Streaming Transmission of Poisson Traffic: Won S. Yoon

Streaming Transmission of Poisson Traffic
Won S. Yoon
wonsyoonalum.mit.edu
ISIT 2008
ISIT 2008 Streaming Transmission of Poisson Traffic – 1 of 29

Outline of Talk
Introduction
Physical Layer ⊲ Introduction: why stream codes?

Network Layer
Discussion

Outline of Talk
Introduction

Network Layer
Discussion ⊲ Physical Layer: random coding error exponents.

Outline of Talk
Introduction

Network Layer
⊲ Network Layer: queueing model, delay analysis.

Outline of Talk
Introduction

Network Layer
⊲ Network Layer: queueing model, delay analysis.
⊲ Discussion: stream codes vs. block codes.

Introduction
Physical Layer
Network Layer
Discussion
Introduction

Classical Information Theory
Introduction Traditionally,
Physical Layer
• Assume an infinite backlog of data.
Network Layer
• Try to maximize throughput and minimize error.
Discussion
• Solution: encode a large number of bits into a long codeword.

Classical Information Theory
Introduction Traditionally,
Physical Layer
• Assume an infinite backlog of data.
Network Layer
• Try to maximize throughput and minimize error.
Discussion
• Solution: encode a large number of bits into a long codeword.
However, this does not consider:

1. Delay constraints.
2. Bursty arrivals of data.

What If Data is Bursty?
Introduction Packets arrive every τ slots, deterministically.

Physical Layer
Network Layer
Discussion


Physical Layer
Consider a block code: let d1 = decoding delay for one packet.
Network Layer
Discussion


Physical Layer
Network Layer
Discussion • For slow arrivals τ > d1 :


Physical Layer
Network Layer

– Just encode each packet independently.
τ τ τ τ
d1 d1 d1 d1


Physical Layer
Network Layer

τ τ τ τ
d1 d1 d1 d1
• For faster arrivals τ < d1 :


Physical Layer
Network Layer

τ τ τ τ
d1 d1 d1 d1

– Need to jointly-encode multiple packets per codeword.
τ τ τ τ τ τ τ τ
d3 d3


Physical Layer
Network Layer

τ τ τ τ
d1 d1 d1 d1

– Need to jointly-encode multiple packets per codeword.
d3 d3
– Faster arrivals → longer codewords → larger delays.

A Cross-Layer Perspective
Introduction Fundamental tradeoff between physical layer and network layer:

Physical Layer
For a fixed error probability:
Network Layer
Discussion


Physical Layer
Network Layer
Discussion • Information theory: aggregation is good.

– Long codewords → increase throughput.


Physical Layer
Network Layer

• Queueing theory: aggregation is bad.

– Small chunks of data → reduce delay.


Physical Layer
Network Layer

• Queueing theory: aggregation is bad.

– Small chunks of data → reduce delay.
Objective: minimize delay for fixed throughput.

Disadvantages of Block Coding
Introduction For large throughput, need to jointly-encode many packets.

Physical Layer
Network Layer
Discussion


Physical Layer
• Long codewords → increase both queueing (encoding) delay
Network Layer
and transmission (decoding) delay.
Discussion
d3 d3


Physical Layer
Network Layer
Discussion
d3 d3
Idea: relax the block-coding constraint


Physical Layer
Network Layer
Discussion
d3 d3
Idea: relax the block-coding constraint → stream codes.


Physical Layer
Network Layer
Discussion
d3 d3
• Encode/decode in a continuous stream.

– Generalizes block codes: performs at least as well.


Physical Layer
Network Layer
Discussion
d3 d3
• Encode/decode in a continuous stream.

– Generalizes block codes: performs at least as well.
– Can streaming do strictly better?

Rate-Adaptive Stream Coding
Introduction Rate control: adapt the rate of packets encoded into the stream.
Physical Layer
Network Layer Fundamental tradeoff:

Discussion • Faster encoding → less queueing, longer transmission.

Physical Layer

– Analogy: network admission control.

Physical Layer

End-point solutions:
1. Admit one packet at a time (patient): per-packet block code.
2. Admit all packets at once (greedy): maximally-joint block code.

Physical Layer

End-point solutions:
1. Admit one packet at a time (patient): per-packet block code.
2. Admit all packets at once (greedy): maximally-joint block code.
Streaming achieves the continuum of solutions in-between.

Problem Formulation
Introduction
Physical Layer Transmitter
Network Layer
Rate Encoder Channel Decoder
ctrl
Discussion queue
data
packets
• Data packets arrive randomly, queued at the transmitter.

Problem Formulation
Introduction
Network Layer
ctrl
Discussion queue
data
packets

• Design the rate control: number of packets encoded per slot.

Problem Formulation
Introduction
Network Layer
ctrl
Discussion queue
data
packets

• Design the rate control: number of packets encoded per slot.
• Objective: minimize delay for fixed error and throughput.
– Delay is measured end-to-end (queueing + transmission).

Previous Work
Introduction Studies of adaptive code rate for bursty delay-constrained data.

Physical Layer
Network Layer • Block codes

Discussion

Previous Work

Physical Layer

Discussion – In queueing literature: this is known as batch service.

Previous Work

Physical Layer

– Information-theoretic view: [MusyTelatar06℄ studied rate
control for single-batch and Poisson arrivals.

Previous Work

Physical Layer

• Stream codes

Previous Work

Physical Layer

• Stream codes
– Information-theoretic view: coding and error exponents.

Previous Work

Physical Layer

• Stream codes
– No known studies of delay for bursty data.

Previous Work

Physical Layer

• Stream codes
– No known studies of delay for bursty data.
Relevant but not related:

• Rateless codes / RCPCs: variable decoding, block encoding.
• Streaming turbo codes: a practical implementation?

Preview of Our Analysis
Introduction Two-step analysis:

Physical Layer
Network Layer
Discussion


Physical Layer
Network Layer
1. Physical layer: random-coding error exponents.
Discussion


Physical Layer
Network Layer
1. Physical layer: random-coding error exponents.
Discussion
2. Network layer: queueing model, end-to-end delay.

Introduction
Physical Layer
Network Layer
Discussion
Physical Layer

The Encoder
Introduction Rate-adaptive convolutional shift-register.

Physical Layer
Network Layer
Discussion

The Encoder

Physical Layer
In time slot t,
Network Layer
Discussion

The Encoder

Physical Layer
In time slot t,
Network Layer
Discussion
• Slide the next packet into register as message Ut .
– If no packet, encode null symbol (known at receiver).

The Encoder

Physical Layer
In time slot t,
Network Layer
Discussion
• Encode the entire contents of register into a channel symbol,

Xt = f (U1 , . . . , Ut ).

The Encoder

Physical Layer
In time slot t,
Network Layer
Discussion
• Encode the entire contents of register into a channel symbol,

Xt = f (U1 , . . . , Ut ).
Code tree representation:

Dynamics of the Tree
Introduction Tree grows at a time-varying rate depending on the rate control.

Physical Layer
Network Layer
N(t)
Discussion


Physical Layer
Network Layer
N(t)
Discussion
• Faster rate control, faster tree growth.


Physical Layer
Network Layer
N(t)
Discussion

• If queue is empty, erase code memory and re-initialize tree.


Physical Layer
Network Layer
N(t)
Discussion

• Tree has alternating cycles of “busy” and “idle” periods.


Physical Layer
Network Layer
N(t)
Discussion

• Tree has alternating cycles of “busy” and “idle” periods.
Required memory ∼ number of packets in a busy period.

• This is guaranteed to be finite as long as the system is stable.

The Decoder
Introduction At time t, MAP-decode all packets in parallel: (Û1 , . . . , Ût )

Physical Layer
Network Layer
Discussion

The Decoder

Physical Layer
Network Layer • Observe new channel output Yt = Xt + Nt .

Discussion

The Decoder

Physical Layer

Discussion
• Decode stage t − d using the entire history, (Y1 , . . . , Yt )

The Decoder

Physical Layer

Discussion
• Define MAP probability of packet t − d as
Λ(y1t ) = max Pr(Xt−d = x|Y1t = y1t )

x

The Decoder

Physical Layer

Discussion

x
– Proportional to sum likelihoods for paths sharing branch t−d:

X X
Λ(y1t ) ∼ max ··· Pr(Y1t = y1t |X1t = xt1 )
xt−d
x1 /xt−d xt

The Decoder

Physical Layer

Discussion

x
– Proportional to sum likelihoods for paths sharing branch t−d:

X X
Λ(y1t ) ∼ max ··· Pr(Y1t = y1t |X1t = xt1 )
xt−d
x1 /xt−d xt
• MAP ⇐⇒ maximize aggregate likelihoods.

The Decoding Algorithm
Introduction • Decode packet t − d by comparing groups of trees.

Physical Layer T
Network Layer Ut 1 0 1 1 0 0 1
t
Discussion
0
subtrees
0 with correct
time-T bit.
1


Physical Layer T
t
Discussion
0
subtrees
0 with correct
time-T bit.
1
• Select the group with maximum aggregate likelihood.


Physical Layer T
t
Discussion
0
subtrees
0 with correct
time-T bit.
1
• Select the group with maximum aggregate likelihood.

• Do this for all packets: decode in parallel.
– Packets may depart out-of-order.

Sequential Decoding
Introduction What if we use sequential decoding (e.g., the Fano algorithm)?

Physical Layer
Network Layer
Discussion

Sequential Decoding

Physical Layer
• Find the best path in an iterative manner.
Network Layer
Discussion

Sequential Decoding

Physical Layer
Network Layer
– Limited by reliability of first stage: first-in-first-out.
Discussion
– “Slow-truck” effect: large burst slows down smaller ones.

Sequential Decoding

Physical Layer
Network Layer
Discussion
⇒ In general, sequential has larger delay than MAP.

Sequential Decoding

Physical Layer
Network Layer
Discussion

• If rate control is binary (same burst sizes): MAP is also serial,
same error-delay performance as sequential.

Sequential Decoding

Physical Layer
Network Layer
Discussion

• If rate control is binary (same burst sizes): MAP is also serial,
same error-delay performance as sequential.
Quantify this error-delay performance.

Error Exponents: A Brief Review
Introduction • Block Codes:

Physical Layer
Network Layer
Discussion


Physical Layer
– Random coding bound: decoding m bits after a delay d,
Network Layer
Discussion


Physical Layer
Network Layer
Pe ≤ βe−dEr ( d )
m
Discussion
Er (R) is Gallager’s block RCE, maxρ∈(0,∞) [E0 (ρ) − Rρ].


Physical Layer
Network Layer
m
Discussion
– To achieve Pe ≤ ǫ, need large enough delay d.

⊲ For AWGN channel with average power constraint:
mρ m
dm ≥ log ǫ + = σ+ (setup cost + workload)
E0 µ


Physical Layer
Network Layer
m
Discussion

mρ m
E0 µ
• Stream Codes:


Physical Layer
Network Layer
m
Discussion

mρ m
E0 µ
• Stream Codes:
– For ML [Forney74℄ and sequential [Jelinek74,Anderson92℄
fixed-rate decoding, the block-coding RCE applies.
– What about for per-packet MAP decoding?

Error Exponents: Stream Codes
Introduction
Physical Layer
For MAP decoding of stage t after delay d:
Network Layer "N #
Xt
Discussion
Pe,t,d ≤ e−dj E0 (ρ)+jmρ e−dE0 (ρ)+(1+Nd )mρ
k=0

Error Exponents: Stream Codes
Introduction
Physical Layer
For MAP decoding of stage t after delay d:
Network Layer "N #
Xt
Discussion
Pe,t,d ≤ e−dj E0 (ρ)+jmρ e−dE0 (ρ)+(1+Nd )mρ
k=0
• Nt packets encoded in the past [0, t).

• Nd packets encoded in the future [t, t + d].
• dj : backwards delay to j -th previous packet.
Nt Nd
0 t t+d

Streaming Decoding Delay
Introduction • For reliable transmission over an AWGN channel:

Physical Layer PNt −dj E0 +jmρ
1

Network Layer
log ǫ + mρ + log j=0 e
d ≥
Discussion E0 − λmρ


1

Network Layer
d ≥
1. First term: workload of a single packet.


1

Network Layer
d ≥

2. Second term: workload due to the past.


1

Network Layer
d ≥

3. Denominator: effect of the future.


1

Network Layer
d ≥

3. Denominator: effect of the future.
• Simplified reliability condition:

h i
σ + µ + wpast (dN
1
1 )
t
d ≥ λ
1− µ
E0 1 1
– µ= mρ and σ = E0 log ǫ
PNt
– wpast = log j=0 e−dj E0 +jmρ

Introduction
Physical Layer
Network Layer
Discussion
Network Layer

Queueing Model
Introduction Parallel-service model:

Physical Layer
Network Layer
Discussion

Queueing Model

Physical Layer
Network Layer 1. A new packet enters service.

Discussion

Queueing Model

Physical Layer

Discussion – Its workload is increased by past packets.

Queueing Model

Physical Layer

– In turn, it increases the workload for past packets.

Queueing Model

Physical Layer

2. Every packet receives one unit of service per slot.

Queueing Model

Physical Layer

Appears similar to processor-sharing, but two key differences:

Queueing Model

Physical Layer


– Future data impose more workload than past data.

Queueing Model

Physical Layer


– Future data impose more workload than past data.
– Aggregate workload is sub-additive (joint coding efficiency).

Queueing Model
Introduction
Physical Layer
Network Layer
Discussion

Queueing Model
Introduction
Physical Layer
Network Layer
Discussion

Queueing Model
Introduction
Physical Layer
Network Layer
Discussion

The Effect of the Past
1
Introduction The future is simple: each packet imposes another µ workload.
Physical Layer
Network Layer
Discussion

1
Physical Layer
However, the delay due to the past is more complicated:
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ
past mρ E log j=0 e
Dstream =
µ−λ

1
Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ
Dstream =
µ−λ
Bounds:
 
N
X
E max Sj ≤ E log eSj 
0≤j≤N
j=0

1
Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ
Dstream =
µ−λ
Bounds:
    
N
X  XN 
E max Sj ≤ E log eSj  ≤ log E  eSj 
0≤j≤N  
j=0 j=0

1
Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ
Dstream =
µ−λ
Bounds:
    
N
X  XN 
0≤j≤N  
j=0 j=0
• Sj is a backwards random walk with steps ∆i = (−τi + µ1 ).

1
Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ
Dstream =
µ−λ
Bounds:
    
N
X  XN 
0≤j≤N  
j=0 j=0
• Sj is a backwards random walk with steps ∆i = (−τi + µ1 ).
– Note: similar to random walk in G/G/1 queueing delay:

queue 1
DG/G/1 = max Sj′ ′
with ∆j = −τi + σ +
0≤j≤n µ

Case I: Deterministic Arrivals
Introduction Assume that packets arrive once every 1/λ.

Physical Layer
Network Layer
Discussion


Physical Layer
Network Layer Optimal streaming policy: encode immediately upon arrival.

Discussion
1
For λ < d1 :


Physical Layer

Discussion
1 ∗ 1
For λ < d1 : Dstream = σ+ µ


Physical Layer

Discussion
1 ∗ 1 ∗
For λ < d1 : Dstream = σ+ µ = Dblock


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 :


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
∗
Dstream =
µ−λ


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
1
log 1
∗ mρ 1−β
Dstream = +
µ−λ µ−λ


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
1
log 1
∗ mρ 1−β
Dstream = +
µ−λ µ−λ
1
Compare with block codes in the regime λ > d1 :


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
1
log 1
∗ mρ 1−β
Dstream = +
µ−λ µ−λ
1
∗ σµ + 1
Dblock =
µ−λ


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
1
log 1
∗ mρ 1−β
Dstream = +
µ−λ µ−λ
1
∗ σµ + 1 1 σµ + 1
Dblock = +
µ−λ 2 µ−λ


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
1
log 1
∗ mρ 1−β
Dstream = +
µ−λ µ−λ
1
∗ σµ + 1 1 σµ + 1
Dblock = +
µ−λ 2 µ−λ
• Comparing the second terms (delay due to the past),


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
1
log 1
∗ mρ 1−β
Dstream = +
µ−λ µ−λ
1
∗ σµ + 1 1 σµ + 1
Dblock = +
µ−λ 2 µ−λ
1 1
mρ log 1−β 1 σµ + 1
<
µ−λ 2 µ−λ


Physical Layer

Discussion
1 ∗ 1 ∗
1
For λ > d1 : σµ + 1
1
log 1
∗ mρ 1−β
Dstream = +
µ−λ µ−λ
1
∗ σµ + 1 1 σµ + 1
Dblock = +
µ−λ 2 µ−λ
1 1
mρ log 1−β 1 σµ + 1
<
µ−λ 2 µ−λ
∗ ∗
=⇒ Dstream < Dblock for high throughput.

Case II: Poisson Arrivals
Introduction Packets arrive randomly according to a Poisson process.

Physical Layer
Network Layer
Discussion


Physical Layer
• In general, difficult to find the optimal rate control.
Network Layer
Discussion


Physical Layer
Network Layer
Discussion
• Use a technique from [MusyTelatar06℄: find bounds and show
that bounds are very close.


Physical Layer
Network Layer
Discussion
σµ + 1 ∗
≤ E[Dstream ]
µ−λ


Physical Layer
Network Layer
Discussion
σµ + 1 ∗ σµ + 1
≤ E[Dstream ] ≤
µ−λ µ−λ


Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ
σµ + 1 ∗ σµ + 1 mρ E log j=0 e
≤ E[Dstream ] ≤ +
µ−λ µ−λ µ−λ


Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ
Compare with block codes [MusyTelatar06℄:


Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ

σµ + 1 ∗
≤ E[Dblock ]
µ−λ


Physical Layer
Network Layer
Discussion
h P i
1 Nt −dj E0 +jmρ

σµ + 1 ∗ σµ + 1 1 σµ + 1
≤ E[Dblock ] ≤ +
µ−λ µ−λ 2 µ−λ

Poisson Arrivals
50000
Introduction
Physical Layer
Network Layer
40000
Discussion
30000
delay (slots)
20000
10000
0
0.0 0.2 0.4 0.6 0.8 1.0
–––– block l.b. utilization factor

——— stream u.b.
······ stream l.b.

Introduction
Physical Layer
Network Layer
Discussion
Discussion

Summary and Future Work
Introduction 1. For deterministic and Poisson arrivals,

Physical Layer
– Streaming can improve delay.
Network Layer
– Greedy rate control is nearly optimal.
Discussion
– Streaming becomes more beneficial when power-limited.
2. Compare the effect of the past at physical-layer vs. at network-

layer (effect of future appears to be the same).
Future work:
• Simplify the past workload expression.
• Examine more bursty arrivals.
– Greedy policies not optimal anymore?
• Other applications for streaming transmission.

Streaming Transmission of Poisson Traffic: Won S. Yoon

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Streaming Transmission of Poisson Traffic: Won S. Yoon

Uploaded by

Copyright:

Available Formats

Streaming Transmission of Poisson Traffic

ISIT 2008 Streaming Transmission of Poisson Traffic – 1 of 29

Physical Layer ⊲ Introduction: why stream codes?

ISIT 2008 Streaming Transmission of Poisson Traffic – 2 of 29

Physical Layer ⊲ Introduction: why stream codes?

Discussion ⊲ Physical Layer: random coding error exponents.

ISIT 2008 Streaming Transmission of Poisson Traffic – 2 of 29

Physical Layer ⊲ Introduction: why stream codes?

Discussion ⊲ Physical Layer: random coding error exponents.

⊲ Network Layer: queueing model, delay analysis.

ISIT 2008 Streaming Transmission of Poisson Traffic – 2 of 29

Physical Layer ⊲ Introduction: why stream codes?

Discussion ⊲ Physical Layer: random coding error exponents.

⊲ Network Layer: queueing model, delay analysis.

⊲ Discussion: stream codes vs. block codes.

ISIT 2008 Streaming Transmission of Poisson Traffic – 2 of 29

ISIT 2008 Streaming Transmission of Poisson Traffic – 3 of 29

ISIT 2008 Streaming Transmission of Poisson Traffic – 4 of 29

However, this does not consider:

ISIT 2008 Streaming Transmission of Poisson Traffic – 4 of 29

Introduction Packets arrive every τ slots, deterministically.

ISIT 2008 Streaming Transmission of Poisson Traffic – 5 of 29

Introduction Packets arrive every τ slots, deterministically.

ISIT 2008 Streaming Transmission of Poisson Traffic – 5 of 29

Introduction Packets arrive every τ slots, deterministically.

Discussion • For slow arrivals τ > d1 :

ISIT 2008 Streaming Transmission of Poisson Traffic – 5 of 29

Introduction Packets arrive every τ slots, deterministically.

Discussion • For slow arrivals τ > d1 :

ISIT 2008 Streaming Transmission of Poisson Traffic – 5 of 29

Introduction Packets arrive every τ slots, deterministically.

Discussion • For slow arrivals τ > d1 :

• For faster arrivals τ < d1 :

ISIT 2008 Streaming Transmission of Poisson Traffic – 5 of 29

Introduction Packets arrive every τ slots, deterministically.

Discussion • For slow arrivals τ > d1 :

• For faster arrivals τ < d1 :

ISIT 2008 Streaming Transmission of Poisson Traffic – 5 of 29

Introduction Packets arrive every τ slots, deterministically.

Discussion • For slow arrivals τ > d1 :

• For faster arrivals τ < d1 :

– Faster arrivals → longer codewords → larger delays.

ISIT 2008 Streaming Transmission of Poisson Traffic – 5 of 29

Introduction Fundamental tradeoff between physical layer and network layer:

ISIT 2008 Streaming Transmission of Poisson Traffic – 6 of 29

Introduction Fundamental tradeoff between physical layer and network layer:

Discussion • Information theory: aggregation is good.

ISIT 2008 Streaming Transmission of Poisson Traffic – 6 of 29

Introduction Fundamental tradeoff between physical layer and network layer:

Discussion • Information theory: aggregation is good.

• Queueing theory: aggregation is bad.

ISIT 2008 Streaming Transmission of Poisson Traffic – 6 of 29

Introduction Fundamental tradeoff between physical layer and network layer:

Discussion • Information theory: aggregation is good.

• Queueing theory: aggregation is bad.

Objective: minimize delay for fixed throughput.

ISIT 2008 Streaming Transmission of Poisson Traffic – 6 of 29

Introduction For large throughput, need to jointly-encode many packets.

ISIT 2008 Streaming Transmission of Poisson Traffic – 7 of 29

Introduction For large throughput, need to jointly-encode many packets.

ISIT 2008 Streaming Transmission of Poisson Traffic – 7 of 29

Introduction For large throughput, need to jointly-encode many packets.