Ming Mao, Marty Humphrey

CS Department, University of Virginia

A Performance Study on the VM Startup Time in the Cloud

Cloud 2012

(June 25, Honolulu, Hawaii)

1

Introduction

One great advantage of the cloud is dynamic scalability, BUT… Only meaningful when the compute resource is available in time

The fact is that acquired instances are NOT ready immediately

Why study the cloud VM startup time Important to time sensitive applications

Important to resource auto-scaling mechanisms

Inconsistent definitions across cloud providers

To set up a reference for the community

To know the difference from two years ago

2

Experiment setup

AWS, Windows Azure, Rackspace

Oct 15th, 2011 – Feb 15th 2012

A client periodically collects the instance startup information

By different factors – time, image size, machine type, location, etc.

3

Type OS Image Location

Amazon EC2 m1.small Linux(Fedora) ami-48aa4921 us-east-1a m1.small Windows (Win Server 2008)

ami-fbf93092 us-east-1a

Windows Azure Small WebRole

default WebRole app in Azure SDK South Central US

Small WorkerRole default WorkerRole app in Azure SDK

South Central US

Small VMRole Win Server 2008R2

South Central US

Rackspace Type IV Linux (Fedora) flavor 71 N/A Type IV Windows (Win Server 2008R2) flavor 28 N/A

VM Definition of VM startup time Linux VMs The first successful ssh login

Windows VMs The first successful remote desktop connection WebRoles The first successful http request

WorkerRoles The first successful logging

Results (1) – by time

4

Cloud OS Average VM startup time EC2 Linux 96.9 seconds EC2 Windows 810.2 seconds Azure WebRole 374.8 seconds Azure WorkerRole 406.2 seconds Azure VMRole 356.6 seconds Rackspace Linux 44.2 seconds Rackspace Windows 429.2 seconds

Results (2) – by image size

5

Cloud Average Data Transfer Rate between VM and Image Store

EC2 10.9 MB/s Azure 1.1MB/s Rackspace 22.5 MB/s

Results (3) – by machine type

6

Azure, Rackspace – the VM startup time goes longer as the instance type goes larger

EC2 – no significant differences among different instance types, except the micro ones

Results (4) – by location

No significant differences among locations

EC2 - The newly established data center has a slightly slower startup performance and greater variance

7

Results (5) – multiple instances

8

EC2 – VM startup time is relatively constant.

Azure – the last VM instance sometimes took significantly longer than the first one.

RS – does not support instance acquisition in batch

Results (6 - 1) – spot instances, by time

Spot instances have a longer waiting time but similar VM startup performance compared to on-demand instances

9

Results (6 - 2) – spot instances, by vm type

Similar to on-demand instances, the VM startup time shows no significant differences across different machine types

10

Results (6 - 3) – spot instances, by location

The US east region shows longer spot instance acquisition time

11

Results (6 - 4) – spot instances, multiple instances

Similar to on-demand instances, spot instances show constant performance across all the 16 machines acquired at the same time

12

Results (6 - 5) – spot instances, by real-time price

No significant relations between the VM startup time and the

real-time price

13

Result (7) – other facts

14

Cloud Success Rate VM Release Time EC2 99.2% 3-8 seconds Azure 99.6% 8-21 seconds Rackspace 92.0% 3-8 seconds

Spot instances are not always cheaper than on-demand VMs

Acquisition requests are not always successfully served (unusable machines).

Summary

Summary VM startup time is independent of time of the day.

EC2, Rackspace - Windows instances take 9 times longer than Linux instances. Azure - all three Role instances show similar performances.

Both the size of the OS image and the instance type can largely affect the VM startup time.

spot instances - longer VM startup time and greater variance

RS had a higher instance acquisition failure rate (8%) > EC2 (0.8%) & Azure (0.4%).

Azure – 200-second improvement and smaller variances; EC2 – performance does not change.

15

Questions

Thank you!

16