Tuning engines is one of my all time favorite activities be they internal
combustion or database.
Rare is the database engine that cannot be tuned for another 20%.
orders of magnitude improvement is not uncommon
for a well organized tuning
session. This is not due to any mysterious talents of mine. Rather, that systems go out of
tune rather quickly or were never tuned in the first place.
Below is a list of suggested settings for Informix Engine Configuration parameters I
have gathered over the years. I keep a copy in my organizer for easy field reference.
I have collected many books that discuss tuning Informix databases. It has been my
observation that they conflict more than they agree. In my mind there are several reasons
Changes to underlying engine architecture obsoletes recommendations with tremendous
Unix implementations differ wildly at granularities important to the tuner.
The rate of hardware advances quickly overtakes information learned.
The first item above was demonstrated in an early release of version 7. The Informix
engine disk IO subsystem changed so markedly that all previous NUMAIOVPS recommendations
A wonderful example of item 2 above is the efficiency of the implementation of the Unix
process manager. A series of tests I performed at Informix's benchmark labs in Menlo
Park demonstrated configuring
as many CPUVPS as hardware CPUs was optimal for
HP/UX. The version of Solaris that we were testing preferred the recommended 1:1
relationship for this.
My favorite conflicting pronouncements are on page 6-134 of INFORMIX-OnLine Dynamic
Server Performance Tuning Training Manual (2/97). When discussing configuring NUMAIOVPS
for systems that do not use kernel IO the page states a 1:1 relationship with controllers
third sentence following
states one per disk.
Note that these examples discuss wide variability in what are arguably the
This should leave the reader with three important conclusions.
It is dangerous to speak of such things in absolute terms.
Each important end user environment must be measured and tuned individually i.e. cookie
cutter approaches will cause trouble for the DBSA.
More importantly, due to the fluidity of this most fundamental knowledge set DBSA should
continue to be well compensated.
What follows is my collection of notes for each onconfig parameter. Most are direct out
of a book. Comments in
are from my own
Where more than one suggestion is listed they are in chronological order, oldest first.
The assumption is made that the system under test (SUT) is a dedicated DB server.
it isn't, it should be.
Informix should be thought of as an Operating System unto
itself. All it really wants from Unix is slices of CPU time. If it must share it's sandbox
with applications, other database engines, or even other Informix instances it is
operating out of the bounds of it's target design.
Acknowledgement that this level of isolation is not always possible in the real world
has been addressed recently with new onconfig parameters. Important adjustments for tuning
in a "mixed environment" are marked "
Any mission critical DB should be on a dedicated machine.
"-" implies the onstat command.
Sometimes the DSBA does not have the luxury of an extended tuning session. He must make
a quick change to a production box to get it back within acceptable parameters. I have
noted parameters that allow for a "
big bang for the buck
" with a
These parameters are grouped logically.
Sometimes the DBA does not have the luxury of an extended tuning session. He must make
a quick change to a production box to get it back within acceptable parameters. I have
noted parameters that allow for a "big bang for the buck" with a
Logical Logs + [Phys logs] + [Temp Tables] + [Data] + [On-Arc.catalogs] + crtl info
This is a basic size calculation formula. For
systems of any import most of these logical units should of course be on separate
unique for each engine on CPU.
On a dedicated server turn this on (1). The parameter is
ignored if it is not supported.
In a mixed environment turning this on may make Informix
"not play well with it's freinds"
#HWCPUs - 1 if #HWCPUs > 3
#scan thrds (frag issue) SB a multiple or factor of NUMCPUVPS determine w/ glo ath rea.
For data loading = #HWCPUs unless HPL used, then = HWCPUs - CONVERTVPS (onpload config
I have seen systems that enjoy 1:1 VCPUS to HW CPUs
relationship and I have seen some that prefer 2:1:1
Both of these turn on different housekeeping mechanics. If
NUMCPUVPS is set to one it is
very important to turn both of these off,
or 0 and 1
respectively (yech). You will see no benefit to configuring one CPUVP unless these are
both off. This is overlooked more often than not!
I would suggest that these should really be considered as one
parameter together as if one is adjusted the other should be adjusted likewise.
1 if supported this parameter turns off Unix nicing (this is
good). Nicing is the Unix mechanism that lowers a processes priority over time to ensure
equality in a mixed environment. As Informix just wants machine cycles a dedicated
server should have this enabled.
Tuning for a mixed environment is not so black and white. It
may be best to have this off initially and turn it on if it is desired to give Informix
more CPU time.
CPU number to start binding to. mpstat will provide number of
CPUs and their associated number. This numbering system seems whimsical.
Number of hardware CPUs to bind to.
NUMCPUVPS = HW CPUs - AFF_SPROC.
This can be very beneficial for systems that are not dedicated
to the DB.
Other non-Informix processes may be allowed to run on the HW
CPUs identified here, but the CPUVPS will be restricted to those identified.
0 internal timer is faster.
I have never seen OS timing used.
-g glo There should be a 10:1 ratio between time spent in usr
vs. sys for CPUVPS. (This has become perhaps 3:1 with KAIO). If sys is too high on a
system with just a few HW CPUs try using just 1 CPUVPS.
-g rea If this shows seven or more threads waiting adding
CPUVPS can bring it down.
Very marginal tool. load average is a combination of system
resource measurements. I have seen slow systems with a la of 2 and I have seen systems
that seemed find at double digits.
Useful only relative to an earlier measurement on the same
KEY OLTP TUNING FIELD
start w/20% RAM, may be up to 50% RAM.
On a dedicated machine
why not start at 50%?
increase size till increase in cache hits insignificant or excess system paging occurs,
use sar or vmstat to determine excess paging.
OLTP Target 95% read, 85% write cache hits
buffers smaller than largest table for DSS will force light scans. Use -g lsc to
measure light scans
Maximize for data loading (50% or more) (except HPL express mode)
More buffers can mean longer checkpoints
- Suggestions for this
seem to change with the weather.
1 per db disk + 1 for ea. chunk accessed freq.
If KAIO is used allocate 1 + 2 for each cooked chunk
of any cooked chunks)
For KAIO systems 2 for OnLine 1 per controller containing cooked chunks
For systems w/o KAIO 2 for OnLine + 1 for each controller then add as indicated.
1 per dbspace
1 per disk
1 per mirrored pair
1 per chunk.
'-g ioq' to monitor IO Qs
DSA spawns one read thread per dbspace (AIO
My suggestion is to get a system that supports KAIO and then
set this to 2.
most machines limit at 30, all limit at 32.
Dig: For systems that do not perform light scans, do not set RA_PAGES higher than 32.
higher for typically sequential DSS
if too high will lower %cached reads
if bufwaits unusually high RA_PAGES may be too high, or difference between RA_PAGES and
RA_THRESHOLD may be too small.
Set close to RA_PAGES e.g. RA_PAGES 32 and RA_THRESHOLD 30, if bufwaits (-p) increase
If most machines limit at 30 won't the
RA_THRESHOLD remain in a constant TRUE state?
Ideally RA-pgsused = (ixda-RA + idx-RA + da-RA)
at least two each on a different drive, more if building large
DSS environments should use HW striping a small number of
TEMPDBS across multiple disks.
Max space required for index build is: non-fragmented tbls (key_size+4) * num_recs *2,
fragmented (key_size+8) * num_recs *2
90 is typical, 100 for SELECT/DELETE only tables
forces initially very compact indices & efficient caching.
50-70% for tbls with high INSERTS to delay need for node splitting
A few years ago the fellow that tests this at Informix posited
on USENET to
use HW mirroring over Informix
. This make sense as
who will have a more intimate knowledge of the devices? HW solutions are always faster
than SW. In order of preference I would suggest HW, OS and then Informix mirroring.
For machines where availability is paramount one can mirror
across controllers and even arrays.
when set to "1" this undocumented parameter will generate read and write
timings in the syschktab SMI table. See Appendix C of DSA Performance Tuning Training
Increasing the Unix priority of AIO processes can improve the performance of data
returned from disk.
Monitor IO with -g ioq
(iof and iov are also worthy).
When AIOs used gfd len SB < 10, maxlen <25.
breaks 25 during engine initialization when it is unimportant so make this distinction
-D will show hotspots at disk level -g ppf at partition level.
When building an important Data Warehouse for my current employer
the Sun Hotshot suggested placing all data in only the middle 2GB sectors of each 4GB
disk, leaving the remaining unused. The highly paid Informix representative felt strongly
that using only the leading 2GB of sectors would perform better. I suggested that we test
and if it was within 5% that this be decided by ease of maintenance. The leading sectors
proved 2% faster than the middle. I do not recall which I ended up implementing.
If your system is IO bound verify if it be controller or disk
bound. The solutions are different.
The use of clustered indices can greatly increase sequential reads.
Informix recommends using fragmentation over HW striping unless the table is a poor
canditate for fragmentation.
I would like to test this statement
I have not been able to test how Kernal IO effects NUMCPUVPS
configuration. From DSA Performance Tuning Manual (2-97)
"If your system
supports kernal aio, onstat -g ath will show one kio thread per CPUVP."
Therefore should NUMCPUVPS be associated with the number of disks, etc
or remain a function of the number of hardware CPUs?
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.
! rootdbs. Place on a separate spindle
= usrthreads * 5 (or size of most freq. blob) * 4
= usrthrds * 5 * 4
UNLESS tblspace blobs in DB w/o logging then usrthrds * size of freq blob pg * 4
In all my manuals I can find no good reference for performance
tuning this parameter.
I have no experience adjusting this save for eliminating
small for greatest recovery, if tape slow, or blobpages volatile
(connects * maxrows (in one trans)) * 512
size of 3 LL buffers in RAM
determines freq of flushing to disk
LOGFILES + 3
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
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
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.
LBU_PRESERVE - Preserve last log for log backup
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.
1 if CPU, additional poll threads assign to NETVPs
300 single HWCPU 350 if more
For data loading one per CPUVP. Each poll thread should be on a CPU class VPS (running a
Do not increase user connects as this will increase work for the poll thread
max # consumed by any query * # concurrent users
OLTP - Big Resident (buffers), small Virtual (SHMVIRTSIZE)
DSS - Small Resident, big Virtual
DSS may be up to 75% RAM if paging is not induced
DSS KEY FIELD
SHMADD (kilobytes) -
10 - 20% of SHMVIRTSIZE
can be used to make Informix more polite reserving resources
for other applications. I have also had to use this when a failing malloc panics oninit
Monitoring: -g seg - One of the first things I check as
consolidating shared memory segments can be huge bang for the buck.
Note that onmode -F (used to free memory segments) can cause
system failures on an active system and should be avoided.
Sar -g 3 3 - will show paging activity
To calc shared memory segments corresponding to a DB instance shmid - 52564801*.0001
if > 0, will enhance parallelization of index builds
after 7.2 are always parallel to some degree)
Each sort thread gets quantum/#sort threads memory
A users effective priority = (pdqpriority/100) * ( MAX_PDQPRIORITY/100) where
pdqpriority is set by the environment variable or the SET PDQPRIORITY statement.
0 As many systems that I have seen fail due to Informix filling
/tmp the "out of the box" default should be off.
Oncheck -pr can be used to replace lost onconfig file
Resident - Buffer Pool
(Pool is for residents use only - I
needed a mnemonic for this.)
Virtual - Light Scan Area
(Virtually no logging - this
28 bytes per data page is used by the engine
4 byte slot entry for each row
Scans and Sorts for index builds are always parallel with 7.2. Idx builds on fragmented
tables add parallel B-(sub)tree builders
UPDATE STATISTICS - single most important SQL statement for Q perf.
lead columns in each index
all columns queried with equality filters (=)
all join columns
col to uniquely distinguish a composite idx from another on same table
and all cols preceding
all other columns
all idx cols. Not run through on HIGH
To speed up update statistics set PSORT_NPROCS to 2, use DBSPACETEMP (duh) do NOT set
DBUPSPACE (limits RAM for US)
The following came from a class handout. I have no idea the original author.
Database Tuning the Informix Way.
Establish performance objectives.
Measure database activity and use of resources.
Identify performance problems such as excessive use of CPU, memory, or disks.
Tune the operating system.
Tune the Online Dynamic Server
Optimize the placement of logs, sort space, and temporary space.
Optimize table placement, sizes of extents, and fragmentation
Make sure the indices are appropriate.
Optimize background activities such as logging, checkpoints, and page cleaning.
Schedule backups and batch jobs for off-peak hours.
Review application programs to make sure appropriate access methods are used to retrieve
data and algorithms are efficient.
Repeat steps 2 through 11.
My additions would be:
.5 and 2.5 Set users expectations
3 & 4 Find a seasoned Systems Administrator for
11 could be performed much earlier in the cycle as more often
than not, the largest performance gains come from applications tuning.
Set aside as much time as possible to dedicate to this task.
Three days is a minimum. In a development environment ensure that two weeks are dedicated
to this at the end of the project. This will invariably get squeezed to three days which,
as above, is the absolute minimum to perform a thorough job.
Calculating maximum number of extents for a particular table:
Max#extents <= (pagesize - ((4 * number of columns in a table) + (8 * number of
BLOB and VARCHAR columns + 136) + (12 * number of indices) + (4 * number of columns in the
indices) + 84))