network coding for three unicast sessions: …athina/papers/allerton2010-slides.pdfnetwork coding...
TRANSCRIPT
Network Coding for Three Unicast Sessions:Network Coding for Three Unicast Sessions:Interference Alignment Approaches
Abinesh Ramakrishnan*, Abhik Das†, Hamed Maleki*,Athina Markopoulou*, Syed Jafar*, Sriram Vishwanath†
(*) UC Irvine and (†) UT Austin
OutlineOutline
o Motivationo Network Alignment (NA: NC+IA)
o Network Alignment Approaches o at the edgeo in the middle
Wh N o When is NA necessary? o Can other schemes achieve >=½ rate or more?
C l io Conclusion
2
AlignmentAlignment
• Interference Alignment (IA) – IA originally introduced for wireless interference channels
• as a systematic way to guarantee half the rate per user
– allows to solve fewer equations for some unknowns h • needed when messages are mixed
• Network Alignment (NA=NC+IA)– IA techniques can also be applied in networks– Applied to repair problem in dist. storage [WD’09, SR’10, CJM’10]– Applied to network coding for multiple unicasts [DVJM’10]
3
The Network as a ChannelA lAnalogy
S1 D1
Multiple Unicast Network Interference Channel
S2 D2
S3 D3
• Both represented by a Linear Transfer Function [Mij]. • +: [Mij] no longer determined by nature but defined by us• -: Spatial dependencies introduced by the graph. Feasibility?
4
Network Alignment ApproachesNetwork Alignment ApproachesCoding in the Middle Coding at the Edge
5
Alignment by Coding at the EdgeAlignment by Coding at the Edge• Asymptotic scheme: Symbol Extension [CJ’08]
– recently applied to network coding for multiple unicastsrecently applied to network coding for multiple unicasts[Das, Vishwanath, Jafar, Markopoulou, ISIT’10]
S1 D1(2n+1) – symbol extension( ) yz1
(n+1) x 1
S2 D2z2
n x 1
S3 D3z3
6
z3n x 1
Alignment by Coding at the EdgeAlignment by Coding at the Edge• Asymptotic scheme: Symbol Extension [CJ’08]
– recently applied to network coding for multiple unicasts [DVJM’10]recently applied to network coding for multiple unicasts [DVJM 10]
S1 D1(2n+1) – symbol extension( ) yz1
(n+1) x 1V1
(2n+1) x (n+1)
S2 D2z2
n x 1V2
(2n+1) x n
S3 D3z3 V
(2n+1) x n
7
z3n x 1 (2n+1) x n
V3
Alignment by Coding at the EdgeAlignment by Coding at the Edge• Asymptotic scheme: Symbol Extension [CJ’08]
– recently applied to network coding for multiple unicasts [DVJM’10]recently applied to network coding for multiple unicasts [DVJM 10]
S1 D1(2n+1) – symbol extension( ) yz1
(n+1) x 1 x1=V1z1
S2 D2z2
n x 1 x2=V2z2
S3 D3z3
8
z3n x 1 x3=V3z3
Alignment by Coding at the EdgeAlignment by Coding at the Edge• Asymptotic scheme: Symbol Extension [CJ’08]
– recently applied to network coding for multiple unicasts [DVJM’10]recently applied to network coding for multiple unicasts [DVJM 10]
S1 D1(2n+1) – symbol extension
M11V1z1
( ) yz1
(n+1) x 1M12V2z2 + M13V3z3
x1=V1z1
S2 D2z2
n x 1 x2=V2z2
M22V2z2
S3 D3z3
M21V1z1 + M23V3z3
M33V3z3
9
z3n x 1 x3=V3z3
M33V3z3
M31V1z1 + M32V2z2
Alignment by Coding at the EdgeAlignment by Coding at the Edge• Asymptotic scheme: Symbol Extension [CJ’08]
– recently applied to network coding for multiple unicasts [DVJM’10]recently applied to network coding for multiple unicasts [DVJM 10]
S1 D1(2n+1) – symbol extension
z1
( ) yz1
(n+1) x 1 x1=V1z1
S2 D2z2
n x 1 x2=V2z2
z2
S3 D3z3
z3
10
z3n x 1 x3=V3z3
Feasibility is a challenge (unlike wireless) due to dependencies in [Mij]
Feasibility conditionsy[DVJM’10]
By construction:• mij(ξ), i,j =1,2,3, non-trivial polynomialsj j p y• mii(ξ) ≠ c mij(ξ) for any c in Fp\{0}
Notation:
Sufficient conditions for asymptotic alignment [DVJM’10] : • for all n, and pi, qi (i=0,1,2,..n) it should be:
[M11V1 M12V2] is full rank
11
QuestionsQuestions
When can we align:• Intuition behind conditions? Relation to Network Structure?• Infinitely many and complicated conditions to check. Simplify?• How mild are these conditions? (e.g. hold almost always in wireless)
Why should we align? • How much benefit/loss compared to alternatives?• Is alignment necessary? Is alignment necessary?
How to align?• Is asymptotic alignment the only way? Algorithms?
Focus mostly on: 3 unicasts, min-cut =1 per session.
12
Understanding the conditionsUnderstanding the conditions
• Let’s look at a subset of all conditions– (p0=1, qo=1) and (p0=1, q1=1)
• Interestingly, these are also necessary for any scheme to achieve rate >=1/2 in the wireless interference channel [CJ, ToIT’09]f [ J, ]
13
Understanding the conditionsUnderstanding the conditions
1
2
1
2
3
2
3
2
14
3 3
Conditions and network structureConditions and network structure
1 11
2
11
2
1
2 2
33 3
2
1 11 1 1
2
11
2
1
15
2
3
2
3
Simplifying the conditions?Simplifying the conditions?
• Conjecture:jThe “small” conditions are sufficient for asymptotic alignment.
?
16
Alignment Approaches g ppto Code at the Edge or in the Middle?
Coding in the Middle Coding at the Edge
Symbol Extension Method:17
Symbol Extension Method:• intelligence at the edge, middle is simple• widely applicable• large number of symbols and finite field
Example of Coding in the MiddleExample of Coding in the Middle• “Ergodic’’ Alignment
– inspired by [Nazer Gastpar Jafar Vishwanath “Ergodic IA” ISIT’09]inspired by [Nazer, Gastpar, Jafar, Vishwanath, Ergodic IA , ISIT 09]– choose coding coefficients over two time slots so that:
⎥⎥⎤
⎢⎢⎡
= )1(1312
)1(11
)1( mmmmmm
M ⎥⎥⎤
⎢⎢⎡
= )2(1312
)2(11
)2( mmmmmm
M
– then subtract and obtain 1 symbol in 2 time slots
l d
⎥⎥
⎦⎢⎢
⎣
=)1(
333231
232221
mmmmmmM
⎥⎥
⎦⎢⎢
⎣
=)2(
333231
232221
mmmmmmM
• Feasibility condition: – each mii is not a function of the mij‘s– more restrictive than the condition for asymptotic alignment
St n ths:• Strengths:– 2 time slots are enough (no need to wait for channel opportunities)– smaller field size and number of symbols than asymptotic scheme
• Limitation:Limitation:– Intelligence resides in the network. May be complex for large networks.
18
How mild are the conditions? Random Graphs
19
How mild are the conditions? Real Topologies
20
How mild are the conditions? Real Topologies
21
OutlineOutline
o Motivationo NC+IA=NA
o Network Alignment Approaches o At the edgeo In the middle
o When is NA necessary? o Can other schemes achieve ½ the min-cut or more?
C l io Conclusion
22
Example: Extended ButterflyExample Extended Butterfly
• K unicast sessions going through the same bottleneck• Routing achieves rate = 1/k Routing achieves rate 1/k • Network coding (with all side links) achieves rate = 1 • Alignment (with sufficiently rich side links) achieves rate =1/2
1 1
ax1+bx2+cx3
x1
1 1
x2
2 2
23x33 3
Example 1 3 unicast sessions, min-cut=1
• No side links ⎥⎤
⎢⎡
bcba
M• Rate 1/3 by any scheme⎥⎥⎥
⎦⎢⎢⎢
⎣
=cbacbaM
1 1
ax1+bx2+cx3
x1
1 1
x2
2 2
24
x33 3
Example 13 unicast sessions, min-cut=1
• One side link ⎥⎤
⎢⎡ +
bccba
M13
• Still rate 1/3 by any scheme⎥⎥⎥
⎦⎢⎢⎢
⎣
=cbacbaM
1 1
ax1+bx2+cx3
x1
1 1
x2
2 2
25
x33 3
Example 1 3 unicast sessions, min-cut=1
• Two receivers have side links (which ones matter) ⎥⎤
⎢⎡ +
bccba
M13
• Here: still 1/3 rate, NA not possible⎥⎥⎥
⎦⎢⎢⎢
⎣ +=
ccbacbaM
32
c
c
c
c
1 1
ax1+bx2+cx3
x1
1 1
x2
2 2
26
x33 3
Example 1 3 unicast sessions, min-cut=1
• Two receivers have side links ⎥⎤
⎢⎡ +
bccba
M13c c
• 1/2 min-cut achievable in 2 slots• by alignment or • or by sharing between session 2 and butterfly 1-3
⎥⎥⎥
⎦⎢⎢⎢
⎣ +=
cbcacbaM
31c c
1 1
ax1+bx2+cx3
x1
1 1
x2
2 2
27
x33 3
Example 1 3 unicast sessions, min-cut=1
• Three receivers have one side link.The optimal rate is ½ the min cut • The optimal rate is ½ the min-cut,
• NA achieves it; routing does not; there are no butterflies• Alignment needed if intelligence is allowed only at the edge• Coding in the middle (RNC + deterministic at 1’,2’,3’) can also achieve ½ the min-cut
1 1
1’
ax1+bx2+cx3
x1
1 1
2’
x3
x2
2 2
3’
x1
28
x33 3
3
x2
When is NA necessary?summary
• Arbitrary network, 3 unicast sessions with min-cut=1:y– Theorem: Whenever NA is possible, another approach (e.g., routing,
butterflies, NC in the middle w/o alignment) can also achieve half the min-cut.P f li – Proof outline:
• Sparsity bound S=1/3: the extended butterfly examples extend to any network• Sparsity bound >=1/2: consider networks where routing rate <= 1/2, and construct a
deterministic NC scheme (in the middle, w/o alignment) that achieves ½ the min-cut
• NA necessary (depending on the topology) to achieve ½ for:– K>3 unicastsK– or min-cut>1
29
Example 2:4 unicast sessions, min-cut=1
• Alignment needed: • receiver 3 has 2 equations, 4 unknowns in 2 slots
d 3’ h 3 4 k 2 l• even node 3’ has 3 equations, 4 unknowns in 2 slots• alignment achieves ½ min-cut, no other scheme does
αx1+βx2+γx3+δx4
x1
αx1 βx2 γx3 δ 4
1 1
x2
2 2
x33 3
3’
30x4
4 4
Example 2:5 unicast sessions, min-cut=1
• Alignment needed - effect amplified• all receivers have 2 (nodes in the middle have 2) equation,s with 5 unknowns in 2 slots
li hi ½ i h h d• alignment achieves ½ min-cut, no other scheme does
1 1x1
1
2
1
2x2
2
3 3
2
x33
x4
4 4
31x5
5 5
Example 3: 3 unicasts sessions, min-cut = 2
• Alignment needed:½ ½• It achieves ½min-cut, which is optimal; no other scheme achieves ½ rate.
a1x+b1y+c1zPick c3 s.t.:ba2x+b2y+c2z
32
1
2
1
ccc
bb
+=
1 1c3z
3
2 2
32
3
OutlineOutlineo Motivation
o NC+IA=NA, analogy and dependencies
o Network Alignment Approaches o At the edgeo In the middle
o When is NA possible/necessary? o Can other schemes achieve ½ the min-cut or more?
o Conclusion
33
ConclusionConclusion• Network Alignment (NA=NC+IA)
– a systematic approach for network coding across multiple unicasts– guarantees half the min-cut per session
• Future Directions:
– Characterize Feasibility and Performance of NA and their relation to Network Structure
– Develop practical Alignment Algorithms.
34