Disks in the Internet

1
Disks in the Internet

• Dr. Eric A. Brewer
• Inktomi UC Berkeley

2
Internet Growth
3
Inktomi at a Glance

• Company Overview
• INKT on NASDAQ
• Founded 1996 out of UC Berkeley
• 180 Employees

• Applications
• Search Technology
• Network Cache
• Online Shopping

• Partners
• Traffic Server -- AOL, _at_Home, Digex, NTT,
  Telenor, ...
  
• Search -- Microsoft, Yahoo, CNET, Wired, NTT,
  ...
  
• Platform -- Sun, Digital, SGI, Intel

4
Search Engine
5
Global Architecture
Microsoft Start Page (98) Redmond, WA (eventuall
y world-wide)
Wired Digital HotBot San Francisco, CA
C-Net Snap! San Francisco
NTT goo Tokyo, Japan
Inktomi Engine166-Node Cluster
110M full-text docsSanta Clara, CA
Others coming...
N2H2 US Schools
100 NodesVienna, VA
Yahoo! ??
OzEmail Anzwers Sydney, Australia
6
Inktomi Architecture
Database

• Heterogeneous Cluster (Solaris Ultra and x86)
• Best cost/performance at time of purchase
• Database spread across the system(proportional
  to node performance)
  
• All nodes run same software and handle HTTP

7
Santa Clara Cluster
Sun

• No disks!
• Very uniform
• No monitors
• No people
• No cables

8
Network Cache
9
(No Transcript)
10
Traffic Server 2.0 Compelling Benefits
Faster Response

Savings
30Bandwidth Saved
Distributed Content
40-50 Cache Hit Rate
Flexibility
Scalability
Foundation for New Services and Incremental Rev
enue
Efficiency
Pay as You Grow
Surge Protection for Servers and Network
11
Scalability AOL Today

• Traffic Server deployed at AOL
• 13 million users
• 1 billion ops/day
• about 10000 ops/sec on average
• peaks about 50000 ops/sec
• Worlds largest caching system by far
• Can any cache keep up with Internet growth?

12
Traffic Server Throughout the Network
13
My disk is bigger than...
14
Disk Capacity Trends
Historical performance data from 720 disks.
Assumption Disk capacity will increase at
60/year,
with a baseline of 9GB in 1998
15
Disk Seek Trends
Historical performance data from 720 disks.
Assumption The seeks/second for one disk will
increase at 10/year, with a baseline of 80 seeks
/disk in 1997.
16
What I really want...

• Not more capacity. We already mostly use 4GB
  disks (to get enough seeks BW)
  
• Very small disk packs
• 6-16 disks (or platters with indep. heads)
• Maximize seeks BW per cubic inch
• RAID optional (!)
• The return of drums
• No moving heads, one head per track

17
The Bar Head

• CMOS bar that spans the disk
• Heads are etched in the bar (dont move)
• Read many tracks simultaneously
• write would be nice too, but less important
• could avoid writing adjacent tracks together
• Would reduce density (at first)
• heads probably farther from disk
• So what!!! Who cares just wait a year.
• Much better at small reads/writes and random I/O
• Longer lifetime and lower cost...

18
Inktomi Customers Partners
America OnLine

• 12 Million Users
• 30 of US Dial Up Web Traffic