The Power of Prediction:

Cloud Bandwidth and Cost

Reduction

Eyal Zohar Israel Cidon

Technion

Osnat (Ossi) Mokryn

Tel-Aviv College

Traffic Redundancy Elimination (TRE)

Traffic redundancy stems from downloading

same or similar information items.

We found around 70% redundancy in

end-clients traffic, compared with past traffic

and local files.

2SIGCOMM 2011

TRE Importance

Moving to the cloud => higher e2e traffic.

Cloud users pay for traffic used in practice =>

incentive to use TRE.

Cloud ProviderCloud Provider

Cloud User

Pay for Use

End-user

Application

TRE

Cloud Traffic

3SIGCOMM 2011

How TRE Works

Chunk 1 Chunk 2 Chunk 3

Byte stream

Anchor 1 Anchor 2 Anchor 3 Anchor 4

Sign. 1 Sign. 2 Sign 3

Rolling hash

SHA-1 signature

Server parses the outgoing stream to content-

based chunks and signs with SHA-1

Chunk 1 Chunk 2’ Chunk 3Insertion example

New bytes

4SIGCOMM 2011

Problems in Existing Solutions

In the cloud environment:

1. High processing costs in the cloud.

2. Scalability – remember each client.

3. Elasticity - unaware of data from other sources.

4. Do not handle long-term repeats (days/weeks).

ReceiverServer 1

Server 2

5SIGCOMM 2011

Our Solution: PACK (Predictive ACK)

Redundancy detection by the client.

Repeats appear in chains.

Tries to match incoming chunks with a

previously received chain or local file.

Sends to the server predictions of the future

data.

6SIGCOMM 2011

PACK: The Client Prediction

Each prediction:

1.TCP seq. – no server parsing

2.Hint – spare unnecessary SHA-1

3.SHA-1 signature

Chunk

SHA-1

7

Last-byte

hint

TCP seq.

SIGCOMM 2011

Chunk 1 Chunk 2 Chunk 3

Sign. 1 Sign. 2 Sign 3

SHA-1 signature

Stream chunks

Chain of chunks

Received Prediction

3

PACK: Server Operation

The server compares the hint with the last-byte to sign.

Upon a hint match it performs the expensive SHA-1.

PACK saves cloud’s computational effort in the absence

of redundancy.

First receiver-based TRE: the server does not parse. It

signs with >99% confidence.

ClientServer

Local

storage

1 21 2 3

2,3?

Chain

2,3V

8SIGCOMM 2011

PACK Benefits

Minimizes processing costs induced by TRE.

– Signs with SHA-1 in the presence of redundancy.

Receiver-based end-to-end TRE => suitable for cloud

server elasticity and client mobility.

– Does not require the server to continuously maintain

clients’ status.

9SIGCOMM 2011

Server Effort Experiment

Several data-sets in 3 modes: baseline no-TRE,

PACK and a sender-based TRE.

0%

20%

40%

60%

80%

100%

120%

140%

0% 10% 20% 30% 40% 50%

Sin

gle

Serv

er

Clo

ud O

pera

tional C

ost

(100%

=w

ithout

TR

E s

yste

m)

Redundancy Elimination Ratio

EndRE-like

PACK

25%-30% redundancy:

common to many

data-sets

10SIGCOMM 2011

Sender-based

0%

5%

10%

15%

20%

25%

30%

35%

0.0

0.5

1.0

1.5

2.0

2.5

3.0

PA

CK

TR

E (

Rem

oved R

edundancy)

All

YouT

ube T

raff

ic (

Gbps)

Time (24 hours)

YouTube Traffic

PACK TRE

YouTube Redundancy

Traces of 40k clients, captured at an ISP.

Found 30% end-to-end (personal) redundancy.

11SIGCOMM 2011

0%

10%

20%

30%

40%

50%

60%

70%

80%

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33

Avera

ge R

edundancy o

f D

aily

Tra

ffic

Days Since Start

Unlimited

1 Hour

24 Hours

Social network: eliminated 30% with one hour cache

and 75% with a long-term cache.

12

Long-Term TRE

SIGCOMM 2011

Gmail account with 1,000 Inbox messages.

Found 32% static redundancy (higher when

messages are read multiple times).

Cloud Email Redundancy

0

50

100

150

200

250

300

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Tra

ffic

Vo

lum

e P

er

Mo

nth

(M

B)

Month

Redundant

Non-redundant

13SIGCOMM 2011

Implementation

Linux with Netfilter Queue, 25k lines of C and

Java, available for download.

Receiver-sender protocol is embedded in the

TCP Options field.

Transparent use at both sides.

14SIGCOMM 2011

Processing Effort in the Client

SIGCOMM 2011 15

Laptop experiment: PACK-related CPU consumption is

~4% when playing HD video (9 Mbps with 30%

redundancy).

Smartphone experiment: PACK consumes ~3% of the

battery power when processing 1 GB video (avg.

monthly data plan).

Virtual traffic saves

the client the need

to chunk or sign.

New Chunking Algorithm

SIGCOMM 2011 16

Most existing solutions use Rabin fingerprint.

New Chunking Algorithm64 bits

nn-1

n-2n-3

n-4

n-5n-6

n-7n-8

n-40n-41

n-42n-43

n-44n-45

n-46n-47

Mask=00 00 8A 31 10 58 30 80

SIGCOMM 2011 17

Summary

Current TRE solutions may not reduce cloud cost.

PACK is the first receiver-based TRE – leverages the

power of prediction.

Minimizes processing costs induced by TRE.

Suitable for cloud server migration and client mobility.

Implementation is available for download.

18SIGCOMM 2011