ISM Book Chapter 3: RAID 5

I like seeing RAID5 used for the tablespace files of a database or email attachments;

It's good for medium sized data that is accessed frequently but not constantly. The distributed parity of RAID5 helps distribute the workload of parity calculations, and you only lose 25% of your disk space in a 4 disk RAID set. RAID6 can also be used in place  of RAID5, but you get a little more protection against data loss.

As compared to RAID1 or RAID10 where small files that are constantly accessed (like log files for database or OLTP applications), but unfortunately you lose half your raw disk space to the mirror.

RAID3 is an interesting little creature in that it uses a dedicated parity disk, which supply's great read speed for large files, (drawings, pictures, movies, ect..) but has a difficulty when it comes to random small IOs , and rebuild time after failures is high.

Hi,

Yes, agreeing with everything Joe has said. RAID-5 configured correctly can be perfect for many random IO profiles, especially when implemented on a dedicated hardware RAID array (read RAID-5 engine - optimized for RAID-5 performance and RAID-5 algorithms).

In addition, there is an interesting feature of classic RAID-3 which is not always mentioned (not sure, without checking, if the ISM book references this). Several classic RAID-3 implementations (as well as having a dedicated parity disk), also have a hardware feature where the disks that are members of the RAID-3 stripe are hard jumpered so that the spindles are synchronized together (this is a drive to drive signal using a pin on the drive interface). Spindle sync reduces the latency in switching from one drive to another - which is perfect for RAID-3 and sequential I/O profiles). With the spindles synchronized, all the read-write heads across the disks can be accessing the same logical cylinder number for their respective drive. When you attend ISM you learn that the drive is formatted down a cylinder, before moving to the next track, and then formatting down that cylinder - this way of formatting improves the sequential performance of the drive - for sequential access to the drive we first read or write down the cylinder (an electro-magnetic operation), before moving to the next track (an electro-mechanical operation). If the spindles are also synchronized together, then this further improves sequential access to the RAID-3 (imagine there are 4 disks in the RAID-3, when we access the RAID-3 logical volume in a sequential manner, then, in parallel, we are accessing down each of the 4 cylinders, before the electro-mechanical operation of moving all the heads to the next track). Classic RAID-3, which is less commonly needed, can absolutely blast along for some sequential I/O profiles; so much so, that on some smaller modular arrays available in the industry (with maybe only a couple of GB of cache), cache can actually get in the way - if you fill up a small cache with sustained high performing sequential I/O, then you only run as fast as the cache 'drain rate'.

As Joe has said, RAID-3 performs very badly for random I/O (randon I/O is probably the more common commercial I/O profile).

Raid 5 parity is distributed across all disks so it overcomes the write bottleneck hence it can be used in applications like messaging, rational database management system implementation & data mining