The Linux RAID-1, 4, 5 Code

We want to provide a RAID-1+ personality that basically would be RAID-1 plus check pointing; this will let us synchronize the disks at boot time after a disk crash. Unfortunately, PC computers are not shipped with an NVRAM device that would allow us to do this with a minimum amount of work.

The RAID-4 and RAID-5 code can be enhanced so it can determine when a complete stripe is being written to disk. When this is detected, the driver does not need to read any of the old information found on the disk. We can write all of the data sectors as they are found on the request queue, compute a new parity bit sector and add it to the queue as well. This is the preferred method for improving RAID-4 and RAID-5 disk write speed as discussed in “Striping in a RAID Level 5 Disk Array” (see Resources).

Currently, the low-level, request-queue and ordering code writes using the elevator algorithm. Preliminary work in this area shows that the algorithm can be improved in a per-device fashion by algorithms that are aware of the disk geometry. Virtualizing the ordering algorithm for the request queue in a per-device fashion may provide a good performance gain.

When there is a system crash, the RAID-4 and RAID-5 systems need to recompute the parity information on the disk array, since the parity blocks may contain old information. To improve this situation, we want to make use of non-volatile RAM (on those systems that have this kind of hardware) for check pointing the state of the disk array. With this feature, we can recompute the information for any blocks that may have incorrect information when starting up the disk array.

Resources