The Effect of Combining Network and Server QoS Parameters on End-to-End Performance

Abstract: Application hosting is becoming a popular business, Application Service Providers (ASPs) need however to keep up with the increasing pace of the market. This implies that they have to provide infrastructure to an increasing number of clients, and at the same time give QoS guarantees to these clients. One solution for ASPs to both guarantee a certain service level (QoS) for their clients and keep expanding would be to have so many resources as to be able to provide more than the maximum aggregate need of their clients. This may turn out to be an expensive or even an impossible solution - sharing infrastructure between clients and offering some means of resource reservation, and using charging to insure that clients only reserve the resources they need, is an alternative. It is however not an easy problem to solve, particularly if the procedure of adding new clients is to be automated, and the resources dynamically allocated. The ICorpMaker framework being developed at the IBM Zurich Research Laboratory offers a solution to the above named problems. In the ICorpMaker framework dynamic resource allocation is achieved by letting clients modify the amount of resources allocated to them in a simple manner, requesting more or less resources than their current allotment. The difficulty in achieving the end-to-end performance the client desires, lies in the fact that it is not certain how modifying resource allocation at the network respectively server level will combine and affect the end-to-end performance experienced by the end users of the service. The aim of this thesis project was to study the correlation between different network and server QoS parameters and the resulting end-to-end performance by making measurements. The results obtained from these measurements give an answer to the question of how to change the network and server resource allocations, when a client's application does not perform in a satisfactory way and hence the client requests more resources. Certain optimizations for the resource (re)allocation were also suggested based on the results.