SCAP: Smart Caching in Wireless Access Points to Improve P2P Streaming Enhua Tan1, Lei Guo1, Songqing Chen2, Xiaodong Zhang1 1The Ohio State University 2George Mason University ICDCS’07, Toronto,

Download Report

Transcript SCAP: Smart Caching in Wireless Access Points to Improve P2P Streaming Enhua Tan1, Lei Guo1, Songqing Chen2, Xiaodong Zhang1 1The Ohio State University 2George Mason University ICDCS’07, Toronto,

SCAP: Smart Caching in Wireless
Access Points to Improve P2P
Streaming
Enhua Tan1, Lei Guo1, Songqing Chen2,
Xiaodong Zhang1
1The
Ohio State University
2George Mason University
ICDCS’07, Toronto, Canada
1
Background

Wireless access to Internet is pervasive:



On campus, in offices, at home, and public utilities
Most are supported by Wireless LANs
Peer-to-Peer applications are widely used:



Streaming: PPLive, Joost, etc …
VoIP: Skype, etc …
Large file distribution: BitTorrent, etc …
Our Focus: Interaction between wireless
users and P2P streaming applications
2
Wired/wireless Communications
Internet
WLAN
Access Point
(AP)
Wired users
Wireless users
3
P2P Streaming for Wired/wireless Users:
Workflow
Internet
Access
Point
Source Peer
Viewing Peer
Wireless Peer
WLAN
4
P2P Streaming for Wired/wireless Users:
Problems
Downstream traffic
for other wireless
users AFFECTED
Internet
Source Peer
WLAN
Generating
upstream
traffic
Streaming content
Wireless Peer
(Relay/Viewing)
Other packets
Viewing Peer
Streaming quality
degraded
5
Problem Summary

Peers in WLAN may relay streaming content by
uploading a lot of traffic:


Congest the WLAN due to channel competitions
Provide low quality of service to the Internet peers
Downstreams have lower priority due to
upstreams
 Extra
upstream
traffic:
Major
problem
source:
upstream relay traffic

 further
increaseupstream
the number
of transmission
errors
 Can
we minimize
traffic
with low overhead?
 to
increase
cost
of contention window back-off
improve the
WLAN
throughput


to improve service quality for Internet peers
6
P2P Streaming for Wired/wireless Users:
Workflow
The same content is
transferred twice in
the WLAN!
 Duplicated traffic
Internet
Access
Point
Source Peer
Viewing Peer
Wireless Peer
WLAN
7
Contributions



Our measurements show that > 75% upstream
traffic is duplicated with the downstream traffic
for three representative applications
SCAP: Smart Caching in the Access Point for
minimizing upstream traffic: design & prototype
implementation
Evaluation results show SCAP can improve the
throughput of the WLAN by up to 88%:

SCAP also reduces the delay to Internet peers
8
Outline


Problem Summary and Contributions
Measurement & Analysis of P2P Streaming
Traffic

SCAP Design & Implementation

Evaluation

Summary
9
Measurement & Analysis of P2P
Streaming Traffic

Aim to answer two questions:



How much duplicated traffic in practice?
How much overhead in identifying such
duplications?
Measurement:


Collect traces of three representative P2P live
streaming applications: PPLive, ESM, and TVAnts
In LAN (100Mbps) and WLAN (802.11b)
10
Workload Statistics

Downstream throughput is typically 300~400Kbps

Upstream traffic to downstream traffic:


Can be as large as 10 times for PPLive due to its popularity

Between 2 to 4 times for TVAnts

Not too much for ESM
PPLive and ESM: most in TCP

TVAnts: 74% in UDP for WLAN
11
Duplication Detection Methods:
Fixed Hashing

Offline workload analysis:

Fixed Hashing (FH)

Compute only 1 fingerprint (hash value) for a downstream
packet; store this fingerprint in a hash table, and cached the packet
in FIFO buffer

For each upstream packet, also compute the fingerprint, and look
it up in the hash table to locate the duplicated downstream
packet; If found the same fingerprint, do further byte-to-byte
comparison
Upstream
packet
Upstream
packet fingerprint
Lookup
Downstream
packet
fingerprint
hash table
Downstream packet
FIFO buffer
12
Duplication Detection Methods:
Rabin Fingerprinting

Rabin Fingerprinting (RF)


A unique hash function: produce fingerprints for a
continuous data stream quickly (NSDI’07 BitTyrant)
We scan the whole packet and only store fingerprints
ending with 8 zeros over 64 bytes content


averagely 5 fingerprints for a 1400 bytes packet (1/28)
FIFO Buffer: stores latest 50,000 downstream
packets

Buffer + hash table: need about 75MB memory totally
13
100
Dup Ratio & Tput
90
70
Offline analysis
processing throughput of
RF is less than FH:

60
50
RF
FH
40
RF-BufAll
30
20

10
0
PPLLAN


PPLWL
TVALAN
TVAWL
ESMLAN
ESMWL
RF can detect more
duplications than FH
All the duplication
ratios are larger than
75%
Still large enough (>
90Mbps) for process
P2P streaming (400
Kbps)
450
400
350
Throughput (Mbps)
Duplication ratio(%)
80
300
250
RF
200
FH
RF-BufAll
150
100
50
0
PPLLAN
PPLWL
TVALAN
TVAWL
ESMLAN
ESMWL
14
Duplication Beginning Offset



FH can only detect the
duplication when the offsets for
up/downstream packets are the
same (no re-packetizing)
ESM does not have any offset
differences  FH performs well
TVAnts has a lot of repacketizing  FH performs the
15
worst
Forwarding Delay
200
seconds
200
seconds
10
seconds

PPLive and TVAnts: most
upstream packets
forwarded in 200 seconds



20
seconds
10 ms
<20 seconds for 70%
ESM: within 10 ms
Implies the downstream
buffer can be quite small
16
Outline


Problem Summary and Contributions
Measurement & Analysis of P2P Streaming
Traffic

SCAP Design & Implementation

Evaluation

Summary
17
SCAP (Smart Caching in Access Points)
Overview
Internet
Access
Point
Metadata upstream packet
Downstreams
buffer
(If duplications found in
downstream buffer)
Relay/Viewing
Peer
Original upstream packet
Downstream
buffer
WLAN
18
Design Issues

Buffer size:


Need 7.5MB for storing recent 200 seconds traffic (in
300Kbps rate), which is affordable for a wireless station
But AP will need to buffer for multiple stations:


AP should dynamically adjust the buffer space for each station
according to its duplication ratios in order to achieve highest traffic
reduction with limited buffer space
Buffer synchronization between AP and station:

If a metadata upstream packet cannot be reassembled on
AP due to a cache miss, TCP flow will be stalled

Wireless station caches several copies of recent sent upstream
packets and resends the uncompressed packet when needed
19
Prototype Implementation

Modified HostAP driver in Linux kernel 2.6.16
for the AP and stations


Wireless card is based on Intersil Prism 2.5 chipset
(802.11b)
Identification of the downstream packet



For AP to locate the packet in decompressing the
upstream packet
Cannot use Sequence Control field (2 bytes)
because it is filled by the firmware
Have to use the first fingerprint value (8 bytes)
20
Outline


Problem Summary and Contributions
Measurement & Analysis of P2P Streaming
Traffic

SCAP Overview

Design & Implementation

Evaluation

Summary
21
Performance Evaluation: LAN Experiment
4.50
4.43 Mbps
4.5
4
3.5
3
2.5
2
1.5
1
0.5
1MB
7MB
70MB
Transfer File Size


8
7
6
4.7
5
Orig
Orig
RF
RF
3FH
FH
4
2
1
0
0

8.9 Mbps
9
Upstream Throughput (Mbps)
Downstream Throughput(Mbps)
5
140MB
1MB
7MB
70MB
140MB
Transfer File Size
Station first receives a file from a server, then sends it back
RF: little overhead for the downstream throughput (1.5%
decrease), and 88% improvement for the upstream throughput
FH: cannot have any improvement due to constant TCP repacketizing
22
Performance Evaluation:
Internet Experiment

Evaluate PPLive, TVAnts, and ESM


Run the applications in a VMWare-based Windows XP
guest OS for HostAP driver to work
Measurement methods:

Because P2P Streaming is a Constant Bit Rate stream:




 Upstream throughput will not change even if we reduces its traffic
Running iperf on another wireless station to observe the impact to
WLAN TCP throughput
Running Ping to observe the impact to response time
Run multiple trials to get comparable P2P downstream
throughput for comparison

Each trial runs for 600 seconds
23
Internet Experiment:
Evaluation Results

RF/FH performs
best for TVAnts
since it has the
largest volume of
upstream traffic:



Increases TCP
throughput by 0.95
Mbps (54% of
upstream traffic)
Decrease Ping
round-trip time by 83
ms (-26%)
Also performs well
for PPLive/ESM
24
Summary




With the increasing popularity of P2P streaming
applications and pervasive deployment of 802.11 WLANs,
more peers will be connected by wireless
We study the impact of wireless peers to the performance
of wireless and Internet users
 Without a proper control of P2P traffic, the
performance of both parties can be significantly
affected
We designed and implemented SCAP (Smart Caching in
Access Points) in order to reduce the upstream traffic for
P2P live streaming applications
Our prototype based evaluation shows the effectiveness
of SCAP:


SCAP improves the throughput of the WLAN by up to 88%
SCAP reduces the response delay to Internet peers as well
25
Thank you!
Enhua Tan: [email protected]
http://www.cse.ohio-state.edu/hpcs/
26
SCAP (Smart Caching in Access Points)
– Basic Idea
Inco
mi n
g
oin
g
t
u
O
g
(1)
(2)
(4)
(3)
Access Point (AP)
Wireless Station

AP stores downstream data in buffer (1)

Station stores downstream data in buffer (2)

Compare upstream packet (3) with (2), upload difference (4)

AP will assemble upstream packet with data in (1) to the
Internet
27
Workflow of SCAP
Downstream
Buffer
HostAP Driver
Downstream packet
Decompressing
Router
Uptream packet
Access
Point
Compressed
upstream packet
Downstream
Buffer
Wireless
Station
Lookup
Duplication
Detection;
Compressing
P2P Streaming Application
28
Rabin Fingerprinting

Rabin Fingerprinting (RF)
A  (a1, a2 ,...,am )
can produce fingerprints
for a continuous data
m1
m2
A(t )  a1t  a2t  ... am
stream quickly:


Advance the fingerprint only
requires an addition, a
multiplication, and a mask
Lack of this property for
other hash functions like
MD5/SHA (and they are
also more complex)
RF( A)  A(t ) modP(t )
29
Some Related Work

XORs in the Air: Practical Wireless Network Coding
(Sigcomm’06)



A Protocol-Independent Technique for Eliminating
Redundant Network Traffic (Sigcomm’00)


Utilizing the broadcasting nature of wireless networks to
improve throughput of multi-hop network (instead of
application characteristics)
Our scheme is utilizing the traffic pattern of P2P
applications
reduces redundant traffic using Rabin Fingerprinting
A Low-bandwidth Network File System (SOSP’01)

Exploits similarities between different versions of a file to
reduce update traffic
30