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

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 bandwidth cost Request / Response Origin Users CDN 2 2

Cache provisioning Traffic class 1 S 1 ? Traffic class Cache 2 hit rate S 2 Traffic class 3 3

Cache provisioning S 1 + + Cache hit rate x 1/2 S 2 x 1/2 4

Traffic classes are diverse Tr S 1 + + Cache hit rate x 1/2 S 2 x 1/2 100s of traffic classes! 5

Traffic classes have different popularity skews Object ct po popul pulari rity ty di distri tributi bution 6

Object sizes vary widely across traffic classes Object ct siz size di distri tributi bution 7

Tr Trace-ba based ed simul ulations ns are e pr prohi hibi bitivel ely expens xpensive S 1 + + Cache hit rate x 1/2 S 2 x 1/2 100s of traffic assignment scenarios! 8

How can we model traffic classes effectively? 9

Footprint descriptors – space-time representation of traffic classes Stack distance Spatial locality : How many unique Joint probability bytes are accessed between distribution successive requests of an object? P(s,t) Inter-arrival time Temporal locality : How often is an object requested? 10

Footprint descriptors yield caching characteristics Stack distance Cache size Hit rate Joint probability distribution P(s,t) Time-to-live Cache size Hit rate = f (size) Cache size = f (Time-to- Inter-arrival time live) Hit rate Hit rate = f (time) Time-to-live 11

How to use footprint descriptors to address cache provisioning challenges? 12

FD 1 addition addition Hit rate + Footprint FD out + FD 2 descriptor x 1/2 calculus Cache size scaling FD 3 + Cache + hit rate x 1/2 S 1 13

Frequency-domain FD 1 addition representation addition + FD out + FD 2 x 1/2 scaling FD 3 FD 1 addition addition + FD out + FD 2 x 1/2 Space-time scaling representation FD 3 14

FD 1 convolutions convolutions * * FD out FD 2 x 1/2 Space-time scaling representation FD 3 15

How do object-disjoint traffic classes 𝜐 1 and 𝜐 2 mix? 𝜐 1+2 s unique bytes 𝜐 1 s 1 unique bytes 𝜐 2 s - s 1 unique bytes Window of duration t P 1+2 (s|t) = P 1 (0|t) P 2 (s|t) + P 1 (1|t) P 2 (s - 1|t) + … + P 1 (s|t) P 2 (0|t) = ∑ 2 P 1 (s 1 |t) P 2 (s−s 1 |t ) 2 3 45 16

𝜐 1+2 s unique bytes 𝜐 1 s 1 unique bytes 𝜐 2 s - s 1 unique bytes Window of duration t P 1+2 (s|t) = P 1 (0|t) P 2 (s|t) + P 1 (1|t) P 2 (s - 1|t) + … + P 1 (s|t) P 2 (0|t) = ∑ 2 P 1 (s 1 |t) P 2 (s−s 1 |t ) 2 3 45 convolution P 1+2 (s | t) = P 1 (s | t) * P 2 (s | t) P 1+2 (s,t) SD + P 1 (1|t) P 2 (s - 1|t) by definition P 1+2 (s,t) = P 1+2 (s | t) P 1+2 (t) IAT 17

Frequency-domain FD 1 products representation products FD out FD 2 Inverse x 1/2 Fast scaling Fourier FD 3 Transform Fast FD 1 convolutions Fourier convolutions Transform * * FD out FD 2 x 1/2 Space-time scaling representation FD 3 18

How accurate are footprint descriptors in practice? 19

FD calculus is accurate in production setting x 84.1 % x (production) 86.6 % (calculated) 86.6 % (simulated) x 20

How are footprint descriptors used in production setting? 21

FD 1 addition addition Hit rate + FD out + FD 2 x 1/2 Cache size scaling FD 3 + Cache + hit rate x 1/2 S 1 22

FD 1 addition addition Hit rate + 1. FD out + FD 2 Compute FD - x 1/2 every few days Cache size scaling FD 3 + Cache + hit rate x 1/2 S 1 23

Parallelizing footprint descriptor computation using calculus For 800 million requests, Sub trace 7hrs to 28 mins = 15x speedup! FD 1 𝜐 1 + Sub trace FD 12 FD 2 𝜐 2 Input trace 𝜐 + FD Sub trace FD 3 𝜐 3 FD 34 + Sub trace FD 4 𝜐 4 24

FD 1 addition addition Hit rate + 2. FD out + FD 2 Compute traffic mix x 1/2 - very often Cache size scaling FD 3 + Cache + hit rate x 1/2 S 1 25

Traffic mix evaluation service Download 5 Gbps Video 1 Gbps Download Video Mix Download 91.1 % Video 47.9 % Mix 85.3 % 26

Predict traffic mix at different volumes Download 1 Gbps Video 5 Gbps Download Video Mix Download 68.4 % 91.1 % Video 53.3 % 47.9 % Mix 56.4 % 85.3 % 27

Conclusions 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 • 28

Thank you! 29