Transcript or “Tipping Point Protocols”

(gatorlog.com)
CS 525
Advanced Distributed Systems
Spring 09
(epath.org)
Indranil Gupta
Lecture 7
More on Epidemics (or “Tipping Point Protocols”)
February 12, 2009
1
Question…
What fraction of main roads need to be randomly knocked out before
source and destination are completely cut off?
Source
Destination
2
Tipping Point!
Critical Value? Answer = 0.5
Source
Destination
(Comes from Percolation Theory)
3
“Tipping Point”
[Malcolm Gladwell, The Tipping Point, Little
Brown and Company, ISBN: 0316346624]
Tipping is that (magic) moment when an idea,
trend or social behavior crosses a threshold,
and spreads like wildfire.
4
Epidemic Protocols
•
•
•
•
•
A specific class of tipping point protocols
Local behavior at each node – probabilistic
Determines global, emergent behavior at the scale of the
distributed system
As one tunes up the local probabilities, the global
behavior may undergo a threshold behavior (or, a phase
change)
Three papers:
1. Epidemic algorithms
2. Bimodal multicast
3. PBBF (sensor networks)
5
Epidemic Algorithms for Replicated
Database Maintenance
Alan Demers et. al.
Xerox Palo Alto Research Center
PODC 1987
[Some slides borrowed from presentation by: R. Ganti and P. Jayachandran]
6
Introduction
• Maintain mutual consistency of updates in a
distributed and replicated database
• Used in Clearinghouse database – developed in Xerox
PARC and used for many years
• First cut approaches
– Direct mail: send updates to all nodes
• Timely and efficient, but unreliable
– Anti-entropy: exchange database content with random site
• Reliable, but slower than direct mail and uses more resources
– Rumor mongering: exchange only ‘hot rumor’ updates
• Less reliable than anti-entropy, but uses fewer resources
7
(from Lecture 1)
Epidemic Multicast
Infected
Protocol rounds (local clock)
b random targets per round
Gossip Message (UDP)
Uninfected
8
Epidemic Multicast (Push)
Infected
Protocol rounds (local clock)
b random targets per round
Gossip Message (UDP)
Uninfected
9
Epidemic Multicast (Pull)
Infected
Gossip Message (UDP)
Uninfected
Protocol rounds (local clock)
b random targets per round
10
Pull > Push
pi – Probability that a node is susceptible after the ith round
pi
pi
1
1
2
i
p
Pull
1
pi 1
n
n 1
pi
Push
• Pull converges faster than push, thus providing
better delay
• Push-pull hybrid variant possible (see Karp and
Shenker’s “Randomized Rumor Spreading”)
11
Anti-entropy: Optimizations
• Maintain checksum, compare databases if
checksums unequal
• Maintain recent update lists for time T,
exchange lists first
• Maintain inverted index of database by
timestamp; exchange information in reverse
timestamp order, incrementally re-compute
checksums
12
Epidemic Flavors
• Blind vs. Feedback
– Blind: lose interest to gossip with probability
1/k every time you gossip
– Feedback: Loss of interest with probability 1/k
only when recipient already knows the rumor
• Counter vs. Coin
– Coin: above variants
– Counter: Lose interest completely after k
unnecessary contacts. Can be combined with
blind.
• Push vs. Pull
13
Deletion and Death Certificates
• Absence of item does not spread; On the contrary, it
can get resurrected!
• Use of death certificates (DCs) – when a node
receives a DC, old copy of data is deleted
• How long to maintain a DC?
– Typically twice (or some multiple of) the time to spread the
information
– Alternately, use Chandy and Lamport snapshot algorithm to
ensure all nodes have received
– Certain sites maintain dormant DCs for a longer duration;
re-awakened if item seen again
14
Performance Metrics
• Residue: Fraction of susceptibles left when
epidemic finishes
• Traffic: (Total update traffic) / (No. of sites)
• Delay: Average time for receiving update
and maximum time for receiving update
• Some results:
– Counters and feedback improve delay
– Pull provides lower delay than push
15
Performance Evaluation
Tipping Point Behavior
16
Discussion
Pick your favorite:
• Push vs. pull vs. push-pull
– Name one disadvantage of each
• Direct mail vs. anti-entropy vs. rumor mongering
– Name one disadvantage of each
• Random neigbhor picking
– Disadvantage in wired networks?
– In Sensor network?
17
Bimodal Multicast
Kenneth P. Birman et. al.
ACM TOCS 1999
[Some slides borrowed from presentation by: W. Fagen and L. Cook]
18
“Traditional” Multicast Protocols
19
Vs. Pbcast
Traditional Multicast
• Atomicity: All or none
delivery
•
• Multicast stability:
Reliable immediately
delivery of messages
• Scalability: Bad. Costs >=
quadratic with group size.
•
• Ordering
•
•
Pbcast
Atomicity: Bimodal
delivery guarantee, almost
all or almost none
(immediately)
Multicast stability:
Reliable eventual delivery
of messages
Scalability: Costs
logarithmic w.r.t. network
size. Throughput stability.
Ordering
20
Pbcast: Probabilistic Broadcast
Protocol
•
Pbcast has two stages:
1. Unreliable, hierarchical, best-effort broadcast.
Eg. IP Multicast
2. Two-phase anti-entropy protocol: runs
simultaneously with the broadcast messages
•
•
First phase detects message loss
Second phase corrects such losses
21
The second stage
• Anti-entropy round:
– Gossip Messages:
• Each process chooses another random process and sends a summary
of its recent messages
– Solicitation Messages:
• Messages sent back to the sender of the gossip message requesting a
resend of a given set of messages (not necessarily the original source)
– Message Resend:
• Upon reception of a solicitation message, the sender resends that
message
• Protocol parameters at each node
– # of rounds and # of processes contacted in each round
– Product of above two parameters called fanout
22
Optimizations
• Soft-Failure Detection: Retransmission requests served
only if received recently; protects against congestion
caused due to redundant retransmissions
• Round Retransmission Limit: Limit the no. of
retransmissions in a round; spread overhead in space and
time
• Most-Recent-First Retransmission: prefer recent messages
• Independent Numbering of Rounds: Allows delivery and
garbage collection to be entirely a local decision
• Multicast for Some Retransmissions
23
Bimodality of Pbcast
Logarithmic
Y-axis
Almost none
Almost all
24
Latency for Delivery
Logarithmic growth
25
Throughput Comparison
26
Discussion
• Disadvantages of Bimodal Multicast?
– When would wasteful messages be sent?
• What happens when
– Rate of injection of multicasts is very very low?
– IP multicast is very very reliable?
– IP multicast is very very unreliable?
27
PBBF: Probability-Based
Broadcast Forwarding
Cigdem Sengul and Matt Miller
ICDCS 2005 and ACM TOSN 2008
(Originated from a 525 Project)
28
Broadcast in an Ad-Hoc Network
• Ad-hoc sensor network (Grid example below)
• One node has a piece of information that it needs
to broadcast: e.g., (1) code update, (2) query
• Simple approach: each node floods received
message to all its neighbors
– Disadvantages?
29
IEEE 802.11 PSM
A real, stable MAC protocol (similar results for SMAC, T-MAC, etc.)
• Nodes are assumed to be synchronized
• Every beacon interval (BI), all nodes wake up for
an ATIM window (AW)
• During the AW, nodes advertise any traffic that
they have queued
• After the AW, nodes remain active if they expect
to send or receive data based on advertisements;
otherwise nodes return to sleep until the next BI
30
Protocol Extreme #1
N1
N1
N2
A
N2
N3
D
A
D
A
N3
A = ATIM Pkt
D = Data Pkt
31
Protocol Extreme #2
N1
N1
N2
N3
A
N2
N3
D
D
D
A = ATIM Pkt
D = Data Pkt
32
Probability-Based Broadcast
Forwarding (PBBF)
• Introduce two parameters to sleep
scheduling protocols: p and q
• When a node is scheduled to sleep, it will
remain active with probability q
• When a node receives a broadcast, it
rebroadcasts immediately with probability p
– With probability (1-p), the node will wait and
advertise the packet during the next AW before
rebroadcasting the packet
33
• Phase transition
when:
pq + (1-p) ≈ 0.8-0.85
• Larger than
traditional bond
percolation threshold
– Boundary effects
– Different metric
• Still shows phase
transition
Fraction of Broadcasts
Received by 99% of Nodes
Analysis: ReliabilityTipping Point!
q
34
Application: Energy and Latency
Energy
Joules/Broadcast
Latency
Average 5-Hop Latency
Increasing p
q
q
≈ 1 + q * [(BI - AW)/AW]
Ns2 simulation: 50 nodes, uniform placement, 10 avg. neighbors
35
Energy
Adaptive PBBF
Achievable
Region
Latency
36
Adaptive PBBF (TOSN paper)
• Dynamically adjusting p
and q to converge to userspecified QoS metrics
– Code updates prefer
reliability overl latency
– Queries prefer latency
over reliability
• Can specify any 2 of
energy, latency, and
reliability
• Subject to those
constraints, p and q are
adjusted to achieve the
highest reliability possible
1.0
0.5
q
p
0.0
Time
37
Discussion
• PBBF: bond percolation (remove roads from city)
• Haas et al paper (Infocom): site percolation
– Remove intersections/junctions (not roads) from city
• Site percolation and bond percolation have
different thresholds and behaviors
• Hybrid possible? (like push-pull?)
• What about over-hearing optimizations? (like
feedback)
38
Question…
Are there other tipping point protocols…?
Source
Destination
39
Next Week Onwards
• Student Presentations start (see instructions)
• Reviews needed (see instructions)
• Project Meetings start (see newsgroup)
– Think about which testbed you need access to:
PlanetLab, Emulab, Cirrus
• Tomorrow: Yahoo! Training seminar
40