Re: Optimizing Informix IO

Folks, Here's some info about this situation. Ray Garrison <apsnet@ix.netcom.com> wrote in article <6gdq0q$58q@sjx-ixn6.ix.netcom.com>... > In article <6gde5t$g4t$1@news.xmission.com>, > "Narkinsky, Pat" <Pat.Narkinsky@rivhs.com> wrote: > > > > >System Manifest: > >INFORMIX-OnLine Version 7.14.UD1 > >1) Have no upgrade path for hardware; we are at the top of the line > >now. > >2) Are still experiencing 0% Idle Time (according to sar). You are (based on the next statement) waiting on I/O. You could verify that you are having long Q lengths with onstat -g ioq. Long (depends on the system, but GT 100 is prob bad) > >3) Have to add approximately 100 users in the next six months. > >Our problems are clearly IO related: examination of the output of sar > >indicate we are spending approximately 80% of our time waiting for IO. > > If your system is mostly (or all OLTP), then you need to tun it is such. To determine if it has kaio running, do: onstat -g ath and check for "kaio" threads. If you see em, they are running. If not, then you'll need more AIO VP's. How many, is the question. I suggest that you set up as best a test env. that you can (with a very repeatable test) and increase the amount of VP's. But that's not all. Other stuff to consider ( for this OLTP system?) is the setting for the LRU parameters. The system defaults are 8 LRU's and LRU_MAX_DIRTY=60 and LRU_MIN_DIRTY=50. I have seen cases in an OLTP system where addding AIO VPS (say 16), and more LRU's (16) and smaller LRU_MAX_DIRTY (10) and MIN_DIRTY(5) helped. I have even seen MIN and MAX set lower, but these numbers are a good first step. > >First question: we are running with NUMAIOVPS equal to 1. Could this be > >aggravating our IO problems? How can I tell if we are running KAIO? See above. > > > >Second question: I've heard mention on this list to the effect "Informix > >doesn't like RAID". Could someone clue me in on the considerations > >here? > > Informix likes raid just fine. It just so happens that they have a very effective partitioning scheme (what they call fragmentation) that allows you to locate data in specific spaces based on data values. RAID takes away some of that control. And remember the cost of RAID 5, between 15 and 20% on writes compared to RAID 0. > >Third: I've got 8GB of memory, however I am only using approximately > >1.5GB. Would it be possible/advisable to try to get the engine to use > >more of that memory? Will this make my checkpoints last too long? How long are your checkpoints now? If you are waiting on I/O, I doubt you can help matters by adding more memory. You need to tune the backend of the backend. > > > >Fourth: One of the busiest dbspaces on the system is "tempdb1" > >(according to 'onperf'). I have the following line in my onconfig file: > > DBSPACETEMP tempdb1:tempdb1:tempdb2 It shouldn't be listed twice. Informix is probably using it twice. It might not hurt to have more temp dbspaces. Another question to ask is why is this OLTP app using temp space. Can we help it out with an index? > >What is the significance of the double listing for tempdb1? Is this > >probably intentional, or is it the result of a typo on the part of my > >vendor? Would this account for the fact that tempdb1 is MUCH busier > >than tempdb2? Should I change it? Probably. > > > >Thanks! > > > >Patrick > > Question the First -> under HP/UX when I bounce the engine with KAIO > enabled, I get this written to the online log: > > DR: DRAUTO is 0 (Off) > HPUX Version B.11.0 -> Using select based KAIO > INFORMIX-OnLine Initialized -- Shared Memory Initialized > > Question the Second -> (again HP) we tried using an HP model 20 array > with RAID 5. Performance was sooooo slooooow we yelled at HP so loudly > they took back the array and credited full amount towards purchase of > a bunch of small disks. Informix much happier with 40 2Gig disks than > it was with one 80 Gig array. > > Question the Third -> I've never been able to figure out how adding more > memory speeds things up, as long as you're not swapping and have plenty > of buffers. I see no difference on our box between 1 Gig RAM and 2 Gig RAM. > > Question the Fourth -> Cool! Never thought of specifying the same temp > space twice on the DBSPACETEMP line. Am guessing Informix treats this as > it would three temp spaces, simply uses the same dbspace for the first > two temps. May be way to maximize leverage when have some fast and some > slow disks - specify each fast disk twice, each slow disk once. Have to > try this... > > Hope some of this is useful... > > Ray Garrison PaddyDD