Title: QoS in Web-Servers
Team Members: Nitin Gupta and Sandhya Gupta
Abstract:

All the webservers today consider the incoming connections on equal
ground i.e. they provide no means of differentiating the incoming
requests nor providing any special treatment to incoming
connections. As long as the server is lightly loaded there is not be a
large difffernce in the response time but as the server becomes
heavily loaded the response times may increase dramatically. There is
no doubt in the fact that web-sites (such as Amazon, Yahoo &c) would
like to give preferential treatment to their regular customers
(generating revenue) and provide faster response time to them over
other users who are merely browsing through their website.

We don't intend to deal with the problem of identification of these
preferred users but instead we will be dealing with the problem that
once these preferred users are identified how to provide them with
better quality of service. The former problem may be solved by
assuming that IP addresses of the customers remain same. The problems
that we identified in loaded servers which lead to lower response time
are

(1) The Accept queue does not distinguish between preferred requests
and ordinary request to be serviced. This causes the preferred
requests processing to be delayed. In case of loaded servers, these
queues may be quite large.

(2) There is a need to distinguish the preferred requests over
non-preferred ones for allocation of CPU time as well.

(3) The outgoing queue should prioritize sending the preferred
requests packets in the network. Presently, OS's such as Linux, have
only a single queue without any priority.

We intend to solve problems (1) & (3) by introducing priority queues
where the kernel would be able to de-multiplex packets based on their
IP addresses into various service classes. These classes would be user
defined. Problem (2) could be solved by having more processes and/or
higher priority processes servicing these requests.

We intend to study the impact of the suggested solutions above in
terms of response time and other related parameters.