Adaptive Resource Allocation in Tiered Storage Systems

Increased consolidation in virtualized datacenters and public clouds has raised the importance of allocating shared server resources fairly among multiple tenants. In the storage domain, tiered storage made up of heterogeneous memory and storage devices are now the norm in high-end systems. In this paper we consider a two-tiered storage system made up of SSDs and hard disks (HDs), and address some of the challenges in achieving both high utilization and fair allocation. Our work is complementary to fairness versus efficiency tradeoffs studied in the context of sequential versus random IOs (e.g. [6, 5]). For tiered storage, [1, 2, 3] emphasized fairness but did not explicitly consider system utilization. The storage system being considered is composed of SSDs and HD arrays. A client makes IO requests that may be served from either the SSD (a hit) or the HD (a miss), based on the hit ratio. The hit ratio will change with different application phases, but is relatively stable within a phase. The client hit ratio is monitored and used as an input to our resource allocator and scheduler. We illustrate the tradeoff between utilization and fairness with a simple example. Suppose the HD and SSD have throughputs of 100 IOPS and 1000 IOPS respectively. A fair scheduler assigns throughputs in proportion to the assigned weights of the clients. Suppose two continuously-backlogged clients 1 and 2 have equal weights, and hit ratios of h1 = 0.5 and h2 = 1.0. The access pattern using a fair scheduler is shown in Figure 1(a). The throughputs for the two clients are 200 IOPS each. The utilization of the disk is 100% but the SSD is only 30%. To fully utilize the SSD, the client weights are changed to 2:9. With this ratio, the throughput of client 1 is still 200 IOPS but 2’s throughput increases to 900 IOPS (see Figure 1(b)). While the relative allocations are no longer 1 : 1, the system throughput increases from 400 IOPS to 1100 IOPS. Furthermore, in this example, the throughput of client 1 is not reduced by the increased allocation to 2. In other cases, the allocation of a client may decrease significantly when the weights are changed to increase utilization. For instance, suppose the HD throughput is 200 IOPS, the clients had hit ratios h1 = 0.1 and h2 = 0.9, and equal weights. In this case, their throughputs under fair scheduling would be 200 IOPS each, but the SSD utilization is only 20%. Changing the ratio of the weights to 1 : 11 results in 100% utilization of both devices, but the allocation of client 1 falls to 100 IOPS while the other increases to 1100 IOPS.