ADAPTIVE ADMISSION CONTROL FOR OVERLOADED WEB SERVERS

1
ADAPTIVE ADMISSION CONTROL FOR OVERLOADED WEB
SERVERS
Aims
Response
Time

• Protecting web servers from overload. Overload
  can lead to high response times and low
  throughput (Fig. 2)
• Providing admission control and service
  differentiation under overload

requests
Front End
Load
Web Servers
Back ends
Fig. 2 Server Overload
Main notions
Fig. 1 Typical Web Server

• Kernel-based architecture (Fig. 3) for efficiency
  and scalability (Fig. 4)
• Admission control based on network-level (e.g.
  clients IP address) and application-level (e.g.
  requested URL) information

user space
kernel space
non-static HTTP reqs.
URLs,
Current issues
cookies
appl. level control parsing, static req.

• Overload control service differentiation
  modules
• rates
• policies
• rate adaptation

• Adapting rates using a PI-controller
• Definition of policies

accept,
reject
HTTP reqs.
TCP SYNs
TCP/IP stack connection req. control
Expected Results
drop,
accept
Architecture for predictable overload control and
service differentiation of web servers
Conn reqs.
rate policed client (reqs/sec)
Fig. 3 Kernel-based Architecture
Fig. 4 Kernel vs User Space Performance
Thiemo Voigt and Per Gunningberg (DoCS)
Sponsored by PAMP/ARTES/SSF