Archives of the TeradataForum

Message Posted: Thu, 15 Jan 2004 @ 15:46:57 GMT

	<Prev	Next>		<<First	<Prev	Next>	Last>>

Hi Phil,

What you have to consider is the I/O bandwidth, storage and system reliability.

The I/O bandwidth is a function of the disk and controller performance. For a read RAID 1 will have similar performance to no RAID as in both cases the read data can be read once. In theory the RAID 1 can be better as data that is on one disk with no RAID cab be read from two disks with RAID 1 this is dependent on the controller firm ware. For a write there the data has to be written twice to different disks. A well configured system will have sufficient I/O band with to keep the cpus busy so increasing it will not significantly increase performance except for work loads that are very I/O biased.

You will get twice as much storage without RAID.

For reliability there is no comparison. With no RAID if and disk fails you loose data, fallback can mitigate this. With RAID 1 a disk must fail and it mirror fail before it is replaced. Assume a mean time between failure for each disk of 4 years and a system with 100 disks with no RAID you will loose data on average every 15 days. For a system with100 disk pairs, the same MTBF, 1 day to replace failed disks RAID 1 will loose data on average every 30 years. This is no the whole story, there are controller and power supply failures to also consider so the figures will be worse than this.

A less drastic option is to consider RAID 5 which performs worse than RAID 1 is less reliable but give 50% more storage.

Fallback with RAID 1 will give improved reliability. It means that even if you loose a disk pair or a pair of controllers or a power you do not loose data. This is worth considering for key data areas as it can be used selectively, by default DBC is in fallback. It is also worth considering for very large systems as the number of disk grows the chances of a failure go up. A system with 1000 disk pairs would have a disk MTBF of 3 years.

The performance will be less as the data will now be written to two different disk pairs. The performance will be dependent on how much of the data is in fall back.

There will be less storage available but this more than 50% as spool and temporary tables are not in fallback.

Teradata have a tool that can calculate the change in throughput expressed in T-Perfs and storage in GB of uncompressed user data capacity. As an example 4 Node 5380 with 42 x 36GB disks per node:

0% FB, 32.26 T-Perf, 1,665GB
50% FB, 29.23 T-Perf , 1,227GB
100% FB, 26.19 T-Perf, 971GB

These figures are based on averages and assumptions but give an indication the effects.

I don't have figures for the effect of journaling.

These spring to mind:

* Disk Pairs

* Controller Pairs

* Power Supplies (2 or three depending on disk cabinets)

* Human error, eg removing wrong disk of failed pair or switching of cabinet

Daniel O'Hara
Project Manager
Teradata a Division of NCR

	<Prev	Next>		<<First	<Prev	Next>	Last>>