LOGGING

CLEANERS*

1 per disk if < 20 disks
1 per 2 disks if 20 to 100
4 per disk if > 100
UPDATE at least one per LRU queue pair. This has most recently been best for me.
If –f indicates that all cleaners are active allocate more.

PHYSDBS

! rootdbs. Place on a separate spindle

PHYSFILE

= usrthreads * 5 (or size of most freq. blob) * 4

PHYSBUFF

= usrthrds * 5 * 4
UNLESS tblspace blobs in DB w/o logging then usrthrds * size of freq blob pg * 4

LOGFILES

> 3
In all my manuals I can find no good reference for performance tuning this parameter.
I have no experience adjusting this save for eliminating anomalous behavior.

LOGSIZE

> 200kb
small for greatest recovery, if tape slow, or blobpages volatile
(connects * maxrows (in one trans)) * 512

LOGBUFF

size of 3 LL buffers in RAM
determines freq of flushing to disk

LOGSMAX

LOGFILES + 3

CKPTINTVL

adjust interval of checkpoints, see LRU.
For data loading (except HPL express mode) and parallel index builds 3000
a large interval will allow size of physlog (i.e. amount of work accomplished) to determine when chkpts occure.
–F will show if writes are LRU driven or CHKPTINTVL driven
–l and –m will determine if checkpoint interval is driven by Physlog = 75% full or this param

LRUS*

max(4, NUMCPUVPS)
if -R shows #dirty pages > LRU_MAX add LRU Qs. If no change increase CLEANERS
UPDATE: 4 per CPUVP (this relationship has most recently been best for me.)
1 per 500-750 buffers, up to 128 (max)
More LRUs better support large number of users by reducing buffwaits

To monitor –g spi shows contention for individual LRU queues

LRU_MAX[MIN]_DIRTY - %buffers assigned to modified queue

lower to decrease checkpoint duration
For data loading (except HPL express mode) & parallel index builds 70 & 80%, allow to almost fill before flushing
 

Monitor: –R queue length, -F writes forced by this parameter by CLEANER thread

LTAPEDEV

Set to /dev/null Informix will just marked the LL as backed up, not even going through the motions. This should be done only on systems where logical log recovery is not required and the user has been made aware of the implications.

LTXHWM

50

LTXEHWM

60

LBU_PRESERVE - Preserve last log for log backup

NOTES:
If physlog frequently fills decrease CKPINTVL
phys logging - buffsize/pages IO SB >75%, if near 100% increase physbuff size
physical and logical log buffers should be about 75% full when flushed.
Huge bang for the buck makes Checkpoints another early thing to look at.
 

 

 

Introduction

Main & CPU

Disk IO

Logging

Memory & DSS

Miscellaneous

Notes

Print Version

bamph

                     

Web design and content bamph