age-based cooperative caching in information-centric networks

Age-based Cooperative Caching

Age-based Cooperative Cachingin Information-Centric Networks

Junxiao ShiApr 17, 2012

Contents

Background Age-based Cooperative Caching scheme Simulation Discussion

Background

In-network Caching

One of the most important commonalities of ICN designs is the universal caching.

In-network caching would improve performance.● Network delay and publisher load can be reduced

by having content requests satisfied at network edge as much as possible.

● Popular contents should be cached near the users.

CDN

In IP network, CDN is widely deployed to deliver HTTP contents.● Content is cached for time specified in Cache-

Control header's max-age field.

NDN

Every NDN router can cache contents, like a CDN node.● CCNx protocol defines the FreshnessSeconds field,

which serves the same purpose as max-age.

Cache Replacement

There isn't enough cache storage to keep every piece of content until they expire.

Cache is full, new content is coming, should I:● Remove some old content, and cache this one?● Don't cache this one?

Ideal Choice

It is desirable to distribute contents hierarchically in the network.● Most popular contents at the edge.● Less popular contents near the edge.● Least popular contents even further.

In this way, the aggregated cache hit rate of the network will be maximized.

Least Recently Used

Least Recently Used (LRU)● Keep track of “last

used” time for each cached content

● Content with earliest “last used” time is replaced by new content

Problem with LRU

Age-based Cooperative Caching

The Scheme

Dynamically configure content's age in an implicit cooperation manner among ICN routers.● Each replica has an age.● The replica obtains its age when it is added into the

cache.● The replica is removed from the cache when the

age expires.● Routers implicitly collaborate by modifying the age.

The Scheme

Age obeys the following rules:● The closer a replica is to the network edge, the

longer age it has.● The more popular a replica is, the longer age it has.

How the scheme works?

Age-based cooperation

Age decision

N contents C1,C

2,...,C

N with associated

popularity P1,P

2,...,P

N

The popularity weights can be formulated as

● Total popularity weights of all contents in the system is 1.

Popularity weights

weight i=Pi

∑ j=1,. .. , NP j

A Case Study

LRU, least recently used

ABC, age-based cooperative caching

Case Study Results

C1 delay 2 6 2 4

C2 delay 6 2 4 2

Average delay 4 4 3 3t si

t ci2

t ci1

DABC=∑i=1

4

t si /4=7

DLRU=∑i=1

2

t si /2=8with better age increment strategy

DABC '=∑i=3

4

t si /2=6

Simulation

CERNET2 topology

wkdlife.comserver

Trace information

One-month trace in 2011

● 76414 requests sent by 36432 users from 20 cities 8369 distinct URLs, access count distribution follows the Pareto

Principle

Subscribers in Beijing contribute a majority of the population

Top 25 URLs

Simulation methodology

Each user generates content requests according to a Poisson process of intensity 0.1 req/s

Caching schemes: ABC, LRU, FIFO

Router cache capacity: 100M

Size of content: 1M

Link delay: 10ms per hop

ABC: BASE_AGE=10 sec, MAX_AGE=60 sec

Result: aggregated network delay

∑i=1

UserCount

DelayOfUser i

UserCount

Result: end user delay

∑i=1

SimTime

∑j=1

rc i

DelayOfRequest j /rci

SimTime

Result: publisher load

Summary

There are significant gains through using ABC. The users will benefit from ABC as the total

download latency will be lowered. Content providers will be able to greatly reduce

the traffic to save bandwidth and computing resources.

Discussion

Consistency

CDN refresh API● POST http://ccms.chinacache.com/index.jsp

with a list of URLs to remove from caches

Internet contents are mostly static● Streaming service: contents will remain static during

their lifetime● Non-streaming service: a large portion of contents

do not change with time● Time critical applications: set a short, non-

increasing age

Paper shortcomings

Popularity weights is not known and cannot be accurately estimated, by network or publisher.

Case study and simulation is limited to one publisher and one user distribution.