Download - Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

Transcript
Page 1: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

1

Distributed Verification and Hardness of Distributed Approximation

Atish Das Sarma Stephan Holzer

Danupon Nanongkai

Gopal Pandurangan David Peleg

WeizmannGoogle ResearchLiah Kor

Roger Wattenhofer

ETH Zurich

U. of Vienna & Georgia Tech

Nanyang Technological University& Brown University

ETH ZurichWeizmann

Amos KormanU. Paris 7

Page 2: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

2

PLANResult summary

Techniques Overview

From communication complexity to distributed algo. lower bound

Page 3: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

3

Distributed network

Page 4: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

4

Distributed network A graph G of n nodes, diameter D

n= 4, D=2

1

2

3

4

Page 5: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

5

Main issue: LOCALITY and BANDWIDTH

?

1

2

3

4

42

3

Page 6: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

6

Time complexity = number of rounds

1

2

3

4

log n

log n

log nlog n

log nlog n

log n

log nlog n

Page 7: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

7

Example: Spanning tree in O(D) time

1

2

3

4

Page 8: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

8

Weighted distributed network

?

10

2

1

9

59

5

1

2

3

4

42

3

Page 9: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

9

Fundamental problems

• Spanning Tree – Broadcasting, Aggregation, etc• Minimum Spanning Tree – Efficient

broadcasting, leader election, etc. • Shortest path – Routing, etc.• Steiner tree – Multicasting, etc. • Many other graph problems.

Page 10: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

10

How fast can we compute distributively?

Page 11: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

11

Three points of this work

1. A new generic technique to prove lower bounds of distributed algorithms. – Works for approximation algorithms. – Connection to communication complexity

2. New bounds for many problems. Tight in some cases.

3. A systematic study of distributed verification.

Page 12: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

12

Distributed algorithms for the above problems require

W(n1/2+D) time

Page 13: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

13

Two main ingredients

1. Verification Approximation2. Connection to communication complexity.

Page 14: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

14

ShowcaseMinimum Spanning Tree

Page 15: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

15

Time of Distributed AlgorithmsProblems Upper bound Lower bound

Spanning tree (ST)

O(D) W(D)

MST O(D + n1/2) W(D + n1/2)

a-approx. MST W(D + (n /a)1/2)

MST Verification O(D + n1/2) W(D + n1/2)

[trivial] [trivial]

[Garay, Kutten, Peleg FOCS’93] [Peleg, Rubinovich FOCS’99]

[Elkin STOC’04]

[Kor, Korman, Peleg STACS’11][Kor, Korman, Peleg STACS’11]

Page 16: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

16

Time of Distributed AlgorithmsProblems Upper bound Lower bound

Spanning tree (ST)

O(D) W(D)

MST O(D + n1/2) W(D + n1/2)

a-approx. MST W(D + (n /a)1/2)

MST Verification O(D + n1/2) W(D + n1/2)

ST Verification O(D + n1/2)

[trivial] [trivial]

[Garay, Kutten, Peleg FOCS’93] [Peleg, Rubinovich FOCS’99]

[Elkin STOC’04]

[Kor, Korman, Peleg STACS’11][Kor, Korman, Peleg STACS’11]

Page 17: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

17

Implication of our results

Page 18: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

18

Time of Distributed AlgorithmsProblems Upper bound Lower bound

Spanning tree (ST)

O(D) W(D)

MST O(D + n1/2) W(D + n1/2)

a-approx. MST W(D + (n /a)1/2)

MST Verification O(D + n1/2) W(D + n1/2)

ST Verification O(D + n1/2)

[trivial] [trivial]

[Garay, Kutten, Peleg FOCS’93] [Peleg, Rubinovich FOCS’99]

[Elkin STOC’04]

[Kor, Korman, Peleg STACS’11][Kor, Korman, Peleg STACS’11]

W(D + n 1/2)

Page 19: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

19

Previous lower bound proofs• Deterministic : Count the number of states.

Argue that the number is not enough. • Randomized: Come up with a good input

distributions.

Our proof• Simple reduction from communication

complexity.• Avoid complication in proving randomized lower

bounds.

Page 20: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

20

PLANResult summary

Techniques Overview

From communication complexity to distributed algo. lower bound

Page 21: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

21

Approx MST lower bound W(n1/2)

Distributed equality verificationlower bound W(n1/2)

ST verification lower bound W(n1/2)

Distributed equality verificationlower bound W(n1/2)

Direct equality verificationlower bound W(n1/2)

Well-known result in communication complexity

Similar to hardness of TSP

Similar to lower bounds of graph streaming algorithms

Three steps of reductionDistributed AlgorithmsCommunication Complexity

simulationtheorem

Page 22: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

22

PLANResult summary

Techniques Overview

From communication complexity to distributed algo. lower bound

Page 23: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

23

Communication complexity of EQUALITY

Page 24: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

24

How many bits do they have to exchange?

Alice Bobx {0, 1}100 y {0, 1}100

x=y?

Yes, x=yYes, x=y

Page 25: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

25

One solution: Alice sends everything ... time=100

Alice Bobx {0, 1}100 y {0, 1}100

x=y?

Page 26: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

26

Theorem: Any algorithm needs ≥100 bits

Alice Bobx {0, 1}100 y {0, 1}100

x=y?

Page 27: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

27

Distributed time complexity of EQUALITY

Page 28: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

28

Alicex {0, 1}100

Boby {0, 1}100

100 green nodes

Alice and Bob are connected by many paths of length 100

Page 29: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

29

Alicex {0, 1}100

Boby {0, 1}100

100 green nodes

In each step, one edge can carry one bit on each direction

Page 30: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

30

How many steps do they need to check whether “x=y”?

Page 31: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

31

Alice Bob

100 green nodes

A: 100 steps because the network diameter is 100

Page 32: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

32

Let’s make the diameter smaller

Page 33: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

33

Alice Bob

100 green nodes

10 green nodes 10 green nodes

Now the diameter is 30How many steps do we need?

Page 34: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

34

Claim: Need > 50 steps.

Page 35: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

35

Proof: Assume there is a distributed algorithm A that uses

≤ 50 steps

A

Page 36: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

36

Alice Bobx {0, 1}100 y {0, 1}100

A50 bits

x=y x=y

Contradiction

Page 37: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

37

Proof: Assume there is a distributed algorithm A that uses

≤ 50 steps

AGoal: Show that Alice & Bob can

use A to compute EQUALITY using 50 bits

Page 38: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

38

Alice

x {0, 1}100

Bob

y {0, 1}100A

x=y x=y

?

?

Page 39: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

39

Alice Bobx {0, 1}100 y {0, 1}100

Alice’s network Bob’s network

Run A Run AA A

Page 40: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

40

A

Alice Bobx {0, 1}100 y {0, 1}100

x y? ?

Alice’s network Bob’s network

0Step

Run A A Run A

Page 41: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

41

In step 0, Alice can run A on all machines except Bob’s

Page 42: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

42

Alice Bob

x y? ?

1Step

Page 43: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

43

Alice Bob

x y? ?

1Step

Page 44: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

44

Alice Bob

x y? ?

1Step

?

?

?

? ?

?

?

?

b1

a1

b1 = bit sent by A run on Bob’s machine

Page 45: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

45

Alice Bob

x y? ?

1Step

?

?

?

? ?

?

?

?

b1

a1

a1

b1

b1 = bit sent by A from Bob’s machine

keep this keep this

Page 46: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

46

Alice Bob

x y? ?

2Step

?

?

?

?

?

?

b2

a2

a2

b2

?

?

?

?

?

?

? ?

b2= bit sent by A from Bob’s machine

Page 47: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

47

Alice Bob

x y? ?

3Step

?

?

?

?

?

?

b3

a3

a3

b3

?

?

?

?

?

?

?

?

?

?

?

?

b3 = bit sent by A from Bob’s machine

Page 48: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

48

Alice Bob

x y? ?

4Step

?

?

?

?

?

?

b4

a4

a4

b4

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

Page 49: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

49

Alice Bob

x y? ?

5Step

?

?

?

?

?

?

b5

a5

a5

b5

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

Page 50: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

50

Alice Bob

x y? ?

50Step

?

?

?

?

?

?

b50

a50

a50

b50

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

? ?

?

?

?

?

?

?

A finishes

x=yx=y

x=yx=y

Page 51: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

51

Alice Bobx {0, 1}100 y {0, 1}100

A50 bits

x=y x=y

Contradiction

Page 52: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

52

Remarks

Page 53: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

53

1. By replacing 100 by n1/2, we can reduce distributed EQUALITY to

ST verification

x=y? Do red edges form a spanning tree?

Page 54: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

54

2. Reduce diameter ...

Page 55: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

55

Alice Bob

n1/2 p

aths

n1/2 green nodes

n1/4 orange nodes

n1/4 green nodes

Diameter = n1/4

Page 56: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

56

Alice Bob

Diameter = log nn1/

2 pat

hs

n1/2 green nodes

Page 57: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

57

3. Getting randomized lower bound

• EQAULITY does not give randomized lower bound.

• Simulation theorem holds for all functions.• Reduce from communication complexity of

HAMILTONIAN CYCLE [Spieker, Raz FOCS’93]

Page 58: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

58

Recap

1. A new generic technique to prove lower bounds of distributed algorithms. – Works for approximation algorithms.

2. New bounds for many problems. Tight in some cases.

3. A systematic study of distributed verification.

Page 59: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

59

Open problems

• Tight bounds of shortest paths, mincut, minimum routing cost spanning tree, Steiner forest, ...

• Lower bounds of algorithms on complete graphs?

• Complexity theory of distributed computing?

Page 60: Distributed Verification and Hardness of Distributed Approximation Atish Das Sarma Stephan Holzer Danupon Nanongkai Gopal Pandurangan David Peleg 1 Weizmann.

60

Thank you!

Related talk at PODCToday 5:10pm

“A tight unconditional lower bound on distributed random walk computation”