Fig. 1. Variable bit-rate video streams are stored across multiple disks of the media server. Multiple clients can connect and start playback sessions via a high-speed network.

Fig. 2. The Shared-buffer Smoothing algorithm generates majorization minimal disk sequence with the disk bandwidth proportion limiting the server buffer proportion.

Fig. 3. We depict the disk time and buffer space proportion in the system (a) before, and (b) after applying Shared-buffer Smoothing. The peak of disk time proportion drops from 11.5% to 8.7%, while the peak of buffer space proportion increases from less than 1% to 8.7%. We assume the disk parameters of Seagate Cheetah and server buffer of 256 MB per disk.

Fig. 4. Using mixed workload, the sustained number of active streams increases almost linearly with the number of disks. Although all the submitted streams are accepted at system load 50%, at system load 90% Shared-buffer Smoothing increases the number of active streams by about 10%. This benefit is maintained across different numbers of disks. Note that the lines of unsmoothed and smoothed streams overlap at load 50%.

Fig. 5. The advantage of Shared-buffer Smoothing when combined with Variable-Grain Striping can exceed 15% (Action) depending on the stream type. We set the load to 90% on 16 disks, and assumed buffer space 256 MB per disk in the system.

Fig. 6. With 16 disks and 90% system load, the average disk time reserved each round increases from less than 80% to over 90% with Shared-buffer Smoothing and server buffer 256 MB per disk. The disk time, that we show for only one of the disks, was similar (typically within 2%) across the different disks of the array.

Fig. 7. With 64MB buffer space per disk, more than half of the total benefit of Shared-buffer Smoothing can be achieved (see also Fig. 5). Increasing the buffer space to 256 MB further improves the number of streams in Science Fiction and Action types although at a diminishing degree.

Fig. 8. Example of stream stored in an array consisting of Seagate and HP disks in alternating order. The low bandwidth of the HP model leads to disks busy over 40% during some rounds. When Shared-buffer Smoothing is applied, disk transfers are adjusted to smooth out the peaks and keep the maximum reservation below 13%. At the same time, the peak server buffer proportion is constrained to never exceed the peak disk time proportion.

Fig. 9. Using the mixed workload, the sustained number of active streams remains the same as the load is raised from 50 to 90% with plain Variable-Grain Striping. In comparison, with Shared-buffer Smoothing the number of streams increases by a factor of 2 at 50% and more than a factor of 3 at 90% load. We assume server buffer space of 256 MB per disk.

Fig. 10. Applying Shared-buffer Smoothing to different stream types, can lead to an increase in the number of accepted streams by more than a factor of 3. We assume load equal to 90%, and server buffer 256 MB per disk.

Fig. 11. The two leftmost bars of each stream show the average reserved disk time for the Seagate (STN) and HP (HPC) disks, assuming plain Variable-Grain Striping and 90% load. The lower transfer bandwidth of the HP disk creates a bottleneck keeping the reserved disk time of the Seagate disk to less than 25% of the round time. As is shown by the two rightmost bars though, with Shared-buffer Smoothing both disk types attain average disk time close to 90% of the round time.

Fig. 12. When we set the server buffer per disk to 64 MB, we obtain most of the benefit of Shared-buffer Smoothing (see also Fig. 10). Scaling the server buffer from 64 to 256 MB increases only marginally the sustained number of active streams.

Fig. 13. We create an array of four disks with Seagate (STN) and HP (HPC) models in alternating order. We show the average and maximum reserved and measured time for the first and second disk of the array with the mixed workload at 90% load. For the STN disks, the reserved statistics are no more than 8% higher than the measured. For the HPC disk, the corresponding difference can get up to 20%. We made the measurements using the detailed disk simulation models by Ganger et al.

Summary of symbols used in the descriptions of our resource reservations schemes

We used six MPEG-2 video streams of 30 min duration each

The coefficient of variation shown in the last column changes according to the content type.

Features of the Seagate SCSI disk assumed in our experiments

Features of the HP SCSI model that is included in the experiments for heterogeneous disk arrays