Upload
wade-mcfarland
View
17
Download
0
Embed Size (px)
DESCRIPTION
Digital Fountains. -Anand Patwardhan. Main Ideas : Distribution of bulk data Reliable multicast, broadcast Ideal digital fountain Erasure codes RMDP Benefits and applicability. CSE 581, Winter 2002 Instructor : Wu-chang Feng. Paper group. - PowerPoint PPT Presentation
Citation preview
Jump to first page
A. Patwardhan, CSE 581
1
Digital Fountains
Main Ideas : Distribution of bulk data Reliable multicast, broadcast Ideal digital fountain Erasure codes RMDP Benefits and applicability
CSE 581, Winter 2002
Instructor : Wu-chang Feng
-Anand Patwardhan
Jump to first page
A. Patwardhan, CSE 581
2
Paper group “A Digital Fountain approach to reliable
distribution of bulk data”, J. Byers, M. Luby et al. ( Feb. ’98)
“Accessing multiple mirror sites in parallel: Using Tornado codes to speed up downloads”, J. Byers, M. Luby, M. Mitzenmacher
“RMDP: An FEC-based Reliable Multicast Protocol for Wireless Environments”, L. Rizzio, L. Vicisano (April ’98)
Jump to first page
A. Patwardhan, CSE 581
3
Reliable distribution of Bulk Data Multicasting with
feedback Reliability : ARQ -
Some solutions use NACK suppression , local recovery – overhead
Scalability – complexity in maintaining group hierarchy
Feedback channel required
Unicasting Reliability : uses ARQ –
Automatic retransmission request
Scalability - Suffers from “Feedback implosion at source”
Efficiency – “repair” packets do not benefit everyone
Requires a feedback channel
Jump to first page
A. Patwardhan, CSE 581
4
An Ideal “Digital Fountain” A Server serving a universe of clients Source data = k packets is encoded to n
= c.k packets (c>1) Server carousels through a stream of n
packets (uses multicast ) Clients drink their fill Clients quenched by any k subset of the
n packets, then disconnect Clients can “drink” at anytime
(asynchronous)
Jump to first page
A. Patwardhan, CSE 581
5
Erasure Codes Also known as Forward Error-
Correcting codes (FEC) Most commonly used : Reed-
Solomon erasure codes k = source data packets encoded n = k + l = c.k
c = stretch factor l = redundant packets
Encoding/decoding complexity increases proportional to k*l*(packetsize)
Jump to first page
A. Patwardhan, CSE 581
6
Tornado Codes
Much simpler than Reed-Solomon Use XOR only Encoding/decoding uses random bipartite graphs n = k+l, but slightly more than k packets have to
be received at the receiver for decoding ( reception inefficiency)
Fast encoding/decoding at the price of reception inefficiency ( usually around 5%)
Encoding/Decoding complexity proportional to (k+l)ln(1/e)*(packetsize)
Jump to first page
A. Patwardhan, CSE 581
7
Comparison of encoding/decoding times
Jump to first page
A. Patwardhan, CSE 581
8
Reception overhead
90% clients were done at 0.06For Tornado A
Jump to first page
A. Patwardhan, CSE 581
9
Comparison of reception efficiency for codes with comparable decoding time
p = probability of loss
Interleaved = data broken into segments and then encodedusing Reed-Solomon codes, k = no of segments
Jump to first page
A. Patwardhan, CSE 581
10
Comparison of reception efficiency for codes with increasing filesize
Jump to first page
A. Patwardhan, CSE 581
11
Simple Mirroring
User has to pick a single site Access intervals often overlap Many to many distribution not
possible
Jump to first page
A. Patwardhan, CSE 581
12
Parallel download from multiple mirror sites Digital fountain mirrors
Scalable, efficient, reliable, tolerant and on-demand
Client collects packets from multiple senders until “quenched”
“Out-of-step” senders and increased stretch factor (n = c.k) minimize duplicate packets
All available bandwidth utilized to speed up download (concept of “disjoint bottlenecks”)
Effectively “Many-to-many” distribution
Jump to first page
A. Patwardhan, CSE 581
13
Reception inefficiency, Speedup & stretch
Duplicates decrease Speedup Increases
Jump to first page
A. Patwardhan, CSE 581
14
RMDP protocol Reliable Multicast data Distribution Protocol Similar to approach of Digital fountains Uses Reed-Solomon, with limited ARQ Authors contend that computational complexity
of Reed-Solomon better than resource expensive Tornado codes in the longer term ( CPUs will improve...)
Optimal values of n,k,l used, claim : expensive but adequate
Feedback suppression, using rules for timeouts
Jump to first page
A. Patwardhan, CSE 581
15
Reliable multicast using erasure codes
Benefits Reliability with little or no feedback Highly scalable Clients can join asynchronously Ideal for wireless, satellite, cable
transmissions with little or non-existent feedback channel
Performance does not decrease with increase in receivers, due to correlated losses
Resilient to bursty losses
Drawbacks Zero feedback at the
cost of bandwidth, buffers and time
Encoding, decoding overhead
Increased buffers Reception inefficieny Data available only after
a minimum set of distinct packets arrive
Does not consider CC
Jump to first page
A. Patwardhan, CSE 581
16
Summary Reliability achieved by “carouseling”
packets encoded using erasure codes A single packet from a block is potentially
useful for a large subset of the receivers Both RMDP and the Digital fountain
approach, use multicasting and erasure codes
Tornado codes simpler to decode/encode but memory requirement is non-deterministic (shown to fare better than RS)
Reed-Solomon codes have fixed memory requirement, but computationally very expensive.
Jump to first page
A. Patwardhan, CSE 581
17
Questions …