Download Perfpmr

[Brayan Jacobsen]

ThePerfPMR package was developed to ensure that reports of suspected performance problems in AIX were accompanied by enough data to permit problem diagnosis by IBM. This makes the shell scripts in PerfPMR useful to other performance analysts as well. PerfPMR is an optionally installable part of the AIX Version 4 Base Operating System. It is located in /usr/sbin/perf/pmr. See the discussion in "Installing AIX Version 4 PerfPMR". A version of PerfPMR is also available for AIX Version 3.2.5. See "Obtaining and Installing AIX Version 3.2.5 PerfPMR".

The script perfpmr is the highest-level script of the package, but it collects data, such as configuration information, that a local performance analyst probably knows already. The lower-level script monitor collects a coordinated set of performance information for a specified number of seconds and summarizes the data. The syntax of monitor is:

One particularly important use of the PerfPMR package is the creation of a configuration and performance baseline prior to a significant change in system hardware or software. Just as you probably back up critical files before such a change, you should make a record of the configurations and the performance they were providing. If a performance degradation should occur after the change, you will have detailed data that will let you perform a rigorous before-and-after analysis of the system.

during the busiest hour of the day. The output files from this measurement run will appear in directory /var/perf/tmp. (If you are running on a pre-Version 4 system, the output files will appear in the current working directory.) Be sure to move these files to a safe haven before beginning the configuration change.

Sem este APAR, o perfpmr pode travar no comando # xm -u do kdb, que pode ser terminado via comand kill. Basta verificar o PID atravs de # ps -ef grep kdb.

O comando # tail -f perfpmr.int pode ser utilizado para verificar a execuo do perfpmr e o ponto em que est travado, caso ocorra.

Nunca rode o perfpmr.sh em background, espere que o prompt do usurio root retorne. A estimativa de trmino sempre o dobro do tempo pedido no comando perfpmr. Ex. perfpmr.sh 600 ir levar 20 minutos para coleta e formataco dos arquivos.

Well, okay, perhaps a bit defensive. Sorry. Not all solutions are

published.This is a combo of experience and many years in support. I alreadysaid that it will work with the values first presented but performancewill be an issue. Its the same with vmo settings, base values will workuntil you get into high loads then you have to tune. In this case, its

lesstuning per sei and more of ensuring sufficient resources are available.I once did an install where there was no cpu or memory considered forthe vio server when the p570 was ordered. We managed to retrieve 1cpu from other partitions and 512 mb of memory. It ran the 9 partitionsbut not well. Booting partitions together took 4 x longer than it should

and this is purely io through virtual scsi adapters.

Well, okay, perhaps a bit defensive. Sorry. Not all solutions are published.

This is a combo of experience and many years in support. I already

said that it will work with the values first presented but performance

will be an issue. Its the same with vmo settings, base values will work

until you get into high loads then you have to tune. In this case, its less

tuning per sei and more of ensuring sufficient resources are available.

I once did an install where there was no cpu or memory considered for

the vio server when the p570 was ordered. We managed to retrieve 1

cpu from other partitions and 512 mb of memory. It ran the 9 partitions

but not well. Booting partitions together took 4 x longer than it should

and this is purely io through virtual scsi adapters.

I have access to an io test routine which will drive all IO at as close to

100% of its

capability. In order to get good io transfer rates, it takes cpu. The tool

is a

confidential product so I cannot share it. We use it to test cpu, memory,

networks,

disk, tape, CD/DVD etc.

As I said earlier, if there is an AIX support agreement, do a perfpmr. IBM

will analyze

under that agreement. If there is no agreement then continue monitoring this

forum

for an answer.

A perfpmr is the best tool to analyze when the problem is present.Having said that, how many instances are you running? What is the

databasea front end to? How many users at one time? Is the database on local orvirtual disk? The numbers I gave will ensure that the power is available

forheavyloads. This forum is to provide assistance and interchange ideas. He has

acpu wait problem and I believe its cpu cycles that are needed. I do

deferto the top line but a perfpmr analysis requires a SW support agreement.

A perfpmr is the best tool to analyze when the problem is present.

Having said that, how many instances are you running? What is the database

a front end to? How many users at one time? Is the database on local or

virtual disk? The numbers I gave will ensure that the power is available for

heavy

loads. This forum is to provide assistance and interchange ideas. He has a

cpu wait problem and I believe its cpu cycles that are needed. I do defer

to the top line but a perfpmr analysis requires a SW support agreement.

Hi Jan,

Why are you capping your VIO servers? If you have licensing issue you

should only cap your Database servers - LPARs which are actually running

licensed code like oracle or DB2. VIOS only run IBM code and if you have

PowerVM license for VIOS you are fully licensed and you should uncap

them. This is what we are doing here for our Oracle servers. Second if

you have p570 p6 and I think if you have HMC 7.3.3 and above and

firmware above 320-040 you should be able to use CPU pools. What you do

is, you created pool with number of CPU equal to your number of

Databases CPU licenses and then only allow your Databases servers CPU

belong to this pool and you should be able to uncap your LPARS and

share CPU between database LPARS.

For a service processor especiall on ethat handles I/o haveing a 0.1 is asking for IO elongation and delays. Not a good move. The answer is depends on your workload charchteristics. What HMC level are you running with ? You can use the hmc utilization manager to determine where your demand is.

I have had good success in using shared uncapped with VIO servers. I

assign enough CPU to support a virtual CPU for each real CPU and leave

it uncapped, this allows the VIO to use any available CPU if it needs it

and any other server to use any available CPU leftover from the VIO. I

have never had to increase desired number for a performance issue. I

use this same method to configure the VIO for all server types.

Depends on the amount of I/O they are processing. We have some as low as

.5 and some as high as 1.5. Remember if your servers are VIO backed and

you reduce the CPUs processing said I/O, you are resource starving them.

You might need to do some testing to find the happy medium since this is

a configuration/site specific question.

Lamar

Depends on the amount of I/O they are processing. We have some as low as .5 and some as high as 1.5. Remember if your servers are VIO backed and you reduce the CPUs processing said I/O, you are resource starving them. You might need to do some testing to find the happy medium since this is a configuration/site specific question.

Lamar

Oracle recommends two times the physical memory for the SWAP ...Sounds like you need to tune your Primary & Secondary Swap for your OS ...More importantly, once you have your Instance up & running,you need to sum all of your SZ or VSZ coulmn from 'ps -o' command ..(Posix

ps).and then multiply that with 4K page size..That will be the total Virtual Set Size for your Instance ...So you will have to boot to a single user mode & resize your swap ..Unlesss you have tools like IGNITE-UX in the HP-UX world...HTHzongk TuOracle Apps DBA/Unix SA

Next, since you get a message in errlog, you can set up a stanza in the error

notify class that can send you a message that this has occurred to you can get a

sense of the problem in real time instead of after the fact. Well, that's not

really true. You can also run "errpt -C" in a window and grep for the message

of interest and run some alert program - so errnotify is not the only way to go.

norm

---

Outgoing mail is certified Virus Free.

Checked by AVG anti-virus system ( ).

Version: 6.0.351 / Virus Database: 197 - Release Date: 4/19/2002

We are experiencing the same problem on an S85 running AIX maint level 09.

Oracle version is 817 . This box has 24GB of Ram and is a 12 way the

performance has been horrible with the 817 database. Tests that we performed

months ago on 734 were 40 percent faster .Here is what we have found and done in order of what had the best impact;1. Recalc the Oracle stats 3 times a day (This has helped tremendously) Why

the stats are getting blown is beyond us and beyond Oracle at this point.In

the past stats were only calculated monthly2. Oracle patched to level 8.1.7.3 3. IBM performance tuned based on performance (perfpmr) taken at peak times

and sent to IBM . They had us make vmstat , syncd changes. 4. We did update the bos.rte.libc to the latest level but our problem did not

seem to be a memory leak.

The re-calc of the stats in oracle seems to have the biggest impact on our

situation. Why we have to do it so often is still a mystery.Any feedback you have found would also be appreciated ...

3a8082e126