4. Conclusion

In this study, we quantified the tradeoff associated with alternative multi-zone techniques when extended to a configuration consisting of heterogeneous disk drives. Ignoring OLT1, our principle result is that no single strategy dominates on all metrics, i.e., throughput, startup latency, cost per simultaneous display, and wasted disk space. All proposed techniques strive to distribute the load of a single display evenly across the available disk bandwidth in order to prevent formation of bottlenecks. They belong to a class of algorithm that is commonly termed non-partitioning. An alternative approach might have been to partition resources into clusters and treat each cluster as an independent server. For example, with a configuration consisting of 3 disk models, we could have constructed three servers and assigned objects to different servers with the objective to distribute the workload of the system as evenly as possible [7]. The system would replicate popular clips across multiple servers in order to prevent formation of bottlenecks [16, 6]. Using this approach, one could optimize system parameters (such as block size) for each configuration independently in order to maximize the performance of each subserver. This is ideal for static workloads that do not change overtime. However, for dynamic workloads, one must employ detective techniques that monitor the frequency of access to objects and replicate popular objects in order to prevent formation of bottlenecks. [30] utilizes a simulation model to show that this approach is generally inferior to a non-partitioning scheme. This is because detective techniques must wait for formation of bottlenecks prior to eliminating them [5].

As a future research direction, one may develop a configuration planner that consumes the requirements of an application and computes system parameters (along with a choice of a non-partitioning technique) that meets the application requirements.