Exact Repair problems with multiple sources: CISS 2014

Post on 25-May-2015

101 views 0 download

Tags:

description

Consider a distributed storage system that stores redundant data to provide reliability in case of node failures. It is also desirable that these systems have exact repair functionality: If one storage node fails, others send it some information such that it reconstruct what it was storing prior to failure. We determine achievable rate regions when there are multiple sources present via a 2-source (3,2,2) exact repair problem.

transcript

CISS 2014, Princeton NJ 1

Exact Repair Problems with Multiple Sources

Jayant Apte*, Congduan Li, John MacLaren Walsh, Steven Weber

ECE Dept. Drexel University

CISS 2014, Princeton NJ 2

Outline

● Problem Definition● Computer assisted proofs: General Structure● Polyhedral bounds on● Polyhedral computation interpretation of rate

region computation● A projection technique for computing

achievable rate region

CISS 2014, Princeton NJ 3

Outline

● Problem Definition● Computer assisted proofs: General structure● Polyhedral bounds on● Polyhedral computation interpretation of rate

region computation● A projection technique for computing

achievable rate region

CISS 2014, Princeton NJ 4

(n,k,d) Exact Repair with multiple sources

CISS 2014, Princeton NJ 5

(n,k,d) Exact Repair with multiple sources

CISS 2014, Princeton NJ 6

2-source (3,2,2) exact repair problem

CISS 2014, Princeton NJ 7

2-source (3,2,2) exact repair problem

2 sources

CISS 2014, Princeton NJ 8

2-source (3,2,2) exact repair problem

3 encoding functions

CISS 2014, Princeton NJ 9

2-source (3,2,2) exact repair problem

3 storage random variables

CISS 2014, Princeton NJ 10

2-source (3,2,2) exact repair problem

3 decoders with different demands

CISS 2014, Princeton NJ 11

2-source (3,2,2) exact repair problem

6 repair encodingfunctions

CISS 2014, Princeton NJ 12

2-source (3,2,2) exact repair problem

3 repair decodingfunctions

CISS 2014, Princeton NJ 13

2-source (3,2,2) exact repair problem

Total 11 random variables

CISS 2014, Princeton NJ 14

Implicit characterization of rate region(Yan et al.)

CISS 2014, Princeton NJ 15

Outline

● Problem Definition● Computer assisted proofs: General structure● Polyhedral bounds on● Polyhedral computation interpretation of rate

region computation● A projection technique for computing

achievable rate region

CISS 2014, Princeton NJ 16

Motivation

SourcesDecoderDemands

SoftwareNetwork

CISS 2014, Princeton NJ 17

Motivation

Software

SourcesDecoderDemands

NetworkRate Region

and optimal codes

CISS 2014, Princeton NJ 18

Software for computer assisted proofs

CISS 2014, Princeton NJ 19

Computer assisted converse

CISS 2014, Princeton NJ 20

Computer assisted converse

Inequalities obtained as an implication of linear Shannon-type,non-Shannon-type, non-linear non-Shannon type inequalities andnetwork constraints

CISS 2014, Princeton NJ 21

Computer assisted converse

CISS 2014, Princeton NJ 22

Computer assisted achievability

CISS 2014, Princeton NJ 23

Software for computer assisted proofs

CISS 2014, Princeton NJ 24

Outline

● Problem Definition● Computer assisted proofs: General structure● Polyhedral bounds on● Polyhedral computation interpretation of rate

region computation● A projection technique for computing

achievable rate region

CISS 2014, Princeton NJ 25

● Closure of set of all 'entropic' vectors arising from N-variable probability distributions

3-D rendition of

CISS 2014, Princeton NJ 26

● Closure of set of all 'entropic' vectors arising from N-variable probability distributions

● Each entropic vector is formed by stacking entropies of subsets of N random variables

3-D rendition of

CISS 2014, Princeton NJ 27

● Closure of set of all 'entropic' vectors arising from N-variable probability distributions

● Each entropic vector is formed by stacking entropies of subsets of N random variables

● Cone:

3-D rendition of

CISS 2014, Princeton NJ 28

● Cannot be expressed as intersection of finite number of linear inequalities for N>3

● For N=4, existence of single nonlinear● non-Shannon inequality(necessary and

sufficient) is known [Liu & Walsh 2014]● Additionally, several hundred linear

non-Shannon inequalities are known[DFZ 2011, Csirmaz 2013]

3-D rendition of

CISS 2014, Princeton NJ 29

CISS 2014, Princeton NJ 30

CISS 2014, Princeton NJ 31

Outline

● Problem Definition● Computer assisted proofs: General structure● Polyhedral bounds on

– Shannon (Outer) bound

● Polyhedral computation interpretation of rate region computation

● A projection technique for computing achievable rate region

CISS 2014, Princeton NJ 32

Shannon Outer Bound

CISS 2014, Princeton NJ 33

Shannon Outer Bound

CISS 2014, Princeton NJ 34

Shannon Outer Bound

CISS 2014, Princeton NJ 35

Outline

● Problem Definition● Computer assisted proofs: General structure● Polyhedral bounds on

– Shannon (Outer) bound

– (Representable) Matroid (Inner) bound(s)

● Polyhedral computation interpretation of rate region computation

● A projection technique for computing achievable rate region

CISS 2014, Princeton NJ 36

(Representable) Matroid Inner bound(s)

CISS 2014, Princeton NJ 37

(Representable) Matroid Inner bound(s)

CISS 2014, Princeton NJ 38

(Representable) Matroid Inner bound(s)

CISS 2014, Princeton NJ 39

(Representable) Matroid Inner bound(s)

CISS 2014, Princeton NJ 40

(Representable) Matroid Inner bound(s)

CISS 2014, Princeton NJ 41

(Representable) Matroid Inner bound(s)

CISS 2014, Princeton NJ 42

(Representable) Matroid Inner Bound(s)

CISS 2014, Princeton NJ 43

Outline

● Problem Definition● Computer assisted proofs: General structure● Polyhedral bounds on

– Shannon (Outer) bound

– Matroid (Inner) bound(s)

– Subspace (Inner) bounds

● Polyhedral computation interpretation of rate region computation

● A projection technique for computing achievable rate region

CISS 2014, Princeton NJ 44

Subspace Inner Bound(s)

CISS 2014, Princeton NJ 45

Subspace Inner Bound(s)

CISS 2014, Princeton NJ 46

Subspace Inner Bound(s)

CISS 2014, Princeton NJ 47

Subspace Inner Bound(s)

CISS 2014, Princeton NJ 48

Subspace Inner Bound(s)

CISS 2014, Princeton NJ 49

Software for computer assisted proofs

CISS 2014, Princeton NJ 50

Polyhedral bounds on rate region

● Using polyhedral inner/outer bound on yields

polyhedral inner/outer bounds on rate region● Lemma 1: Inner bounds on rate region

computed using or are achievable using linear codes

CISS 2014, Princeton NJ 51

Outline

● Problem Definition● Computer assisted proofs: General structure● Polyhedral bounds on

– Shannon (Outer) bound

– Matroid (Inner) bound(s)

– Subspace (Inner) bounds

● Polyhedral computation interpretation of rate region computation

● A projection technique for computing achievable rate region

CISS 2014, Princeton NJ 52

Network Coding constraints

CISS 2014, Princeton NJ 53

Network Coding constraints

● Consider a type 1 or type 2 constraint H● In general, computing extreme rays of given H and

extreme rays of is equivalent to an iteration of Double Description Method of polyhedral representation conversion

● Lemma 2 [Li et al. 2013]: An extreme ray of is an extreme ray of if it is contained in the hyperplane corresponding to H

● Hence, simple membership check suffices to find extreme rays of

CISS 2014, Princeton NJ 54

Software for computer assisted proofs

CISS 2014, Princeton NJ 55

Rate constraints

Storage Bandwidth

CISS 2014, Princeton NJ 56

Rate constraints

Repair Bandwidth

Storage Bandwidth

CISS 2014, Princeton NJ 57

A projection technique for computing achievable rate region

CISS 2014, Princeton NJ 58

A projection technique for computing achievable rate region

CISS 2014, Princeton NJ 59

A projection technique for computing achievable rate region

CISS 2014, Princeton NJ 60

Polyhedral projection via chm

● chm is an implementation of polyhedral projection algorithm called Convex Hull Method by Jayant Apte*

● chmlib v0.x is available at:

http://www.ece.drexel.edu/walsh/aspitrg/software.html

● Rational arithmetic using FLINT: Fast Library for Number Theory

● Rational LP solver based on qsopt

CISS 2014, Princeton NJ 61

Polyhedral projection via chm

● Has been used for– The current work

– Computer assisted converse proofs of rate regions of Multilevel Diversity Coding Systems(a special case of multi-source network coding)

– Finding non-Shannon Information Inequalities via Generalized Copy Lemma of Csirmaz

● Can be used for – Finding necessary conditions for non-contexuality of small

marginal scenarios(Quantum Information)

CISS 2014, Princeton NJ 62

Results

SoftwareNetwork

CISS 2014, Princeton NJ 63

Rate region for H(S1)=1 and H(S2)=1

CISS 2014, Princeton NJ 64

Rate region for H(S1)=1 and H(S2)=2

CISS 2014, Princeton NJ 65

References● X. Yan, R.W. Yeung, and Zhen Zhang. An implicit characterization of the achievable rate region for acyclic

multisource multisink network coding. Information Theory, IEEE Transactions on, 58(9):5625–5639, 2012.● Dougherty, Randall, Chris Freiling, and Kenneth Zeger. "Non-Shannon information inequalities in four

random variables." arXiv preprint arXiv:1104.3602 (2011).● Csirmaz, László. "Information inequalities for four variables." CEU (2013).● Yunshu Liu and John M. Walsh, "Only One Nonlinear Non-Shannon Inequality is Necessary for Four

Variables", submitted to IEEE Int. Symp. Information Theory (ISIT2014)● Congduan Li, J. Apte, J.M. Walsh, and S. Weber. A new computational approach for determining rate regions

and optimal codes for coded networks. In Network Coding (NetCod), 2013 International Symposium on, pages 1–6, 2013.

● Congduan Li, John MacLaren Walsh, Steven Weber. Matroid bounds on the region of entropic vectors. In 51th Annual Allerton Conference on Communication, Control and Computing, October 2013.