The experience of video playback from user-generated video sharing sites suffers from interruptions and delays due to necessary buffering. The authors of this paper present the results of a study using prefetching when watching user-generated videos, which are generally rather short (three to ten minutes).

In a brief introductory section, the authors describe the background and motivation of their work as well as the basic idea of their approach; the major results are summarized. As a motivation in section 2, user experiences with watching YouTube videos are analyzed. The analysis is based on video download traces, which are investigated using the Wireshark network protocol analyzer. The analyzed data represent downloads from 12 different locations and access technologies. From each location, ten videos of different quality levels were downloaded. The presented results comprise: the number of interrupted playbacks, the number of pauses in the interrupted playbacks, and the amount of time that a user had to wait.

Motivated by section 2’s results, in section 3, the authors present their video prefetching scheme, where the prefetching agent (PA) can be located either at the client or at the proxy. In the first case, it receives requests only from one client. The PA stores prefetched prefixes of videos and can perform caching. To determine the set of videos to be prefetched, the authors propose two possible algorithms: one based on users’ search results, and one that uses the YouTube recommendation system. The advantage of both approaches is that they are simple and not computationally expensive.

In section 4, the authors describe the necessary data collection, which is composed of two phases: the data traffic between a campus network and YouTube servers, and the retrieval of two lists (search result and related video lists) from YouTube using the YouTube Data application programming interface (API). The process is described in detail.

Section 5 presents the evaluation of the video prefetching approaches, where the two selection algorithms and two settings (client versus proxy) are compared. To evaluate the approaches, the authors perform a trace-driven simulation based on actual user usage patterns. In the evaluation, two metrics are used: the hit ratio and precision. The first gives the fraction of requests that can be served from the prefetching storage, and the second gives the accuracy of the video selection algorithm (the fraction of the requested videos over the total number of prefetched videos). The experiments are performed both in the case when the PA always has sufficient storage space and when it has only limited storage space. In the sufficient storage case, the algorithm based on the related video lists in combination with the proxy setting gives the best hit ratio (up to almost 76 percent), which is also influenced by the fact that users in the same local network share similar interests. The authors also investigate how many requests from each referrer are hit when the related video lists are used for prefetching, both for the proxy and the client setting. The results show that there is overlap between the videos shown in the related video lists and videos requested through other referrers; a detailed analysis is given. This analysis also shows the advantage of using the related video lists for selecting the videos to be prefetched. The authors also present the combination of the prefetching approach with caching. The results are improved by five to 20 percent compared to the prefetch-only mode. Finally, the authors show that the storage required to achieve the highest hit ratio is within feasible range for a campus network.

In section 6, the authors discuss the influence of limited storage space and the choice of the number of videos to be prefetched from each related video list. Furthermore, the question of how large the prefetched prefix should be, as well as various aspects of the general feasibility of prefetching are discussed. The paper finishes with sections on related work, conclusions, and references.

The paper is well done, despite several typographical errors and some annoying inaccuracies.