powerpoint presentation · burst-erasure channels 6 any code with t (constant), b and r satisfies...

33
Low-Delay Codes Minimizing the Average Delay Among Lost Packets 1 Yuval Cassuto joint work with: Nitzan Adler Technion, EE BIRS 15w5150 [ISIT ’15: Low-Delay Erasure-Correcting Codes with Optimal Average Delay]

Upload: others

Post on 31-Jul-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Low-Delay Codes Minimizing the Average Delay Among Lost Packets

1

Yuval Cassuto joint work with:

Nitzan Adler

Technion, EE

BIRS 15w5150

[ISIT ’15: Low-Delay Erasure-Correcting Codes

with Optimal Average Delay]

Page 2: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Coding performance classification

2

Deterministic Stochastic

worst-case , adversarial expected

average

Do not confuse!

Page 3: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Burst-erasure channel

3

packets of

data & parity

symbols

Burst-erasure

B sequential

packets are erased

Page 4: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Low delay codes

4

Trading-off information rate, error tolerance and decoding delay.

Decoding delay

Information

rate

Error

tolerance

Classical

Error-correcting

codes

Page 5: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

5

Decoding delay

The decoding delay is the number of packets from a packet erasure to its reconstruction.

i

B erased packets The packets that participate

in the reconstruction

T

The ‘i’-th packet is

reconstructed

Page 6: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Review: Low-delay codes for burst-erasure channels

6

Any code with T (constant), B and R satisfies the lower bound:

B [packets]– the burst length

T [packets]– the constant delay

R – the code rate

• Martinian & Sundberg “Burst erasure correction codes with low decoding delay” (2004)

𝑇 ≥ 𝐵 ∙

𝑅

1 − 𝑅 , 𝑅 ≥

1

2

𝐵 , 𝑅 <1

2

Page 7: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Low delay codes – some previous work

7

• Martinian & Trott, “Delay-optimal burst erasure code construction” (2007)

• Li, Khisti, Girod, “Correcting erasure bursts with minimum decoding delay” (2011)

• Badr, Khisti, Martinian, “Diversity embedded streaming erasure codes (de-sco): Constructions and optimality”(2011)

• …

Page 8: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Streaming in the old world

8

6 5 4 3 2 1 6 5 4 3 2 1

Page 9: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Streaming in the new world

9

6 5 4 3 2 1 6 5 4 3 2 1

Page 10: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

10

6 5 4 3 2 1

𝑇 → 𝑇

Low average delay is translated to more information available for processing

6 5 4 3 2 1

Page 11: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Heterogeneous delay codes - Parameter definitions

11

packet : th-’iThe delay of the ‘

Average delay:

𝑇 = 𝑇𝑖𝐵𝑖=1

𝐵

B erased

packets

The packets that

participate in the

reconstruction

T1

Ti

TB

Page 12: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Reconstruction delay

12

:reconstruction delayThe

𝜅𝑖= 𝑇𝑖 − (𝐵 − 𝑖)

i

B erased

packets

The packets that

participate in the

reconstruction

Ti

𝜅i

𝑇 = 𝜅 +𝐵−1

2

Page 13: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Example: Reconstruction delay for constant-delay codes

13

B erased packets The packets that participate

in the reconstruction

𝜅 B

𝜅 B-1

𝜅 2

𝜅 1

𝜅1 + 𝐵 − 1 = 𝜅2 + (𝐵 − 2) = ⋯ = 𝜅𝐵−1 + 1 = 𝜅𝐵= 𝑇

Page 14: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Lower bound on average delay Theorem:

The average delay 𝑇 obtained for a decoding instance following an erasure burst of length B, must satisfy:

𝑇 ≥

𝑅

1 − 𝑅⋅𝐵 + 1

2+𝐵 − 1

2 , 𝑅 ≥

1

2

𝐵

2

𝑅

1 − 𝑅+ 1 ,

1

2> 𝑅 ≥

1

1 + 𝐵

𝐵 + 1

2 , 𝑅 <

1

1 + 𝐵

14

Page 15: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Lower bounds comparison

15

Ex.: 𝐵 = 4

Page 16: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Per packet: S data symbols, P parity symbols

* A similar proof can be done for non-systematic codes.

Lower bound on average delay- Proof

16

1st reconstructed packet 𝜅𝑓𝑖𝑟𝑠𝑡≥

𝑆

𝑃

2nd reconstructed packet

𝜅𝑠𝑒𝑐𝑜𝑛𝑑≥

2𝑆

𝑃

last reconstructed packet

𝜅𝑙𝑎𝑠𝑡≥

𝐵𝑆

𝑃

𝜅 ≥

𝑆𝑃+

2𝑆𝑃

+⋯+𝐵𝑆𝑃

𝐵

Page 17: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Lower bound on average delay- Proof

17

𝑹 ≥𝟏

𝟐:

𝜅 ≥

𝑆𝑃+

2𝑆𝑃

+⋯+𝐵𝑆𝑃

𝐵≥

𝑆𝑃+2𝑆𝑃+⋯+

𝐵𝑆𝑃

𝐵=

𝑅

1 − 𝑅⋅𝐵 + 1

2

𝑇 ≥

R

1 − R⋅𝐵 + 1

2+𝐵 − 1

2

The bound may only be met for 𝑅 =𝑚

1+𝑚 , 𝑚 ∈ ℤ+

Page 18: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Lower bound on average delay

18

𝟏

𝟏+𝑩≤ 𝑹 <

𝟏

𝟐:

𝜅 ≥

𝑆𝑃+

2𝑆𝑃

+⋯+𝐵𝑆𝑃

𝐵≥ … ≥

𝑅

1 − 𝑅⋅𝐵

2+1

2

𝑇 ≥

𝐵

2

𝑅

1 − 𝑅+ 1

The bound may only be met for 𝑅 =1

1+𝑚 and

𝐵

𝑚∈ ℤ+ , m∈ ℤ+

𝑹 <𝟏

𝟏+𝑩:

κ ≥ 1 (trivial)

𝑇 ≥

𝐵 + 1

2

Page 19: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Desired reconstruction schedule

19

𝐦𝐁 Burst

erasure

𝐦

time

Page 20: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Construction with constant delay [Martinian et. al 2004]

20

𝒅𝒂𝒕𝒂 𝒑𝒂𝒓𝒊𝒕𝒚

𝐦 = 𝟐 , 𝐁 = 𝟑

R =𝟐

𝟑

Burst

erasure

time

𝑅 =𝑚

1 +𝑚

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

𝑖

P i = 𝑠0 𝑖 − 𝐵 + 𝑠1 𝑖 − 2𝐵

𝑠0 𝑠1

Page 21: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Martinian’s reconstruction schedule

21

𝐦𝐁 Burst

erasure

𝐁

Nothing decoded!

time

Page 22: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Desired reconstruction schedule

22

𝐦𝐁 Burst

erasure

𝐦

Encoder?

Page 23: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

P{ }

P{ }

P{ }

P{ }

Optimal average-delay construction (R>1/2)

23

𝒅𝒂𝒕𝒂 𝒑𝒂𝒓𝒊𝒕𝒚

Burst

erasure

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

𝑅 =𝑚

1 +𝑚

weren’t

erased

P i = 𝑠 𝑖 2 𝑖 − 𝑖 2 − 1 + 𝑠 𝑖 2 𝑖 − 𝑖 2 − 1 − 𝐵 +𝑠 𝑖 2 𝑖 − 𝑖 2 − 1 − 2𝐵

𝐦 = 𝟐 , 𝐁 = 𝟑

R =𝟐

𝟑

Page 24: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

25

Optimal average-delay

𝐦𝐁

κ =mB +m

2 T =

mB +m+ B − 1

2

𝐨𝐩𝐭𝐢𝐦𝐚𝐥!

for any m,B co-prime

Burst

erasure

Page 25: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

P{ }

P{ }

P{ }

P{ }

Decoding a different burst

26

𝒅𝒂𝒕𝒂 𝒑𝒂𝒓𝒊𝒕𝒚

Burst

erasure

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

Page 26: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

27

Burst phase shift

Phase shift (𝝓): 𝜙 = 𝑖 mod m

𝑖

B erased

packets

properly received

packets

properly received

packets

Page 27: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

28

Average decoding delay by phase shift

𝐦𝐁

𝐨𝐩𝐭𝐢𝐦𝐚𝐥

Burst

erasure

𝝓 = 𝟎

𝝓 = 𝟏

𝝓 = 𝟐

𝝓 = 𝟑

𝐰𝐨𝐫𝐬𝐭

Page 28: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

30

Comparison to known codes

Page 29: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

31

Optimal average delay for R<1/2

𝐅𝐨𝐫 𝐞𝐯𝐞𝐫𝐲 𝛟 T =B

2

1

𝑚+ 1

𝐨𝐩𝐭𝐢𝐦𝐚𝐥!

Page 30: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

Decoding order

32

MSB LSB

Page 31: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

33

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

P{ }

BAD decoding order

GOOD news: we know how to fix that!

Page 32: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

34

Conclusion

• We know how to get to optimal average delay

• Method: specify decoding schedule, find realizing encoder

• Open problem: can we have optimal delay for all phase shifts when R>0.5?

• Future work (a nice metric):

Codes with average delays growing gracefully with B

Add random erasures

Page 33: PowerPoint Presentation · burst-erasure channels 6 Any code with T (constant), B and R satisfies the lower bound: B [packets]– the burst length T [packets]– the constant delay

35

Thank you!!!