Technology for Backbone Web Caching Peter B Danzig Network Appliance November 1998
Speakers Peter Danzig -- Network Appliance “Scaling Cache Deployments” Shirish Sathaye -- Alteon Networks “Scaling w/L4-Switches” James Aviani -- Cisco “Scaling WCCP” Bill Maggs -- MCI/Worldcomm “Routing and Caching” Ed Kern Digex “Report on Digex cache deployment”
Rules of thumb for Scaling Individual Cache Nodes Peter B Danzig Network Appliance November 1998
Scaling Individual Cache Nodes Rule of thumb resources required per megabit Simultaneous TCP Connections 500 Disk Needed 4 GB Disk Arms Needed 1/2-1/3 Memory 32 MB
Simultaneous Connections Little’s Law: N = W W: Average time to fetch a web page: 4-10 Seconds URLs/Second: 25 per megabit Number of TCP connections N = 2 * 25 * 10 = 500 connections/megabit
Disk Arms Keep disks below 50% busy A cache behind a cache sees random references Assume all references require 1 disk/ops Depending on OS, a single disk can achieve operations per second Per megabit: 25 ops/S * 2 (for head room) yields anywhere from 1 disk to 1/4 disk per megabit
Disk Space A megabit link can draw 9GB/day Peak to average load is about 4X So a 9GB disk holds the entire reference set of 4- days of operation. Assume 1/2 of references are already cached or are non-cachable…. Use 4GB of disk per megabit
Memory Each megabit means 500, dual 8 KB TCP buffers: 8 MB Proxy state for KB: 16 MB Disk buffer pool: 8MB Leads to 32MB/megabit
CPU Depends on the manufacturer and OS release For example: C720 running NetCache-3.3: 25 mbits/s C760 running NetCache-3.3: 37 mbits/s
Aggregating Cache Nodes Currently have two technologies for clustering cahe nodes L4 Switches Alteon, Arrowpoint, Foundry, and RND WCCP Cisco routers and WCCP licensees