Slide 1 - UCLA

triangledriprockInternet and Web Development

Aug 7, 2012 (5 years and 3 months ago)

727 views


VDN: Virtual Machine Image Distribution
Network for Cloud Data Centers



IEEE INFOCOM 2012

Orlando, Florida USA

Chunyi Peng
1
,
Minkyong

Kim
2
,
Zhe

Zhang
2
,
Hui

Lei
2

1
University of California, Los Angeles

2
IBM T.J. Watson Research Center

Cloud Computing

Infocom 2012

C Peng (UCLA)

2

the delivery of
Computing
as a
Service

Service Access in Virtual Machine Instances

Infocom 2012

C Peng (UCLA)

3

Picture source: http://
www.wikimedia.org


Cloud Clients

Web browser, mobile app, thin
client, terminal emulator, …

Software as a Service (
SaaS
)

CRM, Email, virtual desktop,
communications, games, …

Application

Platform as a Service (
PaaS
)

Execution runtime, database, web
server, development tools, …

Platform

Infrastructure as a Service (
IaaS
)

Virtual machines, server storage,
load balancer, networks, …

Infra

structure

VM

VM

VM

VM

VM

Client Service Requests (e.g. HTTP)

Problem:

On
-
demand VM provisioning

Time for VM Image Provisioning











4

C Peng (UCLA)

Infocom 2012

VM image transfer

Req


process


User request

time

VM

Bootup

Our focus: Transfer time

Why Slow?


VM image files are large (several or tens of GB)


Centralized image storage becomes a bottleneck


Infocom 2012

C Peng (UCLA)

5

ToR

switch

Access

Data Center

Aggregation

Core

Image
-
server

RH5.6

Roadmap


Basic VDN idea: enable collaborative sharing


VDN solution on efficient sharing


Basic sharing units


Metadata management



Performance evaluation


Conclusion

Infocom 2012

C Peng (UCLA)

6

VDN: Speedup VM Image Distribution


Enable collaborative sharing


Utilize the “free” VM images


Exploit source diversity and make full use of network
bandwidth


ToR

switch

Access

Aggregation

Core

Image
-
server

RH6.0

RH5.6

RH5.5

RH5.6

7

C Peng (UCLA)

Infocom 2012

RH5.6

How to Enable Collaborative Sharing?


What is the basic data unit for sharing?


File
-
based sharing: Allow sharing only among same files


Chunk
-
based sharing: Allow sharing of common chunks from
different files



How to manage content location information?


Centralized solution: directory service, etc.


Distributed solution: P2P overlay, etc.


Infocom 2012

C Peng (UCLA)

8

What is the Appropriate Sharing Unit?


Two factors


The number of the same, alive VM image instances


The similarity of different VM images



Conduct real trace analysis and cross
-
image similarity
measurement


VM traces from six operational data centers for 4 months


VM images including different Linux/Windows versions, IBM
services (DB2, Rational,
WebSphere
) etc


Infocom 2012

C Peng (UCLA)

9

VM Instance Popularity



The distribution of image popularity is highly skewed


A few popular images take a large portion of VM instances


Many unpopular images have a small number of VM
instances (< 5)


Few peers can involve in file
-
based sharing


Infocom 2012

C Peng (UCLA)

10

Unpopular VM images

VM Instance Lifetime

Infocom 2012

C Peng (UCLA)

11


The lifetime of VM instance varies


40% instances (more popular VM instances) < 13 minutes


The unpopular VM images have longer lifetime


VM image distribution network should cope with various
lifetime instances



13 min

VM Image Structure

Infocom 2012

C Peng (UCLA)

12


Tree
-
based VM image structure

Linux


Windows


Services

Misc



(60%)

(25%)

(11%)

(4%)

Red Hat

SUSE

(53%)

……

……

Enterprise Linux v5.5 (32bit) (26.6%)

Enterprise Linux v5.5 (64bit) (18.7%)



Enterprise Linux v5.4 (32bit) (4%)



Enterprise Linux v5.6 (32bit) (0.2%)


……

Database ……

IDE ……

……


……

V7.0 B (0.7%)

V7.0.0.11 S P (0.7%)

V7.0.0.11 R B (0.3%)

V7.0.0.11 S B (0.3%)

V7.0.0.11 S D (0.2%)

V7.0 P (0.1%)

(7%)

Web app. server

VM Image Similarity


High similarity across VM images


Chunk schemes: fixed size and Rabin fingerprinting


Similarity:
Sim(A,B
) = |A’s chunks that appear in B| /|A|


Chunk
-
based sharing can exploit cross
-
image similarity



Infocom 2012

C Peng (UCLA)

13

Enable Chunk
-
based Sharing


Decouple VM images into VM chunks


Exploit similarity across VM images


Provide a higher source diversity and sharing opportunity


Infocom 2012

C Peng (UCLA)

14

RH5.6

RH6.0

RH5.6

RH5.6

RH5.6

RH5.5

Questions:

How to maintain chunk location information (metadata)

How to be scalable and also enable fast data transmission


How to Manage Location Information?


Solution I: centralized metadata server


Cons: be simple


Pros: bottleneck at metadata server


Solution II: P2P overlay network, e.g., DHT


Cons: distributed operations


Pros: be unaware of data center topology and may introduce
high network overhead


Infocom 2012

C Peng (UCLA)

15

Internet

I
-
S

Issues in Conventional P2P Practice

Infocom 2012

C Peng (UCLA)

16

One
logic

operation

(lookup/publish)

Multiple
physical

hops

Hop costs (e.g. time) can be high!

Solution:

Reduce # of hops

Reduce the cost of physical hops


Keep it local or

with close buddies


Topology
-
aware Metadata Management


Divide all the hosts into different
-
level hierarchies and
manage chunks in each hierarchy


Utilize static/quasi
-
static (controlled) topology


Exploit high bandwidth local links in hierarchical structure



Infocom 2012

C Peng (UCLA)

17

Internet

I
-
S

L1

H

H

H

L1

L1

L2

L2

L3

VDN: Encourage
Local

Communication


Local

chunk metadata storage


Index nodes maintain only metadata within this hierarchy


Unnecessary to maintain a global view at all index nodes



Local
chunk metadata operation (e.g., lookup/publish)


Ask close index nodes first


Lower operation overhead



Local

chunk data delivery


Enable high bandwidth transmission between close hosts (e.g.
within the rack)



Infocom 2012

C Peng (UCLA)

18

VDN Operation Flows


Recursive operation from lower
-
hierarchy to higher
-
hierarchy

Infocom 2012

C Peng (UCLA)

19

L2

L2

Image
-
server

Local
Cache

L3

A. Metadata update

B. Metadata lookup

C. Data transmission

L1

L1

L1

L1

1.

2.

3A.

3B.

3C.

4A

4B

5

Performance Evaluation


Setting


One
-
month real trace driven simulation


VM image: 128MB~ 8GB


Tree topology: 4x 4
x

8 (128 nodes)



Network bandwidth:


Static throughput for one physical link


Queue
-
based simulation for multiple transmissions on one link



Schemes


Baseline: centralized operation


Local: fetch VM chunks from local host if possible


VDN: enable collaborative sharing


Infocom 2012

C Peng (UCLA)

20

I
-
S

disk I/O: 1Gbps

Net BW: 1Gbps

2Gbps

500Mbps

200Mbps

(4
-
)

(8
-
nodes)

Great Speedup on Image Distribution

C Peng (UCLA)

21


at S6,

Infocom 2012

S1 data center

S6 data center

VM image size = 4GB

Scalable to Heavy Traffic Loads


Adjust time
-
of
-
arrival using factor 1
-
60


C Peng (UCLA)

22

S6, Median

S6, 90th

Infocom 2012

Low Metadata Management Overhead


Compare with three metadata management schemes


Naïve: on
-
demand topology
-
aware broadcast


Flat: manage metadata in a ring (e.g. DHT, P2P)


Topo
: topology
-
aware design (VDN)


Assume the communication cost is 1:4:10 (reverse to
bandwidth)




23

(a) Number of messages

(
b
) Communication cost

C Peng (UCLA)

Infocom 2012

Conclusion


VDN is a network
-
aware P2P paradigm for VM image
distribution


Reduce image provisioning time


Achieve the reasonable overhead



Chunk
-
based sharing exploit inherent cross
-
image
similarity



Network
-
aware operations can optimize the performance
in the context of data centers



C Peng (UCLA)

24

Infocom 2012

THANKs

C Peng (UCLA)

25

Infocom 2012