Mirroring parallel read

Can anyone confirm whether my understanding is correct in regard to Informix mirroring and parallel read performance. Informixs' implementation of parallel reads from mirrored DBspaces is thus: Data from the first half of the chunk is read from the primary DBspace and data from the second half of the chunk is read from the mirror. TIA. Colin ctaberma@gcc.dhl.com DHL -----------== Posted via Deja News, The Discussion Network ==---------- http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own

On Wed, 05 May 1999 08:58:05 GMT, ctaberma@gcc.dhl.com wrote: >Can anyone confirm whether my understanding is correct in regard to Informix >mirroring and parallel read performance. > >Informixs' implementation of parallel reads from mirrored DBspaces is >thus: > >Data from the first half of the chunk is read from the primary >DBspace and data from the second half of the chunk is read from >the mirror. Your understanding is indeed correct! Thus chunks only half full will see no perforamce benefit from their mirrors... Dave

ctaberma@gcc.dhl.com wrote: > > Can anyone confirm whether my understanding is correct in regard to Informix > mirroring and parallel read performance. > > Informixs' implementation of parallel reads from mirrored DBspaces is > thus: > > Data from the first half of the chunk is read from the primary > DBspace and data from the second half of the chunk is read from > the mirror. Correct. Informix calls this 'split-read' behavior and finds that it reduces the average seek time by 50% because both disks only have to seek over half of the chunks cylinders. This assumes a lot about how the OS and SCSI controller and the drive itself organize sectors but is mostly valid. Art S. Kagel

In article <3731B669.5BFE@bloomberg.net>, kagel@bloomberg.net wrote: > ctaberma@gcc.dhl.com wrote: > > > > Can anyone confirm whether my understanding is correct in regard to Informix > > mirroring and parallel read performance. > > > > Informixs' implementation of parallel reads from mirrored DBspaces is > > thus: > > > > Data from the first half of the chunk is read from the primary > > DBspace and data from the second half of the chunk is read from > > the mirror. > > Correct. Informix calls this 'split-read' behavior and finds that it > reduces the average seek time by 50% because both disks only have to > seek over half of the chunks cylinders. > > This assumes a lot about how the OS and SCSI controller and the drive > itself organize sectors but is mostly valid. > > Art S. Kagel > On our platform Siemens Nixdorf RM400 / RM600 & INFORMIX V7.12.UC1 / V7.30.UC3 I can see that mirroring means an almost equal distribution of reads and exactly the same number of writes on Primary and Mirror. This heavily reduces our I/O load, because we have a read:write ratio like 25:1. This is also true on dbspaces only filled by 25% (oncheck -pe doesn't show any significant 'holes') I think you should test this on your platform Best regards Dick -- Richard Kofler debis Systemhaus EDVg Vienna / Austria -----------== Posted via Deja News, The Discussion Network ==---------- http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own

In article <7gsiq4$7l4$1@nnrp1.deja.com>, richard_kofler@my-dejanews.com wrote: > In article <3731B669.5BFE@bloomberg.net>, > kagel@bloomberg.net wrote: > > ctaberma@gcc.dhl.com wrote: > > > > > > Can anyone confirm whether my understanding is correct in regard to Informix > > > mirroring and parallel read performance. > > > > > > Informixs' implementation of parallel reads from mirrored DBspaces is > > > thus: > > > > > > Data from the first half of the chunk is read from the primary > > > DBspace and data from the second half of the chunk is read from > > > the mirror. > > > > Correct. Informix calls this 'split-read' behavior and finds that it > > reduces the average seek time by 50% because both disks only have to > > seek over half of the chunks cylinders. > > > > This assumes a lot about how the OS and SCSI controller and the drive > > itself organize sectors but is mostly valid. > > > > Art S. Kagel > > > > On our platform > Siemens Nixdorf RM400 / RM600 & INFORMIX V7.12.UC1 / V7.30.UC3 > I can see that mirroring means an almost equal distribution of > reads and exactly the same number of writes on Primary and Mirror. > This heavily reduces our I/O load, because we have a read:write ratio > like 25:1. > This is also true on dbspaces only filled by 25% (oncheck -pe > doesn't show any significant 'holes') > I think you should test this on your platform > To clarify this for those who emailed me: I see an almost perfect read balancing having approx. 200 tables in one dbspace consisting of ONE chunk with 2 GB on the small machines. Only few tables are > 25000 pages (pagesize 2 KB). I can not tell if there would still be parallel reads if I had only one big table Best regards Dick -- Richard Kofler debis Systemhaus EDVg Vienna / Austria --== Sent via Deja.com http://www.deja.com/ ==-- ---Share what you know. Learn what you don't.---