Disruptions and delays in the retrieval of continuous motion video data streams from storage devices are unacceptable for video-on-demand (VOD) service. Concurrent VOD services require careful scheduling of accesses to the storage device in real-time. In this paper, we propose a dynamic approach to schedule the retrieval of motion ideo data in interactive VOD servers. Most of the previous VOD scheduling approaches use fixed disk read sizes and pre-determined reading orders. Run-time information is seldom used to fully exploit the potential capacity of the system. To maximize throughput, our approach relaxes admission control by observing run-time information. To improve performance under heavy transients, we introduce mechanisms to dynamically control the read sizes of disk accesses, to pipeline admission of new streams with departure of ongoing streams, and to safely deviate from seek-reducing service sequence to provide early initial service. Most important of all, our scheduler is designed to reduce start-up delay significantly. When the demand on real-time services decreases, our scheduler is also highly efficient in redirecting disk bandwidth to service sporadic requests. We evaluate our technique by comparing its performance with two static seek-reducing techniques via thorough experiments under different buffer capacities and load settings. Results of experiments show our technique's significant improvements in real-time service throughput, initial response time, and sporadic service throughput.