Leiden; Dec 06Gossip-Based Networking Workshop1 Gossip Algorithms and Emergent Shape Ken Birman.
-
Upload
philomena-fowler -
Category
Documents
-
view
214 -
download
0
Transcript of Leiden; Dec 06Gossip-Based Networking Workshop1 Gossip Algorithms and Emergent Shape Ken Birman.
Leiden; Dec 06 Gossip-Based Networking
Workshop 1
Gossip Algorithms and Emergent Shape
Ken Birman
Leiden; Dec 06 Gossip-Based Networking
Workshop 2
On Gossip and Shape
Why is gossip interesting? Scalability….
Protocols are highly symmetric Although latency often forces us to bias
them Powerful convergence properties…
Especially in support of epidemics New forms of consistency…
Probabilistic… but this is often adequate
Leiden; Dec 06 Gossip-Based Networking
Workshop 3
Consistency Captures intuition that if A and B compare
their states, no contradiction is evident In systems with “logical” consistency, we say
things like “A’s history and B’s are closed under causality, and A is a prefix of B”
With probabilistic systems we seek rapidly decreasing probability (as time elapses) that A knows “x” but B doesn’t
…. Probabilistic convergent consistency
Leiden; Dec 06 Gossip-Based Networking
Workshop 4
Exponential convergence
A subclass of convergence behaviors
Not all gossip protocols offer exponential convergence But epidemic protocols do have this
property… and many gossip protocols implement epidemics
Leiden; Dec 06 Gossip-Based Networking
Workshop 5
Value of exponential convergence
An exponentially convergent protocol overwhelms mishaps and even attacks Requires that new information reach relevant
nodes with at most log(N) delay Can talk of “probability 1.0” outcomes
Even model simplifications (such as idealized network) are washed away! Predictions rarely “off” by more than a
constant A rarity: a theory relevant to practice!
Leiden; Dec 06 Gossip-Based Networking
Workshop 6
Convergent consistency
To illustrate our point, contrast Cornell’s Kelips system with MIT’s Chord
Kelips is convergent. Chord isn’t
Leiden; Dec 06 Gossip-Based Networking
Workshop 7
Kelips (Linga, Gupta, Birman)
30
110
230 202
Take a a collection of
“nodes”
Leiden; Dec 06 Gossip-Based Networking
Workshop 8
Kelips
0 1 2
30
110
230 202
1N
N
members per affinity group
Map nodes to affinity
groups
Affinity Groups:peer membership thru
consistent hash
Leiden; Dec 06 Gossip-Based Networking
Workshop 9
Kelips
0 1 2
30
110
230 202
Affinity Groups:peer membership thru
consistent hash
1N
Affinity group pointers
N
members per affinity group
id hbeat rtt
30 234 90ms
230 322 30ms
Affinity group view
110 knows about other members – 230, 30…
Leiden; Dec 06 Gossip-Based Networking
Workshop 10
Affinity Groups:peer membership thru
consistent hash
Kelips
0 1 2
30
110
230 202
1N
Contact pointers
N
members per affinity group
id hbeat rtt
30 234 90ms
230 322 30ms
Affinity group view
group contactNode
… …
2 202
Contacts
202 is a “contact” for 110 in group
2
Leiden; Dec 06 Gossip-Based Networking
Workshop 11
Affinity Groups:peer membership thru
consistent hash
Kelips
0 1 2
30
110
230 202
1N
Gossip protocol replicates data
cheaply
N
members per affinity group
id hbeat rtt
30 234 90ms
230 322 30ms
Affinity group view
group contactNode
… …
2 202
Contacts
resource info
… …
cnn.com 110
Resource Tuples
“cnn.com” maps to group 2. So 110 tells group 2 to
“route” inquiries about cnn.com to it.
Leiden; Dec 06 Gossip-Based Networking
Workshop 12
How it works
Kelips is entirely gossip based! Gossip about membership Gossip to replicate and repair data Gossip about “last heard from” time
used to discard failed nodes Gossip “channel” uses fixed
bandwidth … fixed rate, packets of limited size
Leiden; Dec 06 Gossip-Based Networking
Workshop 13
How it works
Heuristic: periodically ping contacts to check liveness, RTT… swap so-so ones for better ones.
Node 102
Gossip data stream
Hmm…Node 19 looks like a much better contact in affinity
group 2
175
19
RTT: 2
35m
s
RTT: 6 ms
Node 175 is a contact for Node 102 in some affinity
group
Leiden; Dec 06 Gossip-Based Networking
Workshop 14
Work in progress…
Prakash Linga is extending Kelips to support multi-dimensional indexing, range queries, self-rebalancing
Kelips has limited incoming “info rate” Behavior when the limit is continuously
exceeded is not well understood. Will also study this phenomenon
Leiden; Dec 06 Gossip-Based Networking
Workshop 15
Replication makes it robust Kelips should work even during
disruptive episodes After all, tuples are replicated to N
nodes Query k nodes concurrently to overcome
isolated crashes, also reduces risk that very recent data could be missed
… we often overlook importance of showing that systems work while recovering from a disruption
Leiden; Dec 06 Gossip-Based Networking
Workshop 16
Chord (MIT group)
The MacDonald’s of DHTs A data structure mapped to a
network Ring of nodes (hashed id’s) Superimposed binary lookup trees Other cached “hints” for fast lookups
Chord is not convergently consistent
Leiden; Dec 06 Gossip-Based Networking
Workshop 17
Chord picture0
123
199
202
241
255
248
108
177
64
30
Cach
ed
link
Finger links
Leiden; Dec 06 Gossip-Based Networking
Workshop 18
Chord picture
0
123
199
202
241
255
248
108177
64
30
EuropeUSA
0
123
199
202
241
255
248
108177
64
30
Transient Network Partition
Leiden; Dec 06 Gossip-Based Networking
Workshop 19
… so, who cares? Chord lookups can fail… and it suffers
from high overheads when nodes churn Loads surge just when things are already
disrupted… quite often, because of loads And can’t predict how long Chord might
remain disrupted once it gets that way Worst case scenario: Chord can become
inconsistent and stay that way
The Fin
e Prin
t
The scenario
you h
ave b
een shown is
of l
ow pro
bability
. In
all
likelih
ood, Chord
would
repair
itself
after a
ny part
itionin
g failu
re th
at mig
ht really
aris
e. Cave
at em
ptor a
nd all
that.
Leiden; Dec 06 Gossip-Based Networking
Workshop 20
Saved by gossip!
Epidemic gossip: remedy for what ails Chord! c.f. Epichord (Liskov), Bambou
Key insight: Gossip based DHTs, if correctly
designed, are self-stabilizing!
Leiden; Dec 06 Gossip-Based Networking
Workshop 21
Connection to self-stabilization
Self-stabilization theory Describe a system and a desired
property Assume a failure in which code remains
correct but node states are corrupted Proof obligation: show that property is
reestablished within bounded time But doesn’t bound badness when
transient disruption is occuring
Leiden; Dec 06 Gossip-Based Networking
Workshop 22
Beyond self-stabilization Tardos poses a related problem
Consider behavior of the system while an endless sequence of disruptive events occurs
System never reaches a quiescent state Under what conditions will it still behave correctly?
Results of form “if disruptions satisfy then correctness property is continuously satisfied”
Hypothesis: with convergent consistency we may be able to develop a proof framework for systems that are continuously safe.
Leiden; Dec 06 Gossip-Based Networking
Workshop 23
Let’s look at a second example
Astrolabe system uses a different emergent data structure – a tree
Nodes are given an initial location – each knows its “leaf domain”
Inner nodes are elected using gossip and aggregation
Astrolabe
Astrolabe Intended as help
for applications adrift in a sea of information
Structure emerges from a randomized gossip protocol
This approach is robust and scalable even under stress that cripples traditional systems
Developed at RNS, Cornell
By Robbert van Renesse, with many others helping…
Today used extensively within Amazon.com
Leiden; Dec 06 Gossip-Based Networking
Workshop 25
Astrolabe is a flexible monitoring overlay
Name Time Load Weblogic?
SMTP? Word Version
swift 2003 .67 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Versi
on
swift 2011 2.0 0 1 6.2
falcon 1971 1.5 1 0 4.1
cardinal 2004 4.5 1 0 6.0
swift.cs.cornell.edu
cardinal.cs.cornell.edu
Periodically, pull data from monitored systems
Name Time Load Weblogic?
SMTP? Word Versi
on
swift 2271 1.8 0 1 6.2
falcon 1971 1.5 1 0 4.1
cardinal 2004 4.5 1 0 6.0
Name Time Load Weblogic?
SMTP? Word Version
swift 2003 .67 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2231 1.7 1 1 6.0
Leiden; Dec 06 Gossip-Based Networking
Workshop 26
Astrolabe in a single domain Each node owns a single tuple, like the
management information base (MIB) Nodes discover one-another through a
simple broadcast scheme (“anyone out there?”) and gossip about membership Nodes also keep replicas of one-another’s
rows Periodically (uniformly at random) merge
your state with some else…
Leiden; Dec 06 Gossip-Based Networking
Workshop 27
State Merge: Core of Astrolabe epidemic
Name Time Load Weblogic?
SMTP? Word Version
swift 2003 .67 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Versi
on
swift 2011 2.0 0 1 6.2
falcon 1971 1.5 1 0 4.1
cardinal 2004 4.5 1 0 6.0
swift.cs.cornell.edu
cardinal.cs.cornell.edu
Leiden; Dec 06 Gossip-Based Networking
Workshop 28
State Merge: Core of Astrolabe epidemic
Name Time Load Weblogic?
SMTP? Word Version
swift 2003 .67 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Versi
on
swift 2011 2.0 0 1 6.2
falcon 1971 1.5 1 0 4.1
cardinal 2004 4.5 1 0 6.0
swift.cs.cornell.edu
cardinal.cs.cornell.edu
swift 2011 2.0
cardinal 2201 3.5
Leiden; Dec 06 Gossip-Based Networking
Workshop 29
State Merge: Core of Astrolabe epidemic
Name Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Versi
on
swift 2011 2.0 0 1 6.2
falcon 1971 1.5 1 0 4.1
cardinal 2201 3.5 1 0 6.0
swift.cs.cornell.edu
cardinal.cs.cornell.edu
Leiden; Dec 06 Gossip-Based Networking
Workshop 30
Observations Merge protocol has constant cost
One message sent, received (on avg) per unit time.
The data changes slowly, so no need to run it quickly – we usually run it every five seconds or so
Information spreads in O(log N) time But this assumes bounded region size
In Astrolabe, we limit them to 50-100 rows
Leiden; Dec 06 Gossip-Based Networking
Workshop 31
Big systems…
A big system could have many regions Looks like a pile of spreadsheets A node only replicates data from its
neighbors within its own region
Leiden; Dec 06 Gossip-Based Networking
Workshop 32
Scaling up… and up…
With a stack of domains, we don’t want every system to “see” every domain Cost would be huge
So instead, we’ll see a summaryName Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
cardinal.cs.cornell.edu
Name Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Name Time Load Weblogic?
SMTP? Word Version
swift 2011 2.0 0 1 6.2
falcon 1976 2.7 1 0 4.1
cardinal 2201 3.5 1 1 6.0
Leiden; Dec 06 Gossip-Based Networking
Workshop 33
Name Load Weblogic? SMTP? Word Version
…
swift 2.0 0 1 6.2
falcon 1.5 1 0 4.1
cardinal 4.5 1 0 6.0
Name Load Weblogic? SMTP? Word Version
…
gazelle 1.7 0 0 4.5
zebra 3.2 0 1 6.2
gnu .5 1 0 6.2
Name Avg Load
WL contact SMTP contact
SF 2.6 123.45.61.3
123.45.61.17
NJ 1.8 127.16.77.6
127.16.77.11
Paris 3.1 14.66.71.8 14.66.71.12
Astrolabe builds a hierarchy using a P2P protocol that “assembles the puzzle” without any servers
Name Load Weblogic? SMTP? Word Version
…
swift 2.0 0 1 6.2
falcon 1.5 1 0 4.1
cardinal 4.5 1 0 6.0
Name Load Weblogic? SMTP? Word Version
…
gazelle 1.7 0 0 4.5
zebra 3.2 0 1 6.2
gnu .5 1 0 6.2
Name Avg Load
WL contact SMTP contact
SF 2.6 123.45.61.3
123.45.61.17
NJ 1.8 127.16.77.6
127.16.77.11
Paris 3.1 14.66.71.8 14.66.71.12
San Francisco
New Jersey
SQL query “summarizes”
data
Dynamically changing query output is visible system-wide
Name Load Weblogic? SMTP? Word Version
…
swift 1.7 0 1 6.2
falcon 2.1 1 0 4.1
cardinal 3.9 1 0 6.0
Name Load Weblogic? SMTP? Word Version
…
gazelle 4.1 0 0 4.5
zebra 0.9 0 1 6.2
gnu 2.2 1 0 6.2
Name Avg Load
WL contact SMTP contact
SF 2.2 123.45.61.3
123.45.61.17
NJ 1.6 127.16.77.6
127.16.77.11
Paris 2.7 14.66.71.8 14.66.71.12
Leiden; Dec 06 Gossip-Based Networking
Workshop 34
Large scale: “fake” regions These are
Computed by queries that summarize a whole region as a single row
Gossiped in a read-only manner within a leaf region
But who runs the gossip? Each region elects “k” members to run
gossip at the next level up. Can play with selection criteria and “k”
Leiden; Dec 06 Gossip-Based Networking
Workshop 35
Hierarchy is virtual… data is replicated
Name Load Weblogic? SMTP? Word Version
…
swift 2.0 0 1 6.2
falcon 1.5 1 0 4.1
cardinal 4.5 1 0 6.0
Name Load Weblogic? SMTP? Word Version
…
gazelle 1.7 0 0 4.5
zebra 3.2 0 1 6.2
gnu .5 1 0 6.2
Name Avg Load
WL contact SMTP contact
SF 2.6 123.45.61.3
123.45.61.17
NJ 1.8 127.16.77.6
127.16.77.11
Paris 3.1 14.66.71.8 14.66.71.12
San Francisco
New Jersey
Yellow leaf node “sees” its neighbors and the domains on the path to the
root.
Falcon runs level 2 epidemic because it has
lowest load
Gnu runs level 2 epidemic because it has lowest load
Leiden; Dec 06 Gossip-Based Networking
Workshop 36
Hierarchy is virtual… data is replicated
Name Load Weblogic? SMTP? Word Version
…
swift 2.0 0 1 6.2
falcon 1.5 1 0 4.1
cardinal 4.5 1 0 6.0
Name Load Weblogic? SMTP? Word Version
…
gazelle 1.7 0 0 4.5
zebra 3.2 0 1 6.2
gnu .5 1 0 6.2
Name Avg Load
WL contact SMTP contact
SF 2.6 123.45.61.3
123.45.61.17
NJ 1.8 127.16.77.6
127.16.77.11
Paris 3.1 14.66.71.8 14.66.71.12
San Francisco
New Jersey
Green node sees different leaf domain but has a consistent view of the inner
domain
Leiden; Dec 06 Gossip-Based Networking
Workshop 37
Worst case load?
A small number of nodes end up participating in O(logfanoutN) epidemics Here the fanout is something like 50 In each epidemic, a message is sent and
received roughly every 5 seconds We limit message size so even during
periods of turbulence, no message can become huge.
Leiden; Dec 06 Gossip-Based Networking
Workshop 38
Self-stabilization?
Like Kelips, it seems that Astrolabe Is convergently consistent, self-
stabilizing And would “ride out” a large class of
possible failures But Astrolabe would be disrupted by
Incorrect aggregation (Byzantine faults) Correlated failure of all representatives
of some portion of the tree
Leiden; Dec 06 Gossip-Based Networking
Workshop 39
Focus on emergent shape
Kelips: Nodes start with a-priori assignment to affinity groups, end up with a superimposed pointer structure
Astrolabe: Nodes start with a-priori leaf domain assignments, build the tree
What other data structures can be constructed with emergent protocols?
Leiden; Dec 06 Gossip-Based Networking
Workshop 40
Emergent shape We know a lot about a related question
Given a connected graph, cost function Nodes have bounded degree Use a gossip protocol to swap links until
some desired graph emerges Another related question
Given a gossip overlay, improve it by selecting “better” links (usually, lower RTT)
Example: The “Anthill” framework of Alberto Montresor, Ozalp Babaoglu, Hein Meling and Francesco Russo
Leiden; Dec 06 Gossip-Based Networking
Workshop 41
Example of an open problem
Given a description of a data structure (for example, a balanced tree) … design a gossip protocol such that the
system will rapidly converge towards that structure even if disrupted
Do it with bounced per-node message rates, sizes (network load less important)
Use aggregation to test tree quality?
Leiden; Dec 06 Gossip-Based Networking
Workshop 42
Van Renesse’s dreadful aggregation tree
A B C D E F G H I J K L M N O P
A C E G I K M O
B F J N
D L
An event e occurs at H
P learns O(N) time units
later!
G gossips with H and
learns e
Leiden; Dec 06 Gossip-Based Networking
Workshop 43
What went wrong?
In Robbert’s horrendous tree, each node has equal “work to do” but the information-space diameter is larger!
Astrolabe benefits from “instant” knowledge because the epidemic at each level is run by someone elected from the level below
Leiden; Dec 06 Gossip-Based Networking
Workshop 44
Insight: Two kinds of shape
We’ve focused on the aggregation tree
But in fact should also think about the information flow tree
Leiden; Dec 06 Gossip-Based Networking
Workshop 45
Information space perspective
Bad aggregation graph: diameter O(n)
Astrolabe version: diameterO(log(n))
H – G – E – F – B – A – C – D – L – K – I – J – N – M – O – P
A B C D E F G H I J K L M N O P
A C E G I K M O
A E I M
A I
A – B
C –
DE – F
G –
H
I –
J
K –
L
M –
N
O –
P
A B C D E F G H I J K L M N O P
A C E G I K M O
B F J N
D L
Leiden; Dec 06 Gossip-Based Networking
Workshop 46
Gossip and bias
Often useful to “bias” gossip, particularly if some links are fast and others are very slow
AC
D
FZ
E
YB
XRoughly half
the gossip will cross this link!
Demers: Shows how to adjust probabilities to even the load. Ziao later showed that must also fine-tune gossip
rate
Leiden; Dec 06 Gossip-Based Networking
Workshop 47
Gossip and bias
Idea: “shaped” gossip probabilities Gravitational Gossip (Jenkins)
Groups carry multicast event streams reporting evolution of a continuous function (like electric power loads)
Some nodes want to watch “closely” while others only need part of the data
Leiden; Dec 06 Gossip-Based Networking
Workshop 48
Gravitational Gossip
a c
b
Inner ring: Nodes want 100% of the traffic
Middle ring: Nodes want 75% of the traffic
Outer ring: Nodes want 20% of the traffic
Jenkins: When a gossips to b, includes information about
topic t in a way weighted by b’s level of interest in topic t
Leiden; Dec 06 Gossip-Based Networking
Workshop 49
Gravitational Gossip
Leiden; Dec 06 Gossip-Based Networking
Workshop 50
How does bias impact information-flow graph?
Earlier, all links were the same Now, some links carry
Less information And may have longer delays (Bounded capacity messages: “similar”
to long links?) Question: Model biased information
flow graphs and explore implications
Leiden; Dec 06 Gossip-Based Networking
Workshop 51
Questions about bias
When does the biasing of gossip target selection break analytic results? Example: Alves and Hopcroft show that
with fanout too small, gossip epidemics can die out, logically partitioning a system
Question: Can we relate the question to flooding on an expander graph?
Leiden; Dec 06 Gossip-Based Networking
Workshop 52
Can an overlay learn its shape?
The unifying link across the many examples we’ve cited
And needed in real world overlays when these adapt to underlying topology… For example, to pick “super-peers” Hints of a theory of adaptive tuning?
Related to “sketches” in databases
Leiden; Dec 06 Gossip-Based Networking
Workshop 53
Suppose we wanted to use gossip to implement decentralized balanced overlay trees (nodes “know” locations) Build some sort of overlay Randomly select initial upper-level nodes Jiggle until they are nicely spaced
Instance of “facility location” problem
Can an overlay learn its shape?
Leiden; Dec 06 Gossip-Based Networking
Workshop 54
Emergent landmarks
Leiden; Dec 06 Gossip-Based Networking
Workshop 55
Emergent landmarks
Too many
Too few
Leiden; Dec 06 Gossip-Based Networking
Workshop 56
Emergent landmarks
Leiden; Dec 06 Gossip-Based Networking
Workshop 57
Emergent landmarks
Leiden; Dec 06 Gossip-Based Networking
Workshop 58
Emergent landmarks
One per cluster
Centroid
Leiden; Dec 06 Gossip-Based Networking
Workshop 59
2-D scenario: Gossip variant of Euclidian k-medians (a popular STOC topic)
Nodes know position on a line, can gossip In log(N) time seek k equally spaced nodes
Later, after disruption, reconverge in log(N) time
NB: The facility placement problem is just the d-dimensional k-centroids problem
K=3
Leiden; Dec 06 Gossip-Based Networking
Workshop 60
Why is this hard?
After k rounds, each node has only communicated with 2k other nodes
In log(N) time, aggregate information has time to transit the whole graph, but with bounded bandwidth we don’t have time to send all data
Leiden; Dec 06 Gossip-Based Networking
Workshop 61
Emergent shape problem
Can a gossip system learn its own shape in a convergently consistent way? Start with centralized solutions draw
from the optimization community. Then seek decentralized gossip variants
that achieve convergently consistent approximations, (re)converge in log(N) rounds, and use bounded bandwidth
Leiden; Dec 06 Gossip-Based Networking
Workshop 62
… more questions
Notice that Astrolabe forces participants to agree on what the aggregation hierarchy should contain In effect, we need to “share interest” in
the aggregation hierarchy This allows us to bound the size of
messages (expected constant) and the rate (expected constant per epidemic)
Optional content (time permitting)
Leiden; Dec 06 Gossip-Based Networking
Workshop 63
The question
Could we design a gossip-based system for “self-centered” state monitoring?
Each node poses a query, Astrolabe style, on the state of the system We dynamically construct an overlay for
each of these queries The “system” is the union of these
overlays
Optional content (time permitting)
Leiden; Dec 06 Gossip-Based Networking
Workshop 64
Self-centered monitoring
Optional content (time permitting)
Leiden; Dec 06 Gossip-Based Networking
Workshop 65
Self-centered queries… Offhand, looks like a bad idea
If everyone has an independent query And everyone is iid in all the obvious ways Than everyone must invest work
proportional to the number of nodes monitored by each query
If O(n) queries touch O(n) nodes, workload grows as O(n) and global load as O(n2)
Optional content (time permitting)
Leiden; Dec 06 Gossip-Based Networking
Workshop 66
Aggregation … but in practice, it seems unlikely that
queries would look this way More plausible is something Zipf-like
A few queries look at broad state of system Most look at relatively few nodes And a small set of aggregates might be shared
by the majority of queries Assuming this is so, can one build a
scalable gossip overlay / monitoring infrastructure?
Optional content (time permitting)
Leiden; Dec 06 Gossip-Based Networking
Workshop 67
Summary of possible topics… Consistency models Necessary conditions
for convergentconsistency
Self-stabilization Implications of bias Emergent structures
Self-centered aggregation
Bandwidth-limited systems; “sketches”
Using aggregation to fine-tune a shape with constant costs