1/30/2015 just-in-time virtual machine provisioning for cloud offload kiryong ha ([email protected])...

12
1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha ([email protected]) Carnegie Mellon University

Upload: asher-hunt

Post on 18-Dec-2015

227 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

1/30/2015

Just-in-Time Virtual MachineProvisioning for Cloud Offload

Kiryong Ha ([email protected])

Carnegie Mellon University

Page 2: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

2

Motivation: Cloud Offload

Compute-intensive & interactive applications in mobile context

• Apple’s Siri, AR (Augmented Reality) applications

Cloud offload• These applications are partitioned into a client/a server

– Client program: front-end app installed on a mobile device– Server program : back-end program running on a datacenter (e.g., Amazon EC2)

• The client program offloads computation to the server program at Cloud

Today’s cloud is a suboptimal place; high latency and limited bandwidth

Page 3: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

3

Cloud Computing at the edge of the Internet

Cloudlet: Micro Datacenter at the edge of the Internet Bring the cloud closer to the user

WAN

LAN

OS Server progClient prog…

Cloud (Datacenter)

Eiffel tower

VM

Serverprog

Eiffel tower

Nokia Siemens Network& IBM

NVIDIA

How to provision a back-end server program (VM) at a nearby machine?

Augment Reality

Cloudlet (local server)

Page 4: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

4

Goal: Rapid Just-in-time VM Provisioning

Provisioning server program Just-in-time VM provisioning• Preserve the running state of the server by snapshotting• Provision ① disk image ② memory snapshot

WAN

Cloud

Cloudlet

Goal: Rapid Just-in-time VM provisioning for cloud offload

disk.imgmem.snap

VMOS

Server

VMOS

Server

Challenge: the speed of provisioning

disk.imgmem.snap

Page 5: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

5

Approach: VM Synthesis

VM Synthesis : dividing a custom VM into two pieces1) Base VM: Vanilla OS that contains kernel and basic libraries (e.g., Windows7)2) VM overlay: A binary patch that contains only customized parts

Pre-provision Base VM + just-in-time provision VM overlay

Base VM

Customized VM(Launch VM)

Binary delta

Overlay Memory

Overlay Disk

VM overlay

Compress

Base Disk

Base Memory

disk

..

mem

Page 6: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

6

VM Synthesis

Steps for VM synthesis

VM overlay Synthesize VM- Decompress- Apply delta

Resume the VM

User

Offload operations

ready

Cloudletpre-populated

base VM

VM synthesistime

VM overlay

Base Disk

Base Memory

VMOS

ServerEiffel tower

Page 7: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

7

OpenStack Implementation

• Implementation– Grizzly (at Ubuntu 12.04 LTS)– Icehouse (at Ubuntu 14.04 LTS)– Standalone version

• Main Features1. Create a custom VM overlay (nova-compute extension)2. Store Base VM and VM overlay (glance)3. Perform VM synthesis

Page 8: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

8

OpenStack Implementation

Page 9: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

9

OpenStack Implementation

Page 10: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

10

Overview of Optimizations

1. Creating VM overlay: Minimize VM overlay size

2. VM Synthesis: Accelerate VM Synthesis

WAN

Cloud

Cloudlet

VMOS

Server

VMOS

Server 2. Deduplication

1. Reducing Se-mantic Gaps

3. Pipelining

4. Early Start

VM overlay

Page 11: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

11

First Response Time of the Application

Time between starting VM synthesis and receiving the first offload result at client• First-response within 10 seconds except AR

[intro][background][optimization][result][conclusion]

OBJECT FACE SPEECH AR FLUID0

25

50

75

100 144

Baseline synthesis

Fully optimized synthesis

Tim

e (s

)

10

Page 12: 1/30/2015 Just-in-Time Virtual Machine Provisioning for Cloud Offload Kiryong Ha (krha@cmu.edu) Carnegie Mellon University

12

References

• Cloudlet project page– http://elijah.cs.cmu.edu/

• Paper1) Kiryong Ha, Padmanabhan Pillai, Wolfgang Richter, Yoshihisa Abe, Mahadev

Satyanarayanan, “Just-in-Time Provisioning for Cyber Foraging”, In MobiSys 2013http://krha.kr/data/pubs/vmsynthesis2013.pdf

2) Satyanarayanan, M., Bahl, P., Caceres, R., Davies, N. ‘‘The Case for VM-based Cloudlets in Mobile Computing’’ IEEE Pervasive Computing 2009.http://www.cs.cmu.edu/~satya/docdir/satya-ieeepvc-cloudlets-2009.pdf

• Source code1) https://github.com/cmusatyalab/elijah-provisioning2) https://github.com/cmusatyalab/elijah-openstack