Re: Disk Layout

In article <33EACDA4.1C3280EB@informix.com>, "Mark D. Stock" <mdstock@informix.com> writes >Steve Romankiw wrote: >> >> I have been unable to find information reguarding the pros & cons of >> RAID and ODS. >> >> We are running Sun Solaris 2.5.1 with VxVM and ODS 7.23 UC1. We >> planning for some expansive growth in our database(s). In our Sparc >> Storage Arrays (SSA), we are adding 4.1gb drives (largest drives >> supported for this model) to store our ODS database objects. In order >> to get the best possible performance, what have other Informix shops >> done. >> >> Some of the questions I have are: >> >> Q: Should I simply create two partitions per drive? This creates >> two >> 2gb partitions which the max chunk size for ODS. > >Not necessarily. You need to know the through-put of the disk. >For example, if one read process can read 500 Kb/Sec, and two processes >only read 400 Kb/Sec, you can actually achieve 800 Kb/Sec in total. Not neccessarily, you may well get the disk seeking back and forth between two "hotspots" instead. This cost of this seeking back and forth may mean it is cheaper to have one process which moves the heads once across the drive as it seeks. Depends upon throughput, layout of data and how costly disk seek are when >1 cylinder is involved. >Consider the following (admittedly rather linear) example: > > Procs KB/Sec Total > 1 500 500 > 2 400 800 > 3 300 900 > 4 200 800 Hear 4 is less than 3 as the drive is seeking more and so eventally the extra cost of disk seeks reduces the throughput. > >This example would indicate the optimum number of dbspaces to be three >in order to maximise through-put. > >This is a bad example, and not the way you think. I ran a similar test >on a disk in an old Sun box this week, and got a figure of about ten >dbspaces. We fragmented a table over just three dbspaces (yes, on the >same disk) and saw our benchmark script's performance increase by a >factor of three! Now that's linear scaleability. ;-) > Fragment indicies as much as possible, preferably by the key to reduce the amount of the index to be read if fragment elimination is used. >> Q: Should I employ RAID 0+1? Get the benefits of striping and >> mirroring? > >Consider mirroring for your critical dbspaces (rootdbs and any dbspace >that contains logs). Consider striping for small lookup tables that you >consider will not benefit from fragmentation. > >> Q: Is your decision different based upon OLTP vs DSS? > >Yes. :-) > >The main consideration is that OLTP will generally involve simple SQL >resulting in index reads and large transaction activity. Whereas, DSS >will generally involve complex SQL resulting in sequential scans and >little or no transaction activity. If you are fragmenting then remember >to define a strategy for your indexes as a separate exercise to your >data fragmentation strategy. > >Hope that helps, -- David Williams