Hardware

Virtualization Layer

Front-endVM

Mid-endVM

Back-endVM

Multi-Tier AppDesktop

VM

Global siteRegional site

1

Virtualized Distributed Cloud Hosting Various ApplicationsCloud computing is a popular paradigm for hosting online applications.Virtualization techniques are widely used to host applications ranging from Multi-tier web applications to desktop VMs.

A Hypothetical Distributed CloudCloud Platforms have become more distributed in recent years. The hypothetical distributed cloud is expected to evolve as above with a few large global sites and numerous smaller regional sites.To utilize this infrastructure, we have to automatically place different applications to different locations based on their dynamic needs.

DesktopVMDesktopVMDesktopVMDesktopVMDesktopVM

DesktopVM

DesktopVMDesktopVMDesktopVMDesktopVMDesktopVM

DesktopVM

Cloud

Cloud

Cloud

We place each VM in an initial cloud location based on the immediate request. We treat each desktop VM as a black box and infer their location sensitivities based on their network activities.

Periodically, we make intelligent decisions to dynamically adjust the mapping between desktop VMs and cloud locations depending on application needs and environment demand.

Placing Desktop VMs in Distributed Cloud

3

4036.2 km, RTT = 96.15 ms

3633.9 km, RTT = 84.03 ms

558

.4 k

m, R

TT =

14.

37 m

s

Experiment setup: Thin clients connect from Massachusetts, and three desktop VMs running different applications reside separately in Amazon EC2 sites of Virginia and Oregon.

Desktop VM2

Desktop VM3Desktop VM1

VA

VirginiaOregon

Massachusetts

VNCVNC

VNC

5

Hosting virtualized desktop VMs in remote servers or clouds is gaining popularity for either enterprises or third-party provider.

Users are using remote desktop protocol, e.g., Windows RDP and VNC to access such VMs and use various applications.

However, not all desktop VMs are latency or location sensitive. The location-sensitive rank is determined by the applications running inside. And applications that are characterized as high interactivity or refresh rate contribute to higher rank.

Desktop VM

Remote Desktop Protocol

Virtual Desktop gains popularity

2

VMShadow:Optimizing The Performance of Virtual Desktops in Distributed Clouds

Tian Guo, Vijay Gopalakrishnan, K. K. Ramakrishnan, Prashant Shenoy, Arun Venkataramani, Seungjoon LeeUniversity of Massachusetts Amherst, AT&T Labs - Research

Cloud Manager Proxy

Desktop Apps

OS

WAN Live Migrator

NestedHypervisor

Fingerprinting Engine

Greedy ShadowAlgorithm

Greedy Shadow Algorithm:Uses cost-benefit metrics to judiciously move highlylocation-sensitive VMs at the least cost.

WAN Live Migrator:Uses WAN-specific optimizations to migrate VMs' disk and memory state.

Connection Migration Proxy:Transparently migrates active TCP connections to ensure seamless connectivity despite IP changes.

VMShadow Design Goal

Uses black-box VM fingerprintingtechnique to infer the location sensitivity of VMs.

Fingerprinting Engine:

4

VNC

Resp

onse

Tim

e in

Sec

VM1: Watching Internet Streaming

Oregon

Start Migration Finish Migration

Time (Sec)

VirginiaStart Migration Finish Migration

VM3: Using Local Word Editor

Start MigrationFinish Migration

Time (Sec) Time (Sec)

Oregon

VM2: Watching Local Video

Time (Sec)

In this case study, we demonstrate VMShadow’s ability to discriminate between location sensitive and insensitive desktop VMs and to trigger appropriate WAN migrations to improve VNC response time in an artificially constrained cloud environment.

Decisions are made to migrate VM1 and VM2 to VA, to be closer to user. When VA is resource constrained, low-ranked VM3’s resources are reclaimed by migrating it back to OR, after which VM1 and VM2 are migrated to VA.

Case Study

6