randomized distributed decision david peleg joint work with: pierre fraigniaud amos korman merav...

73
Distributed Decision David Peleg Joint work with : Pierre Fraigniaud Amos Korman Merav Parter

Upload: jarrett-ellers

Post on 31-Mar-2015

217 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Randomized Distributed Decision

David Peleg

Joint work with:Pierre Fraigniaud

Amos KormanMerav Parter

Page 2: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

The goal: Solve problems collaboratively by multiple distributed processors

Distributed computing

• Shared memory (multi-cores)• Message passing (network)

Communication mode:

Page 3: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Point-to-point communication network

The distributed network model

V={v1,…,vn} - Processors (network sites)E - bidirectional communication links

Page 4: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Why is it interesting?

Theoretical viewpoint:Several inherent differences between the distributed and the traditional centralized-sequential computational models

Practical viewpoint:Extremely wide applicability on all levels

Page 5: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

In centralized-sequential setting: Processor knows everything (inputs, intermediate results, etc.)

In distributed setting:Processors have only a partial picture

Incomplete knowledge

Page 6: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

• Do not know the network topology• Know only their local portion of the input• Do not know who else participates• Do not know current stage of others

Incomplete knowledge

X1=3

v1

V?

X=??

In particular, processors:

Page 7: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

GIntermediate topology knowledge model

The KTr model:Each processor u sees Br(u), itsr-neighborhood

KT2

(seeing 2-neighborhoods)

(hypothetical model…) B2(u)

u

Page 8: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

We assume a synchronous model(the entire system is driven by global clock)

Timing and synchrony

Processors’ machine cycle: 1. Send messages to some neighbors2. Receive messages from neighbors3. Perform internal computation

Page 9: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

The LOCAL Model

Focus on impact of locality / distances

Message size and internal computationare unbounded

Page 10: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

The LOCAL model

Complexity measure:Time (# of rounds)

G

In r rounds,each processor u can collect complete information on Br(u), its r-neighborhood⇒ r=3 B3(u)

Page 11: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Equivalent viewpoint

Instead of considering r-round algorithms –

G

consider 0-round (no-communication) algorithmsin the KTr model

Page 12: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Local distributed computation

Input:• Graph G(V,E)• Local views of

r-neighborhood

Goal:Compute a global solution for a given problem

G

Page 13: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

*i ,G(V,E)G 10 , | , xxL

Language: (Decidable) collection of pairs

Local distributed computation tasks as languages

X1=3v1

X2=6v2

X3=8v3

X4=2v4

X5=3v5

X6=3v6

X7=3v7

Page 14: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

)()( , | , vuE(G)(u,v)Goloring xxx C

oloringColoringC

Examples

x(v) = color of v

Page 15: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Examples

1)( ,1,0)( | ,Gu

uuG xxxAMOS

AMOS AMOS

At Most One Selected

Page 16: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Examples

1)(,1,0)( | ,Gu

uuG xxxLE

LE

Leader Election

Page 17: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Examples

)()(),(

),( | ),(,

0

0

uvGVv

GVuGonsensus

inout

outin

xx

xxC

Xin=0Xout=1

Xin=1Xout=1

Xin=0Xout=1

Xin=0Xout=1

onsensusC

Page 18: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Examples

.a MIS is 1)( | )( | , vGVvSGMIS xx

MIS

Page 19: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Local computational tasks

Local Construction (LC) Tasks:

Given a global problem π on a graph G,

construct a solution x locally

Local Decision (LD) Tasks:

Given a global problem π on a graph G

and a proposed solution x,

decide (or verify) locally that x solves π on G

Page 20: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Distributed Complexity Theory

Proof Labeling Schemes [Korman, Fraigniaud , P, 05]

Locally Checkable Proofs [Goos, Suomela, 11]

Decidability Classes for Mobile Agent Computing [Fraigniaud, Pelc, 2012]

Locality & Checkability in Wait-free Computing [Fraigniaud, Rajsbaum, Travers, 11]

Local Distributed Decision [Fraigniaud, Korman, P, 11]

On the Impact of Identifiers on Local Decision [Fraigniaud, Halldórsson, Korman, 12] [Fraigniaud, Goos, Korman, Suomela, 13]

Page 21: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Decision rules: Nodes need to collectively decide whether the given instance belongs to the language.

Local decision tasks[Fraigniaud, Korman, P, 11]

Page 22: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

YesNo

Local decision tasks

Decision rules:

• In a legal (“YES”) instance: all nodes output YES

• In an illegal (“NO”) instance: some nodes output NO

[Fraigniaud, Korman, P, 11]

YesYes

Yes

Yes

Page 23: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Local decision tasks[Fraigniaud, Korman, P, 11]

The asymmetry between the two cases looks odd, but…

Decision rules:

• In a legal (“YES”) instance: all nodes output YES

• In an illegal (“NO”) instance: some nodes output NO

Page 24: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Local decision tasks[Fraigniaud, Korman, P, 11]

Note: If every node is required to know the correct answer on both legal and illegal instances, then no local solutions are possible!

Decision rules:

• In a legal (“YES”) instance: all nodes output YES

• In an illegal (“NO”) instance: some nodes output NO

Page 25: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Local decision tasks

Example: Consider the following “very local” task:Input: Binary vector xQuestion: Is v1‘s input x(v1)=1 ?

X1=1v1

X2=0v2

X3=0v3

X4=1v4

X5=0v5

X6=1v6

X7=1v7

Page 26: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Local decision tasks

Algorithm: • All nodes other than v1 say YES.• Node v1 answers according to x(v1).

X1=1v1

X2=0v2

X3=0v3

X4=1v4

X5=0v5

X6=1v6

X7=1v7

Requiring all nodes to know the answer means no local solution is possible…

Page 27: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Decision Tasks

Fault tolerance

Checking correctness of construction algorithm

Platform for distributed complexity theory

Applications:

Page 28: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

A t-round distributed algorithm is a LOCAL Decider for if:

: Every processor says “yes”

: At least one says “no”

Class of languages that have a t-round local decider.

LD(t) (Local Decision)

Local analogue for the class P

LOCAL Decider

Page 29: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Local Computation Tasks

The class of all languages is divided into 4 classes:

1. Hard to construct and hard to decide.

2. Easy to construct and easy to decide.

3. Hard to construct but easy to decide.

4. Easy to construct but hard to decide.

Page 30: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Hard to construct but easy to decide (locally)

)()( , | , vuE(G)(u,v)Goloring xxx C

Deterministic construction of ( in rounds

[Panconesi, Srinivasan, 96]

Page 31: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Hard to construct but easy to decide (locally)

)()( , | , vuE(G)(u,v)Goloring xxx C

Deterministic decision in a single round.

No

Page 32: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Easy to construct but hard to decide (locally)

1)( | ,Gu

uG xxAMOS

0-round construction: Each node marks itself non-selected ( x(u)← 0 )

At Most One Selected

Page 33: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Easy to construct but hard to decide (locally)

1)( | ,Gu

uG xxAMOS

Observation: AMOS is not locally decidable in o(n) rounds.

At Most One Selected

Page 34: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Hard to construct and hard to decide (locally)

Leader Election

1)( | ,Gu

uG xxLE

Construction barrier due to symmetry breaking:O(diameter) rounds are required.

Observation:LE is not locally decidable in o(n) rounds.

Page 35: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Easy to construct and easy to decide (locally)

(v)(u)s.t

neighbor v hasu every | ,_

xx

xGColoringeakW

Theorem [Naor, Stockmeyer, STOC ’93]: Weak 2-Coloring in fixed odd degree graphs can be constructed in O(1) rounds.

Local decision: in single round No

Local construction:

Page 36: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Thm [Naor, 96] : Randomization does not help for 3-coloring a ring:Randomized lower bound = deterministic upper bound = Θ(log*n) rounds.

Thm [Naor, Stockmeyer, 93] : For constant degree graphs, certain randomized labeling algorithms can be de-randomized.

Does Randomization help in local construction?

Low-Degree Graphs

So randomization fails to help?

Page 37: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Does Randomization help in local construction?

Large-Degree Graphs( can be computed

Thm [Alon, Babai, Itai, 86], [Luby, 86] : randomly in O(logn ) w.h.p.

Thm [Panconesi, Srinivasan , 96] : deterministically in time .

So randomization does help?

Page 38: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Yes, No

9 8

3

7

4

5

6

12

10

u

12

13

1415

16

17

18

19

20

9

9

99

9

9

23

Randomized local decision

Page 39: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Does Randomization help in local decision?

1)( | ,Gu

uG xxAMOS

Recall: AMOS is not locally decidable in o(n) rounds.

Page 40: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Does Randomization help in local decision?

1)( | ,Gu

uG xxAMOS

Use randomness to decide!

Recall: AMOS is not locally decidable in o(n) rounds.

Page 41: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Randomized LOCAL Decider

A (p,q)-decider for language L is a local, 2-sided error Monte Carlo algorithm, such that:

: With probability* ≥p, every node returns “yes”

: With probability* ≥q, at least one node returns “no”

* The probabilities are taken over all coin tosses performed by the nodes

Page 42: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Randomized LOCAL Decider

Class of languages that have a t-round (p,q)-decider

BPLD(p,q,t) (Bounded Probability

Local Decision) Local analogue for BPP

Page 43: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Does randomization help in local decision?

[Fraigniaud, Korman, P, 11]

Partial answer:Yes, for: - some families of languages, - some values of p and q

Thm: p2+q=1 is a sharp threshold for hereditary languages*

* Languages that are closed under inclusion.

Page 44: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

p (“yes” probability)

q (“

no”

prob

abili

ty)

Yes

Randomization threshold No

p2+q=1 is a sharp threshold for hereditary languages

p 2+q=1

Does randomization help in local decision?

[Fraigniaud, Korman, P, 11]

Page 45: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Distributed Complexity Classes

NoLD

= class of languages decidable by t-round deterministic algorithm

LD(t)

If problem is here (randomly decidable

with high p and q)

t is in LD (can be decided

deterministically)

Page 46: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Distributed Complexity Classes

NoLD

BPLD

= class of languages with t-round (p,q)-decider s.t. p2+q 1BPLD(t)

∃ problems here (randomly decidable

with low p and q)

t do not belong to LD

= class of languages decidable by t-round deterministic algorithm

LD(t)

Page 47: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

p2+q ≤ 1: randomization helps

1)( | ,Gu

uG xxAMOS

Recall: is not in LD AMOS

Page 48: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

p2+q ≤ 1: randomization helps

0-round (p,q)-decider (code for node u) If unselected, return “yes” with probability 1 If selected, return “yes” with probability p

“Yes” w/prob

YesYesYesYes YesYes

1)( | ,Gu

uG xxAMOS

Page 49: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Yes Yes

“Yes” w/prob

Prob(every node returns “yes”) ≥ p

Legal (“YES”) instance

Yes Yes Yes

Correctness of the Decider

YesYes

Page 50: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Yes Yes

“Yes” w/prob

Prob(at least one node returns “no”) ≥ 1-p2 ≥ q

Illegal (“NO”) instance

Yes Yes Yes

“Yes” w/prob

Correctness of the Decider

Page 51: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Is my t-ball legal?

p2+q > 1: de-randomization possible

G

9 8

3

7

4

5

6

12

10u

12

13

14

1415

16

17

18

19

19

20

9

9

9 9

99

9

9

{Yes, No}

(p,q)-deciderapplied by node u

t

Page 52: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

p2+q > 1: de-randomization possible

9 8

3

7

4

5

6

12

G

10u

12

13

14

1415

16

17

18

19

19

20

9

9

9 9

99

9

9

The required radius t’ depends on how far p2+q – 1 is bounded away from zero.

Is my t’-ball legal?

Deterministic decider applied by node u

{Yes, No} t’

Page 53: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Instance (G,x)

t-round (p,q)-decider

Sketch (on path)

Page 54: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Prob(at least one node returns “no”) < δ2t

-Secure Zone

-Secure zone: a “safe-separator” (of width 2t) with low rejection probability

A -Secure zone divides the instance into two independent sub-instances carrying the mass of rejection probability

Page 55: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

-Secure Zones

tp

R

)1log(

log11sec

t=Run time of the(p,q)-decider

Claim: Every legal sub-path of length contains a secure zone.

Define

𝑹𝒔𝒆𝒄

Page 56: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

By contradiction: consider a legal sub-path P of length ,suppose P does not contain a secure zone

Proof:

-Secure Zones

For every sub-path of length 2t in P, Prob(some node returns “no”) ≥ δ

Prob(all nodes return “yes”) < 1-δ

or

Page 57: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

-Secure Zones

… …2t 2t

P has K= ``independent” sub-paths of length 2t.

Page 58: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

In each of these K sub-paths, Prob(all nodes return “yes”

< )K

𝐾 <log𝑝

log(1−𝛿)Contradiction for proper selection of constants.

-Secure Zones

… …2t 2t

Page 59: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

-Secure Zones

Claim: Every legal sub-path of length contains a secure zone.

Page 60: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

De-randomization of (p,q)-decider for p2+q>1

Given: Hereditary language L. with a t-local (p,q)-decider

A t-local (deterministic) decider for L. :

t’=Rsec (t)

Every node u inspects its radius t’ neighborhood B(u)

If B(u) ∈ L, then u outputs ”yes”, else it outputs “no”.

Page 61: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

De-randomization of (p,q)-decider for p2+q>1

Every node u inspects its radius t’neighborhood B(u)

If B(u) ∈ L, then u outputs ”yes”, else it outputs “no”.

Simulation correctness proof:

Legal instance I ∈ L:As L is hereditary, all neighborhoods B(u) are legal

Page 62: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

De-randomization of (p,q)-decider for p2+q>1

IlLegal instance I ∉ L:

Need to show that at least one ball B(u) is illegal.

Towards contradiction assume all balls are legal.

LI

Maximal legal sub-pathL'I

L)(uBu

sec2R

Page 63: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

L'IL)(uB

secR

L''I

Hence, contradiction to the fact that is the maximal legal sub-path in .

'II

De-randomization of (p,q)-decider for p2+q>1

Claim:

Page 64: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

The Gluing Lemma

The union of two legal instances is legal provided their overlap is sufficiently large

L1I

L2I

L3I

The required overlap size depends on the value p2+q-1

tp

R

)1log(

log11sec

Page 65: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

The Gluing Lemma

The required overlap size depends on the value p2+q-1

Fsatisfying < p2+q-1

Page 66: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

2t

Since the (p,q) decider is a t-round algorithm, L and R are independent!

Proof of the Gluing Lemma

Event L: every node on the left returns “yes”

Event R: every node on the right returns “yes”

The overlap section contains a secure sub-path

3I

Page 67: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

2t

q ≤ Prob(at least one node returns “no”) ≤ +

Proof of the Gluing Lemma

Event L: “yes” Event R: “yes”

Contradiction to the definition of

Assume towards contradiction that

3I

The overlap section contains a secure sub-path

Page 68: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Zooming into the Randomization Region

DeterminismRandomization

p (“yes” probability)

q (“

no”

prob

abili

ty)

[Fraigniaud, Korman, Parter, P, 12]

Page 69: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

= class of languages that have a (p,q)-decider s.t

for integer k

The Bk hierarchy

Bk(t)

Bk

p1+1/k + q 1

Page 70: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Theorem: The Bk hierarchy is strict

BPLD

B2

B

ALL

B3

Determinism (B1)

p (“yes” success probability)

B1(t) ALLq

(“no

” su

cces

s pr

obab

ility

)

p 2+q>1p 3/2+q>1p 4/3+q>1

p+q>1

Determinism

Page 71: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

The At most k selected Language

B2

B

ALL

Bk+1

Determinism q

p

At most Kselected

At most 1 selected

kuGGu

)( | , xxselected- k mostAt

Lemma:

kk BB \1 selected k mostAt

Integer k

Page 72: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Towards Distributed Computational Complexity Theory

Are there intermediate classes between Bk(t) and Bk+1(t)?

Hardness/ completeness: Notions of reductions and complete problems for locality classes

Randomization and non-determinism: Interplay between certificate size and success guarantees.

The role of identifiers [Fraigniaud, Goos, Korman, Suomela, 13]

Complexity theory for the CONGEST model Other combining rules for local decision (instead

of “logical and”)

Page 73: Randomized Distributed Decision David Peleg Joint work with: Pierre Fraigniaud Amos Korman Merav Parter

Randomization

Thank you for your attention