DotSlash

1
DotSlash or how to deal with 15 minutes of fame

• Weibin Zhao
• Henning Schulzrinne
• Columbia University
• CATT/WICAT Annual Research Review
• November 14, 2003

2
Web Hotspots

• A sudden, dramatic surge of request rate
• Hard to predict e.g., external reference to web
  site
• Build up quickly
• Large magnitude
• Short-term overload
• Long delay or no service
• One-time event
• 15 minutes of fame
• Small web sites, e.g., on DSL or cable modem
• Thus, traditional solutions (Akamai, etc.) not
  economical

3
Existing Mechanisms

• Capacity planning
• ? high-end servers, LAN clusters, WAN mirrors
• Fixed capacity, cost
• CDNs
• Advance set-up, cost
• Only effective for large servers
• Caching
• Not fully controlled by the server
• Admission control
• No service for some clients

4
DotSlash Approach

• Enable a site to utilize spare capacity at other
  sites via dynamic collaboration ? Rescue
• Extending capacity dynamically as needed
• Usage models mainly for mutual aid, could be
  commercial
• Focusing on small sites
• Advantages
• Client transparent no browser changes
• Low cost based on existing systems, no
  additional hardware cost

5
Rescue Example-1
HTTP redirect Cache miss
www.origin.com 1.2.3.4
www.rescue.com 5.6.7.8
(8) Reverse proxy
origin.com
rescue.com
(9)
Dynamic DNS
Dynamic DNS
cache
(10)
(7)
(3)
(4) HTTP redirect alias1.rescue.com
origin.com DNS
rescue.com DNS
DNS RR
(5) alias1.rescue.com
Client1
(6) 5.6.7.8
(2) 1.2.3.4
(1) www.origin.com
6
Rescue Example-2
DNS RR Cache hit
www.origin.com 1.2.3.4
www.rescue.com 5.6.7.8
origin.com
rescue.com
Dynamic DNS
Dynamic DNS
cache
origin.com DNS
rescue.com DNS
DNS RR
Client2
(2) 5.6.7.8
(3)
(1) www.origin.com
(4)
7
DotSlash Functions

• Workload monitoring
• Rescue server discovery
• Enhanced SLP mSLP Select/Sort extensions
• Request distribution and redirect
• DNS RR HTTP redirect
• Dynamic replication of content
• Reverse caching proxy
• Rescue relationship management
• Allocate/release rescue servers

8
Basic Architecture

• Distributed system dynamic virtual hosting
• Distributed multiple servers for one site
• Virtual hosting one server for multiple sites
• Components
• Apache module (Mod_dots), DotSlash daemon
  (Dotsd), dynamic DNS (BIND), service discovery
  (enhanced SLP)

SHM
Other Dotsd
Apache
Dotsd
Mod_dots
TCP
HTTP
Client
SLP
DNS
BIND
mSLP
9
Server Status
Origin server Ask help from others
SOS
Allocate rescue servers
Release rescue servers
Regular server Otherwise
Normal
As a rescue server
Not as a rescue server
Rescue
Rescue server Provide help to others
10
Workload Monitoring

• Bottlenecks
• Network bandwidth
• CPU
• Memory
• Metrics
• Output data rate (HTTP)
• KBytes/second
• CPU load average
• Num concurrent conns
• MaxClients

Overload
OR Gate
CPU
Net
Mem
AND Gate
Normal
11
Origin Server

• Workload monitoring
• HTTP redirect
• Use rescue server aliases
• Dont redirect requests from rescue servers
• Redirect Policies
• WRR based on rescue server capacity (load)
• URL Hashing (URL locality) for large files
• DNS round-robin
• Add/remove rescue server IP addresses

12
Rescue Server

• Workload monitoring
• Reverse caching proxy
• Memory caching concurrency control among threads
• Check Host header
• A regular request
• A reverse proxy request to origin server (active
  rescue)
• An HTTP redirect to origin server (expired
  rescue)
• Dynamic DNS update
• Register new aliases

13
Rescue Server Management

• Allocation
• Initial allocation based on origin server load
• Additional allocation based on rescue server
  load
• Policy double current capacity
• Release
• When load is consistently low
• Release one rescue server at a time
• Partial release the min-capacity rescue server
• Final release the last rescue server

14
Other Components

• Testing tool
• Use httperf from HP Labs, and extend it to handle
  HTTP redirects automatically
• Dynamic DNS
• Use BIND 9.2.2, set up a testing domain
  dotslash.cs.columbia.edu, and only allow updates
  from specific hosts
• Service discovery
• Use SLP with mesh-enhancement (mSLP), and Select,
  Sort and Attribute List extensions

15
Testing Environment

• Apache
• 2.0.47
• With proxy and caching modules
• Enable caching and mod_dots
• OS
• Redhat 9, Linux kernel 2.4.20-20.9, hosts from
  Clic Lab
• Redhat 7.3, Linux kernel 2.4.19-6_planetlab,
  nodes from PlanetLab

16
Conclusion

• Preliminary version implemented
• Basic measurement and control flow
• Working on service location
• On-going work
• Performance measurement
• Apply to dynamic contents
• LAMP Linux Apache MySQL PHP