Storage arrays typically had one pipe in to them and one pipe going around them daisy chaining from disk drive to disk drive. Now a new breed of storage array is appearing, one using switching technology to provide direct links to drawers or to individual drives, even, in the future, to both.
The reason for doing this is to avoid the disadvantages of using a single shared pipe to transfer data within the array.
The Christmas tree light problem
Emulex' Brian Reed, VP business development, characterises these disadvantages as the Christmas tree light problem. When one bulb goes out on a household's Christmas tree then all the bulbs go out. You have to find out why the fault happened, locate the failing bulb and replace it before the whole set of bulbs light up again. There's no way of locating the failed bulb except by starting at one end of the wire and seeing if each bulb in turn is working okay.
So it is with shared pipe disk arrays. If one disk fails then the whole array can fail.
Locating the failed drive can be time consuming. Arrays can have fifty or so drives in them.
It gets worse. If several requests come into the array at once then the shared pipe inside the array has to be used by each of the responding drives. The ones that get the pipe after the first respond less quickly. The array's latency increases as the number of requests to individual drives increase.
Storage arrays generally have a RAID controller sitting between the pipe to the outside world and the drawers of disks. There can be up to ten or so drawers and each one has its set of hard drives. The shared pipe inside the array is like a single large loop.
Switching technology puts a switch between the RAID controller and the individual drives or between the RAID controller and the drawers. In the first case the array becomes an SBOD or switched bunch of disks, rather than a JBOD - just a bunch of disks.
In the second case the switch is known as a root switch in Emulex terminology. The main reason for doing either is reliability. With switching the failure of one switched component does not affect the others. Working switched drawers carry on working when one drawer fails. One failed SBOD drive does not prejudice the whole array.
The switches in an SBOD can monitor the I/O to/from a drive and watch its state. Variations can indicate a failing drive. It's not necessary to access the drive's SMART data to do this.
The failed drive can be located more quickly as well.
A switched array also responds faster as the requests mount up. Each request occupies its own internal link from drive to RAID Controller in the SBOD case. Root switched arrays also respond faster as each drawer has its own link. A switched array of fifty drives will respond twice as fast as a RAID controller/JBOD set of fifty drives.
Array vendors are now designing arrays with switching at both drive and dawer level. These can be expected to respond up to four times as fast as a JBOD/RAID controller combination.
Switched arrays are better performers than JBODs and more reliable. Total cost of ownership is much better with switched arrays and they ship more data per time period. In a couple of years time we can expect virtually every storage array to be switched.
Oh, and tape libraries have a similar problem with tape drives sharing a common link inside the library. The answer is to switch them, and Reed calls this arrangment an SBOT, meaning a switched bunch of tapes.