Access Type

Open Access Dissertation

Date of Award

January 2017

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

First Advisor

Nabil Sarhan

Abstract

This dissertation presents an overall solution for interactive Near Video On Demand (NVOD) systems, where limited server and network resources prevent the system from servicing all customers’ requests. The interactive nature of recent workloads complicates matters further. Interactive requests require additional resources to be handled. This dissertation analyzes the system performance under a realistic workload using different stream merging techniques and scheduling policies. It considers a wide range of system

parameters and studies their impact on the waiting and blocking metrics. In order to improve waiting customers experience, we propose a new scheduling policy for waiting customers that is fairer and delivers a descent performance.

Blocking is a major issue in interactive NVOD systems and we propose a few techniques to minimize it. In particular, we study the maximum Interactive Stream (I-Stream) length (Threshold) that should be allowed in order to prevent a few requests from using the expensive I-Streams for a prolonged period of

time, which starves other requests from a chance of using this valuable resource. Using a reasonable I-Stream threshold proves very effective in improving blocking metrics. Moreover, we introduce an I-Stream provisioning policy to dynamically shift resources based on the system requirements at the time. The proposed policy proves to be highly effective in improving the overall system performance. To account for both average waiting time and average blocking time, we introduce a new metric (Aggregate Delay) .

We study the client-side cache management policy. We utilize the customer’s cache to service most interactive requests, which reduces the load on the server. We propose three purging algorithms to clear data when the cache gets full. Purge Oldest removes the oldest data in the cache, whereas Purge Furthest clears the furthest data from the client’s playback point. In contrast, Adaptive Purge tries to avoid purging any data that includes the customer’s playback point or the playback point of any stream that is being listened to by the client. Additionally, we study the impact of the purge block, which is the least amount of data to be cleared, on the system performance.

Finally, we study the effect of bookmarking on the system performance. A video segment that is searched and watched repeatedly is called a hotspot and is pointed to by a bookmark. We introduce three enhancements to effectively support bookmarking. Specifically, we propose a new purging algorithm to

avoid purging hotspot data if it is already cached. On top of that, we fetch hotspot data for customers not listening to any stream. Furthermore, we reserve multicast channels to fetch hotspot data.

Download

Share

COinS