A Web-server farm is a specialized facility designed specifically for housing Web
servers catering to one or more Internet facing Web sites. In this dissertation, stochastic
dynamic programming technique is used to obtain the optimal admission control
policy with different classes of customers, and stochastic
uid-
ow models
are used to compute the performance measures in the network. The two types of
network traffic considered in this research are streaming (guaranteed bandwidth per
connection) and elastic (shares available bandwidth equally among connections).
We first obtain the optimal admission control policy using stochastic dynamic
programming, in which, based on the number of requests of each type being served,
a decision is made whether to allow or deny service to an incoming request. In
this subproblem, we consider a xed bandwidth capacity server, which allocates the
requested bandwidth to the streaming requests and divides all of the remaining bandwidth
equally among all of the elastic requests. The performance metric of interest in
this case will be the blocking probability of streaming traffic, which will be computed
in order to be able to provide Quality of Service (QoS) guarantees.
Next, we obtain bounds on the expected waiting time in the system for elastic
requests that enter the system. This will be done at the server level in such a way
that the total available bandwidth for the requests is constant. Trace data will be
converted to an ON-OFF source and
fluid-
flow models will be used for this analysis. The results are compared with both the mean waiting time obtained by simulating
real data, and the expected waiting time obtained using traditional queueing models.
Finally, we consider the network of servers and routers within the Web farm where
data from servers
flows and merges before getting transmitted to the requesting users
via the Internet. We compute the waiting time of the elastic requests at intermediate
and edge nodes by obtaining the distribution of the out
ow of the upstream node.
This out
ow distribution is obtained by using a methodology based on minimizing the
deviations from the constituent in
flows. This analysis also helps us to compute waiting
times at different bandwidth capacities, and hence obtain a suitable bandwidth to
promise or satisfy the QoS guarantees.
This research helps in obtaining performance measures for different traffic classes
at a Web-server farm so as to be able to promise or provide QoS guarantees; while at
the same time helping in utilizing the resources of the server farms efficiently, thereby
reducing the operational costs and increasing energy savings.
Identifer | oai:union.ndltd.org:tamu.edu/oai:repository.tamu.edu:1969.1/ETD-TAMU-2009-08-7081 |
Date | 2009 August 1900 |
Creators | Goel, Piyush |
Contributors | Gautam, Natarajan |
Source Sets | Texas A and M University |
Language | en_US |
Detected Language | English |
Type | Book, Thesis, Electronic Dissertation, text |
Format | application/pdf |
Page generated in 0.0017 seconds