The only absolute requirement for matching drive specs in a RAID 10 array is that the drives must be of the same architecture, e.g., all SAS or SATA. Mixing and matching drives of differing performance levels will not compromise the ability of the drives to provide the RAID level, but the worst drive among them will lower the performance or capacity of the RAID set significantly. Two example scenarios are as follows.
Speed limitations: There are four drives in my RAID 10 set with data striped across two mirrored pairs. Three drives are 15,000 rpm and the remaining drive is 10,000 rpm. As data is written to the 10K/15K mirror pair, the 15,000 rpm drive will write the data faster than the 10,000 rpm drive, meaning that the 15,000 rpm drive has to constantly wait for the 10,000 rpm drive to catch up before it can send the acknowledgement back to the application and write the next data chunk.
In certain circumstances, the 15,000 rpm drive will have to wait so long that it misses the beginning of the next sector and has to do a full disk revolution before it can write the next data chunk. In a RAID 10 configuration, data logical unit numbers (LUNs) are striped over the RAID 1 mirrors, with data areas spanning all drives in the RAID set.
If an application was writing to all four drives at the same time, the 15K/15K mirror pair would have to wait for the 10K/15K pair to finish writing before carrying out the next write, leading to lower disk performance.
Size limitations: There are four drives in my RAID 10 set with data striped across two mirrored pairs. Three of the drives are 450 GB capacity and the remaining drive is 300 GB. In this scenario the usable data area on each drive is equivalent to the capacity of the smallest drive -- 300 GB in this case.
As RAID 10 uses only half the capacity of the entire RAID 10 set anyway, the usable capacity in this situation would be 600 GB from a raw capacity of 1,650 GB. If four 450 GB drives were utilized, the usable capacity in the RAID set would be 900 GB -- a considerable difference.
This was first published in August 2010