Disk arrays represent a new way to construct high performance I/O systems, by replacing a few large formfactor drives by many small drives. Data is striped across the drives, with data redundancy introduced to improve media availability. A fundamental issue is how best to configure the array for performance, capacity, and availability. We examine three possible array organizations: no redundancy ("simplex" or RAID Level 0), 100% redundancy ("mirroring" or RAID Level 1), and parity redundancy (RAID Level 5). The mirrored approach favors performance over capacity, while the opposite is true for parity protected arrays. Since no single array organization clearly dominates the others, we present several alternative metrics by which to compare them. We argue that a way to simultaneously evaluate capacity and performance is needed, and propose a new storage system metric based on data temperature (IOs/second/GB) for this purpose. We illustrate its use by assessing several example array organizations.