The data collected in the VOSAIC user access log files provides information on user access statistics and general user information such as IP address, access-time, access filename, total connection time and playback information. Since the workload is representative of an exclusive video storage repository, accuracy of data in predicting future access patterns is enhanced. However, in the VOSAIC logs, there is little information on jitter, response latencies to a user request and low level resource utilization. The study of the access logs indicates a large variation in request accesses. The study set consists of 150 video objects in the video server with 126 clients accessing these video objects over the 20 day period.
The graphs in Figure 2 depict some of the variations in user requests and connections in the sample VOSAIC data. From Figure 2, it is evident that while the system can support a limited number of concurrent sessions (approximately 8), the typical operating point is lower (about 4-5 concurrent requests). This may be caused by typical access patterns or by admission control techniques that limit the number of concurrent sessions based on the resource utilization of each session. Traditional HTTP servers do not require admission control due to short bursty transfers, but this is not true for continuous media requests. The admission controller component in the VOSAIC server estimates the resource requirement of an incoming request and admits the request only if sufficient resources are available. The ascending section of the graph indicates that the utilization of the system is based on the availability of requests. The descending section of the graph depicts the resource bottleneck, i.e. as the number of overlapping requests increase, resource utilizations are heavy leading to either a larger number of rejects or longer response times. The graphs in Figure 2 indicate variations in client accesses (few clients request very frequent accesses) and variation in the number of requests per day over the period observed.
Based on the variations and results obtained from the VOSAIC empirical studies, we classify video workloads into multiple classes. We classify video objects into long(L), medium(M) and short(S) objects based on the size of the video and the amount of storage occupied. Another level of classification is based on the degree of access to a video object. We define frequently accessed video objects to be hot(H), moderately accessed video objects to be warm(W) and infrequently accessed videos to be cold(C). Ensuring video QoS requires guarantees that cannot be supported with best-effort traffic. Hence, we classify service requests as guaranteed(G) vs. best-effort(B) requests and apply the results from the empirical studies to determine the guaranteed class. This categorization can be summarized into a few QoS classes that represent the degree(quality) of service that can be expected. The choice of values that determines boundaries among individual classes is based on empirical parameters, for example, those obtained from the VOSAIC system analysis. In the sample logs, out of 150 video objects, 8 have over 50 requests, giving a 5 percent ratio of hot videos in the system. Similarly, from the study conducted, we consider short videos to be those that last less than a minute and long videos to be ones that last longer than a minute. In general, the choice of boundary values for any system is dependent on the workload, user class and other parameters. Therefore, empirical studies of the system will indicate values for separation into QoS classes.