ms thesis defense evaluating cpu utilization in...

Evaluating CPU utilization in a Cloud

Environment

6/9/2017

Presenter

Hafiz ur Rahman

MSCS, KFUPM

Thesis Committee Members

Dr. Farag Azzedin (Advisor)

Dr. Mahmood Khan Naizi

Dr. Salahdin Adam

ICS Department, KFUPM

MS THESIS DEFENSE

MS THESIS DEFENSE Hafiz ur Rahman2 of 37

Outline

Literature ReviewIntroduction Research Objectives

Research Methodology ConclusionResults and Discussion

Present Research Objectiv

es

Conduct Literature Review

Select Appropri

ate Approac

h to achieve

objectives

Conduct Performa

nce Evaluatio

n

Collect Data

Analyze and

Conclude


• Cloud computing allows users and businesses to access their data

and applications from any device and anywhere in the world

• Virtualization is the fundamental element of cloud computing by

which we can deliver resources or data as a service

• A datacenter is a facility composed of networked computers and

storage that organizations use to organize, process, store, and

disseminate large amounts of data

Introduction


Underutilized servers (unproductive)

Maintenance cost

High infrastructure cost

High power consumption

Disaster recovery

Migration challenges

Running legacy applications

Difficulty in management of computing resources

Introduction Why Virtualization?


Virtualization offers many advantages over traditional data centers

✓ Reducing fiscal costs

✓ Running legacy applications

✓ Easing computing resource management

✓ Ease of deployment

✓ Easing system migration

✓ Easing backups and disaster recovery

✓ Efficient resource utilization

Many organizations are adopting virtualization technology to reduce the cost

while maximizing the productivity, flexibility, responsiveness, and efficiency

Introduction Why Virtualization?


Virtualization is a technology to run multiple operating systems on asingle machine and giving an illusion that each OS is running on realhardware

An example of running Windows, Linux, and Mac OSs on the same machine

Hardware

Virtual Machine Monitor

Windows

App App

App

Linux

App App

App

Mac

App App

Introduction What is Virtualization?

Hardware

Windows

App App

App

Regular System Virtualized Environment


• In a regular system, the hardware resources are shared by single operating system

• In virtualization environments, hypervisors are responsible to manage hardware resources and virtual machines

• Virtualization technology inserts an additional abstraction layer between the hardware and the operating system

Introduction Difference between regular & virtualized environment

Hardware

Windows

App App

App

Hardware

Virtual Machine

Monitor

Windows

App App

App


➢ System resources such as CPU, main

memory, and I/O are virtualized to

virtual machine

➢ Every guest operating system is in

charge of virtual resources and

concurrently share and access the

hardware resources, which mainly

decreases the performance of the

system


Virtualization ShortcomingsIntroduction

Such shortcomings may lead to system performance degradation

vCPU1vCPU1

vCPU2

vCPU1

vCPU2

pCPU pCPU pCPU pCPU

Virtual Machine Monitor…

…

VMnVM1 VM2

…

…VM2

vCPU3 vCPU3

vCPU2

vCPU1

vCPU4

Virtual CPUs – Physical

CPU MappingVirtual Machines

Allocation

Virtual CPUs –Virtual

Machines configuration

Virtualized Environment


Virtualization ShortcomingsIntroduction CPU Virtualization & Utilization

➢ CPU utilization is the amount of work handled by a CPU in

a given time

➢ In virtualized environments, the higher the percentage of

the CPU utilization results in the maximum performance

➢ CPU utilization is the metric that represents how busy is a

processor

• CPU is the most significant and critical resource among all the available resources in a system


In virtualized environments, the higher the percentage of the CPU utilization results in the maximum performance

CPU utilization in virtualized environmentsProblem Description

• CPU installed on the host is only one set, but each VM that runs on the host requires their own CPU. Therefore, CPU needs to

be virtualized

Under Allocation Balance Allocation Over Allocation Over Allocation


CPU Utilization

vCPU-VM allocation

vCPU-pCPU mapping

Resource sharing

between physical cores

Resource sharing

between logical cores

The choice of guest OS

The BIOS setting for

power management

Hypervisor

Main factors that have implications on CPU utilization [23]



• The aim is to investigate the best CPU resources allocation and better performance using state-of-the-

are hypervisors

• We focused, on hypervisor selection, vCPU-VM allocation, and vCPU-pCPU pinning strategies to find

out the variation in CPU utilizations

pCPU pCPU pCPU pCPU


…

VMnVM1 VM2

…

…VM2

vCPU1

vCPU2

vCPU1

vCPU2

vCPU1

vCPU2

vCPU1

vCPU4




vCPU1vCPU1

vCPU2

vCPU1

vCPU2

pCPU pCPU pCPU pCPU


…

VMnVM1 VM2

…

…VM2

vCPU3 vCPU3

vCPU2

vCPU1

vCPU4

We are motivated by the fact that virtualization suffers from drawbacks

Virtual CPUs – Physical CPU Mapping

Virtual Machines Allocation

Virtual CPUs –Virtual

Machines configuration

Motivation #01



• There are a variety of vendors for the virtualization environment

• When applying virtualization technology to an infrastructure environment, which hypervisor among others is

better and faster in terms of CPU utilization?

pCPU pCPU pCPU pCPU


…

• All virtualization vendors claim that their virtualization hypervisor is the best, however they depend on the

used application

Motivation #02



• Researchers evaluate and analyze hypervisors without investigating the best vCPU-VM configuration and

vCPU – pCPU mapping through which better CPU utilization and performance for each hypervisor can be

expected

• They assigned vCPU toVM based on non-suitable configuration

vCPU1vCPU1

vCPU2

vCPU1

vCPU2

pCPU pCPU pCPU pCPU


…

VMnVM1 VM2

…

…VM2

vCPU3 vCPU3

vCPU2

vCPU1

vCPU4Prior to deploying VM in a cloud

environment, which vCPU –VM

configuration is the best to

improve CPU utilization?

Motivation #03

After VM deployment, which vCPU-

pCPU pinning strategies is the best

to improve CPU utilization?



Research Objectives


Work Hypervisor Used ArchitectureVirtualization

Type

Benchmarking

Tool

Hypervisor

Comparison

vCPU-VM

Configuration vCPU-pCPU Mapping

Hwang et al. [1]

VMware EXSi,

Hyper-V2008,

KVM 2.6, Xen

CMP HAVRamspeed,

Netperf√ Randomly Randomly

S. Varette et al, [14]

Hyper-V, KVM,

VirtualBox,

VMware

CMPFV,

HAVHPL √ Randomly Randomly

Mancas [33] Vmware, KVM Not Mention Not Mention Passmask √ Randomly Randomly

Literature Review

√ : Authors compared different hypervisors

Babu et al [35]OpenVZ,

XenServer, XenSMP

Container,

PV, FVUnixbench √ Randomly Randomly

Charles David [31]KVM (RHEL 5),

Xen 3.1.2SMP

HAV,

PVPTS √ Randomly Randomly

Zong et al. [16] Xen 3.4.0 SMP PVApache

TPC-H* Randomly Not Randomly

Kourai et al, [10] Xen 4.4.03 SMP PV Tascell * Not Randomly Randomly

Sogand et al. [11] VMware EXSi 5.5 NUMA Unknown

Vcenter

Performace,

SSH + Sar

* Not Randomly Randomly


Present Research

Objectives

Conduct Literature

Review

Select Appropriate Approach to

achieve objectives

Conduct Performance Evaluation

Collect DataAnalyze and Conclude

Proposed Methodology


Proposed Methodology Architecture


➢ N-Queens problem is a classical

combinatorial problem, that have

different workloads

➢ The problem involves placing N queens

on an N x N chessboard such that no

queen can attack any other

➢ A solution requires that no two queens

share the same row, column, or diagonal

➢ As the problem size increases, the

corresponding possible solutions and

the elapsed time to solve the problem

is also increasing

➢ We tested each hypervisor for different

queens size ranges from 4 to 19

➢ For a regular-sized board (8 x 8), there

are 92 distinct solutions,

Experimental Design


Experimental Design


For the clarity of presentation, the Results and Discussion is divided into five sections:

Results and Discussion

(1) The Effect of Virtualization Technology on Performance

Regular System

Hypervisor#01

Hypervisor#02

(2) The Effect of Virtual Machines on

Performance

Under Allocation

Balance Allocation

Over Allocation

(3) The Effect of Virtual CPUs on

Performance

Under Allocation

Balance Allocation

Over Allocation

(4) The Significance of Over Allocation on Performance

Uniform Configuration

Un-uniform Configuration

(5) The Effect of Pinning Strategies on

Performance

Pinning Strategy

No Pinning Strategy



1. The Effect of Virtualization Technology

Hardware

Ubuntu

Low High

M

Hardware

Citrix XenServer

Ubuntu

Low High

M


Hardware

KVM

Ubuntu

Low High

M




The effect of Virtualization Layer using N-Queens benchmark


N-Queens problem with elapsed time (second)



The effect of Virtualization Layer using John-the-Ripper benchmark


E: The maximum error with one degree of confidence,

(alpha): using two tail distribution,

s: standard deviation,

n: number of samples



2. The Effect of Virtual Machines on Performance

The Effect of Virtual Machines on Performance

Under Allocation

Over Allocation

Balance Allocation



VM#1

vCPU1

vCPU3

pCPU1 pCPU2 pCPU3 pCPU4

VM#1

vCPU5

vCPU7

pCPU5 pCPU6 pCPU7 pCPU8 pCPU9 pCPU10 pCPU11 pCPU12

pCPU13 pCPU14 pCPU15 pCPU16 pCPU17 pCPU18 pCPU19 pCPU20 pCPU21 pCPU22 pCPU23 pCPU24

vCPU2

vCPU4

vCPU6

vCPU8

vCPU9

vCPU11

vCPU13

vCPU15

vCPU10

vCPU12

vCPU14

vCPU16

VM#2

vCPU1

vCPU2

vCPU3

vCPU4

vCPU5

vCPU6

vCPU7

vCPU8

vCPU1

vCPU2

vCPU3

vCPU4

vCPU5

vCPU6

vCPU7

vCPU8

VM#1 VM#2

VM#3 VM#4

vCPU1

vCPU2

vCPU3

vCPU4

vCPU1

vCPU2

vCPU3

vCPU4

VM#1 VM#2

VM#5 VM#6

VM#3 VM#4

VM#7 VM#8

vCPU1

vCPU2

vCPU1

vCPU2

vCPU1

vCPU2

vCPU1

vCPU2


2. The Effect of Virtual Machines on Performance <<< Under Allocation >>>




The effect of Virtual Machines on Performance - Under Allocation

2. The Effect of Virtual Machines on Performance <<< Under Allocation >>>



The effect of Virtual Machines on Performance - Balance Allocation

2. The Effect of Virtual Machines on Performance << Balance Allocation >>



2. The Effect of Virtual Machines on Performance << Balance Allocation >>

Linear Regression model for Balance Allocation



The effect of Virtual Machines on Performance - Over Allocation

2. The Effect of Virtual Machines on Performance < Over Allocation >



The effect of Virtual Machines on Performance: Interval Plot of Elapsed time (sec) for Citrix XenServer and KVM Hypervisors (Problem Size: 19)

2. The Effect of Virtual Machines on Performance < Comparison >



The effect of Virtual Machines on Performance (Under allocation, Balance Allocation, and Over allocation comparison)

2. The Effect of Virtual Machines on Performance < Comparison >



The effect of Virtual CPUs on Performance - KVM Total CPU Utilization Hypervisor Level - KVM

The effect of Virtual CPUs on Performance - Citrix XenServer Total CPU Utilization Hypervisor Level - Citrix XenServer

3. The Effect of Virtual CPUs on Performance (Comparison)



pCPU19 pCPU20 pCPU21 pCPU22 pCPU23 pCPU24





VM#1

vCPU1

vCPU3

vCPU5

vCPU7

vCPU2

vCPU4

vCPU6

vCPU8

vCPU9

vCPU11

vCPU10

vCPU12


VM#2

vCPU13

vCPU15

vCPU17

vCPU19

vCPU14

vCPU16

vCPU18

vCPU20

vCPU21

vCPU23

vCPU22

vCPU24






VM#1

vCPU1

vCPU3

vCPU5

vCPU7

vCPU2

vCPU4

vCPU6

vCPU8

vCPU9

vCPU11

vCPU10

vCPU12


4. The Significance of Over Allocation and Pinning

Strategies








VM#1

vCPU1

vCPU3

vCPU5

vCPU7

vCPU2

vCPU4

vCPU6

vCPU8

vCPU9

vCPU11

vCPU10

vCPU12


VM#2

vCPU13

vCPU15

vCPU17

vCPU19

vCPU14

vCPU16

vCPU18

vCPU20

vCPU21

vCPU23

vCPU22

vCPU24






VM#1

vCPU1

vCPU3

vCPU5

vCPU7

vCPU2

vCPU4

vCPU6

vCPU8

vCPU9

vCPU11

vCPU10

vCPU12


vCPU47 vCPU47 vCPU47 vCPU48 vCPU47 vCPU48

: : : : : : : : : : : : : : : : : :


Strategies




Strategies

The effect of Virtual CPUs on Performance - Uniform vCPUs per VMs

The effect of Virtual CPUs on Performance - Non-uniform vCPUs per VMs

CPU Utilization for uniform vCPUs

CPU Utilization for non-uniform vCPUs



The Effect of Hypervisors, Virtual Machines and Virtual CPUs on Performance

(Concluding Remarks)

Mean Elapsed time of Hypervisors, VMs, and vCPUs



Normal distribution of Elapsed Time (sec), uniformed vCPU allocation

The Effect of Pinning Strategies on Performance

(Concluding Remarks)

F i r s t S c e n a r i o

Normal distribution of Elapsed Time (sec), un-uniformed vCPU allocation

S e c o n d S c e n a r i o


Using CPU-bound operations, the results obtained from this evaluation showed that

commercial (Citrix Xenserver) and opensource (KVM) hypervisors showed similar performance

in terms of elapsed time and CPU utilization

The performance of a system would degrade by running many VMs, improper allocation of

vCPUs to VMs, or using unsuitable vCPUs - pCPUs pinning strategies

The elapsed time increases, when there is a massive over allocation of vCPUs

The best performance (elapsed time) was gained when there was only few active VMs with

balance allocation and over allocation (over allocation using no pinning strategies)

Pinning and no pinning strategies have no effect on under allocation and balance allocation of

vCPUs-VMs or overallocation having same workload is running on each VM

Conclusion and Future Work


CPU utilization during live migration of VMs from one host to another is one of the

future directions for this work

The obtained results from our evaluation experiments can be validated by

comparing using other commercial hypervisors (VMware and Hyper V). This will aid

the cloud service providers in choosing the best cloud platform

CPU utilization for I/O bound operations has not been investigated

We suggested that the cloud service providers and researchers should consider the

effects of massive over allocation of vCPUs, VMs, and vCPUs-pCPUs mapping when

they choose deployment strategies for vCPUs-VMs configuration and vCPU-pCPU

mapping

Conclusion and Future Work


I would like to acknowledge the support provided by

King Fahd University of Petroleum and Minerals (KFUPM)

Information & Computer Science Department

Acknowledgements

My Advisor: Dr. Farag Azzedin

Committee Members: Dr. Mahmood Khan & Dr. Salahdin Adam

All friends in KFUPM

Thank You !



The effect of Virtual CPUs on Performance - Citrix XenServer

3. The Effect of Virtual CPUs on Performance (Citrix XenServer)



Total CPU Utilization Hypervisor Level - Citrix XenServer

3. The Effect of Virtual CPUs on Performance (Citrix XenServer)



The effect of Virtual CPUs on Performance - KVM

3. The Effect of Virtual CPUs on Performance (KVM)



Total CPU Utilization Hypervisor Level - KVM

3. The Effect of Virtual CPUs on Performance (KVM)



4. (a) The Significance of Over Allocation << Uniform vCPUs per VMs >>

The effect of Virtual CPUs on Performance - Uniform vCPUs per VMs CPU Utilization for uniform vCPUs

• No Pinning strategy: The hypervisor is free to schedule domain’s vCPUs on any pCPUs

• Pinning strategy: The hypervisor is free to schedule the Dom0 vCPUs on any pCPUs and other active VMs’ vCPUs are statically pinned to user define logical CPUs

• Both strategies have no effect on under allocation and balance allocation of vCPUs-VMs

F i r s t S c e n a r i o



The effect of Virtual CPUs on Performance - Non-uniform vCPUs per VMs CPU Utilization for non-uniform vCPUs

4. (b) The Significance of Over Allocation << Un-uniform vCPUs per VMs >>

S e c o n d S c e n a r i o



Elapsed time to solve N-Queen Problem using Citrix XenServer



Elapsed time to solve N-Queen Problem using KVM


Types of Hypervisor


CMPs use relatively simple single-thread processor cores to

exploit only moderate amounts of parallelism within any one

thread, while executing multiple threads in parallel across

multiple processor cores.

SMP - Symmetric multiprocessing - the case where two or more

processors (identical or non-identical) are connected via an

interconnect and have equal access to all other system resources

like memory, IO ports, etc. ...


NUMA

NUMA - Non-uniform memory access - the case where in an SMP system the processors are not equi-distant from the shared system memory. i.e., a processor may be physically at varying distances from a given set of memory regions.

http://en.wikipedia.org/wiki/Non-uniform_memory_access


The cloud service provider assigns resources to each user

request aiming to minimize resource allocation and fulfill

user requirements

ms thesis defense evaluating cpu utilization in...

Documents