Re: Must we use disk-striping
Posted in 1998
Rob Prop wrote: > > Hi, > > I have a serious question on which I've already got a lot of different > answers. > > When installing/ordering a new machine for IDS should you use the following > disk-architectures: > + Mirroring > Pos: HA + both disks can be used for reading > Neg: All the data should be written two times Yes but most mirror systems perform the two writes in parallel so there is not time cost. > + Striping > Pos: Data load is balanced over a set of disks > Neg: You don't know which data is on which disk so you can't use > fragmantation optimal True and I second Dave K's comment about safety and administration. > + RAID > Pos: Disk reads can be faster (Many READ boxes have caching) and you > have HA > Neg: Writes are slower I assume from your comments that you mean RAID 5. Actually plain striped drives have slightly better read performance than an equivalent RAID 5 setup, but not noticably. Neg: RAID 5 is prone to propagating failing disk sectors to the corresponding parity sector. No read ahead support since one of the optimizations of RAID 5 to better support random I/O is to only read the one drive containing the requested data. Recovery involves reading all drives in the RAID set so that all drives are busy and online performance is reduced by as much as 80% during recovery. You do not mention RAID 3 which is considered a good option for database servers (see postings last two weeks). You also do not mention RAID 10 (also known as RAID 1+0) which is a stripe of N mirrored pairs. This gives you the increased read performance of mirroring, striping supports Informix READAHEAD better than plain mirrored pairs or RAID, write performance does not suffer as in RAID 5, recovery requires only one drive to be actively read to rebuild the replaced mirror drive so performance during recover is reduced 1/N. Overall RAID 10 gives the best performance and supplies the best safety of ANY drive configuration. Only Neg: $$ it is expensive. How much are stellar performance and nearly perfect safety worth? Art S. Kagel