Comparing RAID 5 and RAID 0+1

When you plan your Microsoft Exchange Server system's storage layout, plan on sacrificing some storage to RAID fault tolerance. For RAID 1 and RAID 0+1 volumes, you can expect to give up 50 percent of your storage for fault tolerance. However, for RAID 5 volumes, the percentage of disk space that you sacrifice for fault tolerance equals just one divided by n, where n is the total number of hard disk drives (i.e., spindles) in the array. For example, in a set of five 18GB hard disks, you give up one hard disk, leaving four hard disks (72GB) for storage.

Two types of RAID 0+1 exist. In one type, the data is striped first, then mirrored. In the other type, the data is mirrored, then striped. The latter configuration is slightly more fault tolerant because it can survive a multiple hard disk failure, as long as the failed disks aren't members of the same mirror set. Figure A illustrates how RAID 5 volumes (i.e., stripe sets with parity) and the two types of RAID 0+1 volumes store data.

Although RAID 5 requires a smaller investment in hard disks, RAID 5 can reduce availability in heavy-load situations. The problem with RAID 5 is that to retain fault tolerance, updates to any disk in a stripe set also require an update to the parity information. To update the parity information, the RAID software must read all other members of the stripe set, and those disk-reads take time. Because RAID 1 and RAID 0+1 maintain fault tolerance through mirroring, these approaches don't have parity update problems.

Surprisingly, in large Exchange Server systems, high-performance RAID 0+1 volumes can use fewer hard disks than RAID 5 volumes. How is this possible? In heavy-read situations in which mailboxes are large (e.g., more than 50MB per user), RAID 0+1 outperforms RAID 5 in I/O. To calculate the number of hard disks your Exchange 2000 Server databases require for RAID 5 and RAID 0+1 volumes, download Compaq's Solution Sizers' Exchange 2000 Storage Planning Calculator at http://www.compaq.com/activeanswers.

End of Article