Can I expect informix I/O performance increase?
Hello, Our current setup: Informix 7.30UC7-2 on F50 (2 procs); 512 Mb RAM, 2 SSA Enhanced Raid adapters with F/W cache, 28 GB data in several databases which are put on four 9 GB disks , stripped _file system_ (64 K stripe size) and mirrored by Informix to another (the same) file system on another four 9 GB disks via second SSA adapter (btw, our experience shows that this is fasted than spreading file system over disks (min. size can be 16 Mb :-( ) and mirror it by AIX). I'd like to know your opinion if I can expect a considerable performance _increase_ (more than 10 %) from following: 1) put the databases on 2x8 SSA disks instead of 2x4; 2) put all databases on raw space with KAIO enabled (well, this is AIX. I guess it has very good file systems); 3) reduce stripe size to 16 K on raw partitions; Maybe you can suggest something else? thanks for your attention, Ivars Strazdins State Revenue Service system administrator -----------== Posted via Deja News, The Discussion Network ==---------- http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own
What is your definition of "performance increase"? What do the statistics for your disk service times show at present? My experience of systems, Sybase and Oracle as well as Informix, on adequately specified disk arrays has been that the cacheing has reduced disk service times so substantially (to <10ms) - presumably becuase actual physical I/O is so infrequent - that any further concentraing on this areas through "traditional" ideas of disk placement/configuration etc. is unrewarding. So I'd suggest that you gather stats on your current performance and identify clearly which areas you are seeking to improve, and what performance indicators (Arghhhh, I'm sounding like a management consultant!) you'll use to judge the effect. HTH Neil Truby Londis Holdings Hampton Hill, UK Ivars.Strazdins@dati.lv wrote in message <7fp97f$995$1@nnrp1.dejanews.com>... >Hello, > >Our current setup: > >Informix 7.30UC7-2 on F50 (2 procs); 512 Mb RAM, 2 SSA Enhanced Raid adapters >with F/W cache, 28 GB data in several databases which are put on four 9 GB >disks , stripped _file system_ (64 K stripe size) and mirrored by Informix to >another (the same) file system on another four 9 GB disks via second SSA >adapter (btw, our experience shows that this is fasted than spreading file >system over disks (min. size can be 16 Mb :-( ) and mirror it by AIX). > >I'd like to know your opinion if I can expect a considerable performance >_increase_ (more than 10 %) from following: > >1) put the databases on 2x8 SSA disks instead of 2x4; 2) put all databases on >raw space with KAIO enabled (well, this is AIX. I guess it has very good file >systems); 3) reduce stripe size to 16 K on raw partitions; > >Maybe you can suggest something else? > >thanks for your attention, >Ivars Strazdins >State Revenue Service >system administrator > >-----------== Posted via Deja News, The Discussion Network ==---------- >http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own
In article <7fpd15$t56$1@taliesin.netcom.net.uk>, IN GENERAL, I think it'll be found that using the raw logical volumes will speed up Informix. Most database apps seem to perform better with raw LVs than with file systems. I don't have any numbers, but (since I don't think he's using RAID) I don't think that putting more disks in the SSA loop will slow him down all that much. I found with Sybase that OS mirroring was at a minimum about 10% faster than Sybase mirroring -- but if you want to stripe the LVs, then you don't have much of a choice. Still, if the app is doing sequential writes and reads, then I'd say dump the file systems and go with raw LVs. That in and of itself (moving to raw LVs) should give a substantial performance increase. Again, this is a general, blanket statement, but that's been my experience. "Neil Truby" <ntruby@netcomuk.co.uk> wrote: > What is your definition of "performance increase"? > > What do the statistics for your disk service times show at present? My > experience of systems, Sybase and Oracle as well as Informix, on adequately > specified disk arrays has been that the cacheing has reduced disk service > times so substantially (to <10ms) - presumably becuase actual physical I/O > is so infrequent - that any further concentraing on this areas through > "traditional" ideas of disk placement/configuration etc. is unrewarding. > > So I'd suggest that you gather stats on your current performance and > identify clearly which areas you are seeking to improve, and what > performance indicators (Arghhhh, I'm sounding like a management consultant!) > you'll use to judge the effect. > > HTH > > Neil Truby > Londis Holdings > Hampton Hill, UK > > Ivars.Strazdins@dati.lv wrote in message > <7fp97f$995$1@nnrp1.dejanews.com>... > >Hello, > > > >Our current setup: > > > >Informix 7.30UC7-2 on F50 (2 procs); 512 Mb RAM, 2 SSA Enhanced Raid > adapters > >with F/W cache, 28 GB data in several databases which are put on four 9 GB > >disks , stripped _file system_ (64 K stripe size) and mirrored by Informix > to > >another (the same) file system on another four 9 GB disks via second SSA > >adapter (btw, our experience shows that this is fasted than spreading file > >system over disks (min. size can be 16 Mb :-( ) and mirror it by AIX). > > > >I'd like to know your opinion if I can expect a considerable performance > >_increase_ (more than 10 %) from following: > > > >1) put the databases on 2x8 SSA disks instead of 2x4; 2) put all databases > on > >raw space with KAIO enabled (well, this is AIX. I guess it has very good > file > >systems); 3) reduce stripe size to 16 K on raw partitions; > > > >Maybe you can suggest something else? > > > >thanks for your attention, > >Ivars Strazdins > >State Revenue Service > >system administrator > > > >-----------== Posted via Deja News, The Discussion Network ==---------- > >http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own > > -- Kevin R. Gee MS Student, University of Texas at Arlington, Arlington, TX Senior UNIX Administrator, Origin Technology in Business, Arlington, TX geek@cse.uta.edu -- My comments and opinions are MY OWN... -----------== Posted via Deja News, The Discussion Network ==---------- http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own
Ivars.Strazdins@dati.lv wrote: > > Hello, > > Our current setup: > > Informix 7.30UC7-2 on F50 (2 procs); 512 Mb RAM, 2 SSA Enhanced Raid adapters > with F/W cache, 28 GB data in several databases which are put on four 9 GB > disks , stripped _file system_ (64 K stripe size) and mirrored by Informix to > another (the same) file system on another four 9 GB disks via second SSA > adapter (btw, our experience shows that this is fasted than spreading file > system over disks (min. size can be 16 Mb :-( ) and mirror it by AIX). > > I'd like to know your opinion if I can expect a considerable performance > _increase_ (more than 10 %) from following: > > 1) put the databases on 2x8 SSA disks instead of 2x4; 2) put all databases on > raw space with KAIO enabled (well, this is AIX. I guess it has very good file > systems); 3) reduce stripe size to 16 K on raw partitions; > > Maybe you can suggest something else? Nope, what you are considering is ALMOST EXACTLY what I'd recommend. You are going RAW which removed the overhead filesystems add by copying to the OS buffer cache before writing to the controller (yes even in O_SYNC mode which Informix does use) (15-25% gain). You are reducing the stripe size from 64K to 16K (good this will reduce drive contention and free your spindles faster) (5-10% gain). You are doubling the number of drives which will improve throughput under peak load only (but then by as much as 80%) as long as the controllers are wide enough to be able to receive enough data per second to drive the array at peak performance (depends on spindle speed * average and maximum track size)! The only thing better would be if you could get RAID 10 firmware instead of having to use Informix mirror. Remember by mirroring stripe sets (essentially implementing RAID01 or RAID 0+1) you will have to rebuild the entire stripe set if any drive fails, during rebuild you will be at risk for catastrophic failure if you loose any one of the eight drives on the other side of the mirror, and rebuilding the mirror will cost you as much as 80% of available array throughput. Compare this to striping mirrored pairs (or RAID10): You only rebuild one drive from mirror if a drive fails so you are only at risk for that one drive's mirror to fail which is eight times less likely and only the throughput of that one drive pair is impaired 80% during the rebuild leaving 90% of total throughput capacity available (for an 8 pair stripe, 80% for a 4 pair stripe). Normal performance of RAID01 and RAID10 should be about the same except that firmware mirroring, including RAID10, tends to be enough faster than Informix mirror (which OS mirror is not) that it more than makes up for Informix's more intelligent use of the mirrors and so will be a bit faster. Art S. Kagel