Proactive Web Caches

Large-Scale Active Middleware project

Joe Touch, Katia Obraczcka,
Amy Hughes, Anand Oswal

USC/ISI Computer Networks Division

Presented at the NLANR Web Cache Workshop
Boulder CO, June 1997

Introduction - Proaction

• Benefits
• - Faster than speed-of-light response latency
• - Efficient multicast without requiring long server queues
• Costs
• - Resources - BW, CPU, storage
• - Complexity - contention avoidance for BW, CPU

Proaction

Preliminary Results

• FTP (Infocom `94)
• - without proaction, per-item response averages 2.1 RTTs
• - with proaction 3x lower latency, 0.7 RTT avg. response
• - proaction costs 7x higher BW
• HTTP
• - with proaction, 83% hit within 100ms,
• - without proaction, 14% hit within 100 ms
• - proaction costs 5-8x higher BW

Motivation

Bandwidth too high for interactive...

...or no interactive at all

Unicast Vision

• Interactive response time is useful / critical
• - Everyone talks about the speed-of-light...
• - BW requirement is HUGE
• - `Can't get there from here' at all in some cases
• Idle BW is a terrible thing to waste
• - Exists now on modem, ISDN links
• - Need tagging or squelch tunnels for Internet-wide
• May be expensive, but even Porsche sells cars

Unicast design

Multicast Vision

• Hot-spots are important
• - Significant traffic (conjecture)
• - Important traffic, opportunity for interactive response
• HS's generate communities of interest (COI)
• - Groups of users associated with a group of data
• COI are dynamic
• - Time scale of hours-days-weeks (conjecture and goal)
• - E.g., tell a few friends, they'll hit, etc.
• Content dictates COI
• - Predicted by URLs for now

Tuner Protocol

Publisher / subscriber relationship

• Server
• - "TV-Guide" per server
• - Multiple channels per server dynamically `topic of interest'
• - Requests mcast if in a current `topic of interest'
• Cache
• - Partitioned per channel
• - Tune to TV-Guide of popular servers
• - Allocate partition per popular `topic of interest'
• - (assumes within one server, or labelled)
• - Automatically converges at network aggregation points

Multicast design

Issues

• Cache partitioning
• Item selection
• - Ordering / prioritization
• - At server, and replacement at client
• Group selection
• - At server and client
• Transport issues
• - Lazy multicast reliable transport
• - Background unicast reliable transport
• - Multicast parameter tuning (TTL)