Evaluating VM Live Migration Overhead with Xen Hypervisor Petrônio Bezerra*§, Gustavo Martins*, César Rocha*§, Reinaldo Gomes* and Fellype Cavalcante* Federal University of Campina Grande (UFCG) – Paraíba – Brazil § Federal Institute of Education, Science and Technology of Paraiba (IFPB) – Paraíba – Brazil *

Overview • Introduction • Experiment Setup • Evaluation • Conclusions

Introduction • Cloud computing brings many benefits

found

by

cloud

• “virtual machine (VM) is as a fully protected and isolated copy of the underlying physical machine’s hardware”.

WWW/INTERNET

• Virtualization → easy computing providers

10/2016

• computational structure → customer’s needs

2

Introduction

WWW/INTERNET

10/2016

• Virtual Machine Monitor (VMM) or hypervisor

3

Introduction

• Web server and a DBMS, deployed on a virtual server during a VM live migration process • Series of measurements with Xen hypervisor • Preliminary statistical analysis (real and virtual)

WWW/INTERNET

• Motivation • Measure the overhead that occurs due to the VM live migration process perceived on the client’s side by a decrease in performance over services

10/2016

• Live Migration

4

Experiment Setup

WWW/INTERNET

10/2016

• Two scenarios: virtual hosts and physical hosts

5

Experiment Setup • ab – Apache HTTP Server Benchmarking tool • it was measured the mean number of request per seconds

WWW/INTERNET

• it was measured the number of operations per seconds using the Workload A option

10/2016

• Yahoo! Cloud Serving Benchmark

6

Evaluation

• Large distance → normality assumption is probably false • Lowest value (= 3) was found in data set from ab on VM with live migration. The others > 19

• Suggests that the data sets were not normally distributed

WWW/INTERNET

• First, the distances between the mean and median were analyzed

10/2016

• Seeking statistical normality

7

Evaluation • The Shapiro-Wilk test was chosen and in almost all results the null hypothesis was refuted

WWW/INTERNET

10/2016

• P-values less 0.05 with W < 1.0 were found

8

WWW/INTERNET

10/2016

Evaluation

9

Evaluation

WWW/INTERNET

10/2016

• Box plots – the samples are independent

10

Evaluation

WWW/INTERNET

10/2016

• Box plots – the samples are independent

11

WWW/INTERNET

• The overhead caused by live migration process is significant, noting that on both benchmarks, the number of operations or requests per second were significantly reduced; • The results showed that the ab benchmark had superior performance when running on a fully real environment, regardless of the scenario being with or without live migration • In summary, the overhead caused by the VM live migration process observed from the client's point of view is very impactful, since performance was degraded in all results of the benchmark execution on real environment when compared to virtual environment.

10/2016

Conclusions

12

WWW/INTERNET

• We plan to quantify the effects concerning factors that more increases the overhead, using a 2kr experimental design will be used

10/2016

Future Work

13

Contacts: Petrônio Carlos Bezerra [email protected] César Rocha [email protected]

Gustavo Martins [email protected] Reinaldo Gomes [email protected]

Fellype Cavalcante [email protected]