Dell PowerVault ME5 Series Administrator's Guide

Disk count per RAID level

The controllers allocate virtual volume storage in 4-MiB pages, which are referenced paged tables in memory. There is a sequential write performance penalty when RAID-5 or RAID-6 disk groups are used in a virtual pool and the stripe size of the disk group doesn't align well with the 4-MiB page.

• Example 1: Consider a RAID-5 disk group with five disks. The equivalent of four disks provide usable capacity, and the equivalent of one disk is used for parity (parity is distributed among disks). The four disks providing usable capacity are the data disks and the one disk providing parity is the parity disk. In reality, the parity is distributed among all the disks, but conceiving of it in this way helps with the example.
  NOTE The number of data disks is a power of two (2, 4, and 8). The controller will use a 512-KiB stripe unit size when the data disks are a power of two. This results in a 4-MiB page being evenly distributed across two stripes. This is good for performance. Even better would be a 9-disk RAID-5 disk group with 8 total data disks, allowing one page to fit in exactly one stripe.
• Example 2: Consider a RAID-5 disk group with six disks. The equivalent of five disks now provide usable capacity. Assume the controller again uses a stripe unit of 512-KiB. When a 4-MiB page is pushed to the disk group, one stripe will contain a full page, but the controller must read old data and old parity from two of the disks in combination with the new data in order to calculate new parity. This is known as a read-modify-write, and it is a performance killer with sequential workloads. In essence, every page push to a disk group would result in a read-modify-write.

To mitigate this issue, the controllers use a stripe unit of 64-KiB when a RAID-5 or RAID-6 disk group isn't created with a power-of-two data disks. This results in many more full-stripe writes, but at the cost of many more I/O transactions per disk to push the same 4-MiB page.

The following table shows recommended disk counts for RAID-6 and RAID-5 disk groups. Each entry specifies the total number of disks and the equivalent numbers of data and parity disks in the disk group. Note that parity is actually distributed among all the disks.

To ensure best performance with sequential workloads and RAID-5 and RAID-6 disk groups, use a power-of-two data disks.