECR required to execute a metatask (SVMs).
 Several VMs can be executed on a PM.
 The assignation of VMs to PMs is made in batch mode.
 There are enough CPU cores to supports the CPU requirements
of the SVM.
 It is better to assign a VM to a PM with a user (busy state) or to a
PM (with available cores) that is already executing another VM
(busy state).
 Physical machines are homogeneous.

VM1 VM2 VM3 VM4
Three different strategies/techniques were defined and tested to reduce
the ECR consumed by UnaCloud
Set of Virtual Machines (SVM)
Set of Physical Machines (SPM)
4. TURNED OFF2. BUSY1. IDLE
0% 10% 0% 0%
3. HIBERNATED

1. CUSTOM ROUND ROBIN ALLOCATION
SVM in arbitrary order
• It does not consider the state of the PM.
• It does not use a minimum SPM.
• VMs with low requirements may be initially assigned to PMs with
large CPU capabilities (blocking VMs with larger requirements).
SPM in arbitrary order

2. SORTING VMS AND PMs TO MINIMIZE THE USE OF PMs
SVM descending ordered by cores required and execution time.
SPM ordered by (1) PMs running VMs, (2) PMs in busy state and (3)
PMs with more available CPU cores.

2. SORTING VMS AND PMs TO MINIMIZE THE USE OF PMs
 We identified that if VMs with
similar execution times are
executed on the same PMs
the ECR required to execute
the SVM is lower.
 Every PMs used to executed
the SVM will execute tasks at
peak capacity during a
similar period of time.
PM1 PM2
8 Cores 8 Cores
VM1
4 Cores
10 Hours
VM2
4 Cores
1 Hour
VM3
4 Cores
10.5 Hours
PM1 PM2
8 Cores 8 Cores
VM1
4 Cores
10 Hours
VM2
4 Cores
1 Hour
VM3
4 Cores
10.5 Hours
With Algorithm 2

 We identified that if VMs with
similar execution times are
executed on the same PMs
the ECR required to execute
the SVM is lower.
 Every PMs used to executed
the SVM will execute tasks at
peak capacity during a
similar period of time.
PM1 PM2
8 Cores 8 Cores
VM1
4 Cores
10 Hours
VM2
4 Cores
1 Hour
VM3
4 Cores
10.5 Hours
PM1 PM2
8 Cores 8 Cores
VM1
4 Cores
10 Hours
VM2
4 Cores
1 Hour
VM3
4 Cores
10.5 Hours
With Algorithm 3
3. EXECUTING VMS WITH SIMILAR EXECUTION TIME

3. EXECUTING VMS WITH SIMILAR EXECUTION TIME

ALGORITHM TO CALCULATE THE ECR
 This algorithm is used once the
scheduling process (using one of
the 3 previous algorithms) has
finished.
 The ECR consumed by every PM
is calculated based on the ECR of
each VM.
 The goal of this algorithm is to
calculate the total power
consumed by the VMs.

 We conduct experiments based on
simulations.
 We use workloads based on real
VMs production traces (collected
during one year).
 We consider that a given
percentage of PMs has a user
working on it.
 We vary the percentage of busy
machines from 10% up to 50%
with increments of 10%.

 On average Algorithm 2 and
Algorithm 3 save up to 36% of PMs
more than Algorithm 1 when only
10% of the PMs are busy by users.
 When the number of PMs with
users increases up to 50%,
Algorithm 2 and Algorithm 3
require 26% less PMs than RR.

 For the first scenario, the average
gain on ECR by Algorithm 2 is up
to 19% and up to 20% by
Algorithm 3.
 In scenario 2, Algorithm 2 and
Algorithm 3 save up to 29% and
30% more than Algorithm 1,
respectively.
 Algorithm 3 saves 2% more energy
in both scenarios than Algorithm
2.

 We developed and simulated
different energy-aware algorithms
to allow UnaCloud to operate in an
energy efficient way.
 We will implement the new
proposed algorithms on the
UnaCloud infrastructure.
 We will try more allocation
strategies that consider VMs
shared resources constraints that
can lead to performance
degradation on the quality of
service.
CONCLUSIONS AND FUTURE WORK

THANKS FOR YOUR ATTENTION!

CONTACT INFORMATION
mj.villamizar24@uniandes.edu.co mjvc007@hotmail.com
@mariocloud
http://linkedin.com/in/mariojosevillamizarcano
Mario José Villamizar Cano

Energy-aware VM Allocation on An Opportunistic Cloud Infrastructure

More Related Content

What's hot (17)

Viewers also liked (7)

Similar to Energy-aware VM Allocation on An Opportunistic Cloud Infrastructure (20)

More from Mario Jose Villamizar Cano (20)

Recently uploaded (20)

Energy-aware VM Allocation on An Opportunistic Cloud Infrastructure