Footprint Descriptors: Cache Provisioning in a Global CDN Aditya - - PowerPoint PPT Presentation

ā–¶
footprint descriptors cache provisioning in a global cdn
SMART_READER_LITE
LIVE PREVIEW

Footprint Descriptors: Cache Provisioning in a Global CDN Aditya - - PowerPoint PPT Presentation

Aug 24, 2023 241 likes ā€¢547 views

Footprint Descriptors: Cache Provisioning in a Global CDN Aditya Sundarrajan * , Mingdong Feng + , Mangesh Kasbekar + , Ramesh K. Sitaraman *+ * University of Massachusetts Amherst + Akamai Technologies Cache hits reduce end-user latency and

slide-1
SLIDE 1

Footprint Descriptors: Cache Provisioning in a Global CDN

*University of Massachusetts Amherst +Akamai Technologies

Aditya Sundarrajan*, Mingdong Feng+, Mangesh Kasbekar+, Ramesh K. Sitaraman*+

slide-2
SLIDE 2

Cache hits reduce end-user latency and bandwidth cost

2

Origin Users CDN

2

Request / Response

slide-3
SLIDE 3

Cache provisioning

3

S1 S2 Cache hit rate Traffic class 1 Traffic class 2 Traffic class 3

?

slide-4
SLIDE 4

Cache provisioning

4

S1 S2 Cache hit rate +

x 1/2 x 1/2

+

slide-5
SLIDE 5

5

S1 S2 Cache hit rate +

x 1/2 x 1/2

+

Tr Traffic classes are diverse

100s of traffic classes!

slide-6
SLIDE 6

Object ct po popul pulari rity ty di distri tributi bution

6

Traffic classes have different popularity skews

slide-7
SLIDE 7

Object ct siz size di distri tributi bution

7

Object sizes vary widely across traffic classes

slide-8
SLIDE 8

8

S1 S2 Cache hit rate +

x 1/2 x 1/2

+

100s of traffic assignment scenarios!

Tr Trace-ba based ed simul ulations ns are e pr prohi hibi bitivel ely expens xpensive

slide-9
SLIDE 9

How can we model traffic classes effectively?

9

slide-10
SLIDE 10

Footprint descriptors ā€“ space-time representation

  • f traffic classes

10

Stack distance Inter-arrival time Spatial locality: How many unique bytes are accessed between successive requests of an object? Temporal locality: How often is an object requested? Joint probability distribution P(s,t)

slide-11
SLIDE 11

Footprint descriptors yield caching characteristics

11

Stack distance Inter-arrival time Joint probability distribution P(s,t)

Cache size Hit rate Time-to-live Hit rate Time-to-live Cache size

Hit rate = f (size) Hit rate = f (time) Cache size = f (Time-to- live)

slide-12
SLIDE 12

How to use footprint descriptors to address cache provisioning challenges?

12

slide-13
SLIDE 13

13

FD1 FD2 FD3

addition scaling

FDout

x 1/2

+ +

addition

x 1/2

+ +

S1 Cache hit rate

Footprint descriptor calculus

Cache size Hit rate

slide-14
SLIDE 14

14

FD1 FD2 FD3

addition scaling

FDout

x 1/2

+ +

addition

Space-time representation Frequency-domain representation FD1 FD2 FD3

addition scaling

FDout

x 1/2

+ +

addition

slide-15
SLIDE 15

15

FD1 FD2 FD3

convolutions scaling

FDout

x 1/2

* * convolutions

Space-time representation

slide-16
SLIDE 16

How do object-disjoint traffic classes šœ1 and šœ2 mix?

16

šœ 2 šœ 1+2 šœ 1 Window of duration t s unique bytes s1 unique bytes s - s1 unique bytes P1+2 (s|t) = P1(0|t) P2(s|t) + P1(1|t) P2(s - 1|t) + ā€¦ + P1(s|t) P2(0|t) = āˆ‘

P1(s1|t) P2(sāˆ’s1|t

2 2345

)

slide-17
SLIDE 17

17

+ P1(1|t) P2(s - 1|t) P1+2(s | t) = P1(s | t) * P2(s | t)

convolution

P1+2(s,t) = P1+2(s | t) P1+2(t)

by definition

SD IAT P1+2(s,t)

šœ 2 šœ 1+2 šœ 1 Window of duration t s unique bytes s1 unique bytes s - s1 unique bytes P1+2 (s|t) = P1(0|t) P2(s|t) + P1(1|t) P2(s - 1|t) + ā€¦ + P1(s|t) P2(0|t) = āˆ‘

P1(s1|t) P2(sāˆ’s1|t

2 2345

)

slide-18
SLIDE 18

18

FD1 FD2 FD3

convolutions scaling

FDout

x 1/2

* * convolutions

Frequency-domain representation FD1 FD2 FD3

products scaling

FDout

x 1/2

products Inverse Fast Fourier Transform Fast Fourier Transform

Space-time representation

slide-19
SLIDE 19

How accurate are footprint descriptors in practice?

19

slide-20
SLIDE 20

FD calculus is accurate in production setting

20

x x x

86.6 % (calculated) 84.1 % (production) 86.6 % (simulated)

slide-21
SLIDE 21

21

How are footprint descriptors used in production setting?

slide-22
SLIDE 22

22

FD1 FD2 FD3

addition scaling

FDout

x 1/2

+ +

addition

x 1/2

+ +

S1 Cache hit rate

Cache size Hit rate

slide-23
SLIDE 23

23

FD1 FD2 FD3

addition scaling

FDout

x 1/2

+ +

addition

x 1/2

+ +

S1 Cache hit rate

1. Compute FD - every few days

Cache size Hit rate

slide-24
SLIDE 24

Parallelizing footprint descriptor computation using calculus

24

Input trace šœ Sub trace šœ1 Sub trace šœ2 Sub trace šœ3 Sub trace šœ4 FD1 FD2 FD3 FD4

+

FD12

+

FD34

+

FD

For 800 million requests, 7hrs to 28 mins = 15x speedup!

slide-25
SLIDE 25

25

FD1 FD2 FD3

addition scaling

FDout

x 1/2

+ +

addition

x 1/2

+ +

S1 Cache hit rate

Cache size Hit rate

2. Compute traffic mix

  • very often
slide-26
SLIDE 26

Traffic mix evaluation service

26

Video Mix Download Download 5 Gbps Video 1 Gbps Download 91.1 % Video 47.9 % Mix 85.3 %

slide-27
SLIDE 27

Predict traffic mix at different volumes

27

Video Mix Download Download 1 Gbps Video 5 Gbps Download 68.4 % Video 53.3 % Mix 56.4 % 91.1 % 47.9 % 85.3 %

slide-28
SLIDE 28

Conclusions

28

  • Footprint descriptors model the spatial and temporal localities of traffic

classes

  • Footprint descriptor calculus has high accuracy in production setting
  • Currently used in production setting for evaluating traffic class mixes
slide-29
SLIDE 29

29

Thank you!