SLIDE 1
CDNs serve more than 50% of content
2
Origin Users CDN
2
Request / Response
Midgress-aware traffic provisioning for content delivery Aditya - - PowerPoint PPT Presentation
Midgress-aware traffic provisioning for content delivery Aditya - - PowerPoint PPT Presentation
Midgress-aware traffic provisioning for content delivery Aditya Sundarrajan, Mangesh Kasbekar, Ramesh K. Sitaraman, Samta Shukla CDNs serve more than 50% of content Midgress Egress Request / Response Origin Users CDN 2 2 Performance and
SLIDE 2
SLIDE 3
Performance and cost metrics
3
End-user latency Origin offload ratio Bandwidth cost Cache hit rate
SLIDE 4
4
100s of content providers 100s of 1000s of servers Millions of users
SLIDE 5
5
Traffic classes Requests Users
Traffic provisioning Cache management
SLIDE 6
6
Traffic classes Requests Users
Past work has focused on cache management Traffic provisioning Cache management
SLIDE 7
7
Traffic classes Requests Users
How can we assign traffic classes to reduce midgress? Traffic provisioning Cache management
SLIDE 8
8
S1 S2 Midgress +
x 1/2 x 1/2
+
100s of traffic assignment scenarios!
Traffic provisioning to reduce midgress
SLIDE 9
Traffic provisioning to minimize midgress
9
CDN Min. miss traffic to origin Origin Users Traffic classes Optimize traffic class assignment
SLIDE 10
Eviction age equality
10 10
Head Tail Eviction age Evict Insert
- 1
- 2
- 3
- 4
……
SLIDE 11
Footprint descriptors*
11
Stack distance Inter-arrival time Spatial locality: How many unique bytes are requested between successive requests of an object? Temporal locality: How often is an object requested? Joint probability distribution P(s,t)
* Footprint descriptors: Theory and practice of cache provisioning in a global CDN, A. Sundarrajan et al. in ACM CoNEXT 2017
SLIDE 12
12
Stack distance Inter-arrival time Joint probability distribution P(s,t)
Cache size Hit rate Eviction age Hit rate Eviction age Cache size
Hit rate = f (size) Hit rate = f (eviction age) Cache size = f (eviction age)
Caching properties from FDs
* Footprint descriptors: Theory and practice of cache provisioning in a global CDN, A. Sundarrajan et al. in ACM CoNEXT 2017
SLIDE 13
Traffic mixing using FD calculus
13
+
FD1 FD2 FD1+2 The addition operation is the convolution of joint pdfs which can be efficiently computed using FFT
* Footprint descriptors: Theory and practice of cache provisioning in a global CDN, A. Sundarrajan et al. in ACM CoNEXT 2017
SLIDE 14
Traffic provisioning to minimize midgress
14 14
CDN Min. miss traffic to origin Origin Users FD of traffic classes FD calculus to optimize traffic class assignment
SLIDE 15
2 N-1 1 N
λ1, λ2,…, λT T traffic classes N servers
- Min. ∑ij xij λj mj(cij)
Total miss traffic from cluster ∑j x1jλj … ∑j xNjλj
Traffic provisioning as an optimization problem
Estimate miss rate
- f traffic
mix using FD calculus
15
Cache size, C Traffic capacity, B
MILP – NP Hard!!
SLIDE 16
FD-based local search is faster than MILP
16
2 N-1 1 N
Servers … Predict midgress of traffic mix using FD calculus Traffic classes
- 1. Randomly assign traffic classes
SLIDE 17
FD-based local search is faster than MILP
17
2 N-1 1 N
- 2. Reassign traffic classes using
local search such that midgress is minimized Servers … Predict midgress of traffic mix using FD calculus Traffic classes
SLIDE 18
Metro-level traffic provisioning
18
Cluster 1 … Midgress of metro area Traffic classes Servers Cluster N … Servers …
SLIDE 19
Trace characteristics
19
Number of traffic classes 25 Length of trace 16 days Traffic types Web, media, download
SLIDE 20
Metro-level midgress reduced by 20%
20
10 20 30 40 50 60 100 200 300 400 500 600 Cache miss rate, % Cache size, TB OPT local search baseline fit
SLIDE 21
Traffic provisioning in partitioned caches
21
10 20 30 40 50 60 100 200 300 400 500 600 Cache miss rate, % Cache size, TB OPT-share baseline fit-share OPT-part baseline fit-part OPT baseline fit OPT – part baseline fit – part
SLIDE 22
Conclusions
22
Midgress-aware traffic provisioning reduced midgress by almost 20% in metro area Midgress-aware heuristic performs within 1.1% of OPT but is much faster Midgress-aware traffic provisioning can be extended to work with additional constraints such as minimum redundancy and maximum midgress, any cache management algorithm, and with partitioned caches
SLIDE 23
Th Thank k you!
23