Howto vmstat

[BRAGEUL]

I suppose that i have got a performance problem on my DS25 Tru64 5.1.B . I use vmstat and i note this result (fault value in my case is strong )
root> vmstat 5 3
Virtual Memory Statistics: (pagesize = 8192)
procs memory pages intr
cpu
r w u act free wire fault cow zero react pin pout in sy cs
us sy id
4 209 39 91K 139K 21K 11M 1M 5M 348 2M 0 67 1K 397
3 1 96
3 209 40 91K 139K 21K 2 15 23 0 27 0 399 4K 1K
55 3 42
3 209 40 91K 139K 21K 0 0 0 0 0 0 404 4K 1K
57 3 40

What does it mean 21K in fault value. Is this value normal.

Can you Help me to describe this result
My config :
DS25 mono Proc With 2Go RAM and 2 Database Instance Oracle 9i (50 Users).

[Michael Schulte zur Sur · Accepted solution]

Hi,

to me it looks like, the 21k belong to the memory wire and fault is 11m.

btw, the first row is the accumulation since boot time.

greetings,

Michael

[BRAGEUL]

Thank for your response,

But how can i suspect on my server a performance problem about (memory, Cpu...)
What are the main critical value of result of command like vmstat or collect or sys_check.

Bests Regards

[Michael Schulte zur Sur]

Hi,

collect I don't know, but sys_check creates a html file and makes suggestions, if values are too low or high. Also you can watch the system with top and monitor.
One important value is page out. This should not happen on a regular basis becauses it indicates a memory shortage and makes the system slow.

From what I have seen from the vmstat, your system isn't heavily burdoned, but it is of course only a show period.

How long has the machine been up?

Michael

[Hein van den Heuvel]

That actually looks healthy from a systme perspective. A decent amount of user time, modest system time, reasonable interupt-systemcalls-contecswitch ratios, Free memory, no paging (during vmstat!).

The only worrying component in your post is the opening line: "I suppose that i have got a performance problem". Could it be that all is well, or are your users complaining?

If Oracle is the bulk of your work, then even without serious investigation you may want to consider to give more memory to Oracle. You seem to have 1GB free ?! How big are the SGA's for the 2 instances? Double in size?!

For serious investigations be sure to run Oracle statspack over a representative period of production time and share hightlights of the report(s).

Regards,
Hein.

[BRAGEUL]

Hi Hein,

You told me that you have seen 1Go free of memory . Did you see this value in vmstat stat result ? so is it the free value 139K in vmstat result ? so what is the mathematique formul to obtain 1Go with a result of 139K ?

....
Is oracle StatPack is a additional product (free or not) ?

Thanks

[Johan Brusche]

A standard page in Tru64 contains 8192 octets(bytes). So the 139Koctets under vmstat's "free" column gives 139*1024*8192=1166016512octets=1,16 GB

Cordialement,
Johan.

[Hein van den Heuvel]

Right.

For an other, more detailed, view on (Physical) memory try: vmstat -P

It show some pages to bytes info, and is a nice way to see 'GH' memory should you ever decide to use that Oracle enhancer.

(sysconfig -q vm | grep gh)

Hein.

[BRAGEUL]

Hi,

This is a result of vmstat when my server in full used.

is there any strong value ?

[Hein van den Heuvel]

Ah, that's a different picture.
Now you are almost out of memory.

It is not as bad as it sounds because the ubc had grown to 800M (97098 pages) and can readily be trimmed down if more memory pressure occurs. Considering that this is an Oracle box, you may want to limit max_ubc to say 20% (still 400MB) and give the memory directly to Oracle (SGA) wihout it having to fight the ubc. Also, try GH or bigpages !?


Overall it still looks like a reasonably happy system.

You are doing a little more system time then I would like to see. Do you have timed_statistics on for the Oracle instances? Are you actually collecting stats?
What level?
SQL> show parameter statistics

You do NOT want level TYPICAL as it seems to add little value but uses a lot of system time through getrusage calls. Try:
alter system set statistics_level=BASIC;

You may also want to create a time-device to allow oracle (after restart) to read the time mapped from memory instead of needing the gettimeofday system call:
mknod /dev/timedev c 15 0

Cheers,
Hein.

[BRAGEUL]

Thanks Hein,

So i've got TYPICAL Value in oracle statistic and i will set to BASIC Value as soos as is possible because i don't use this fonctionnality. The device timedev is already created.

I read in a Oracle documentation that GH is only use on Alphaserver GS series and i've got DS20 and DS25 servers. Where can i set the max_ubc to mimimize the %memory use ? is there a specific sysconfigtab paramaters

I got this sysconfigtab parameters actually.
vm:
swapdevice = /dev/disk/dsk1a,/dev/disk/dsk1b,/dev/disk/dsk1g
vm-swap-eager = 1
vm_segmentation = 1

vm_ubcdirtypercent=10
vfs:
noadd_exec_access = 0

bufcache=1
fifo_do_adaptive=0
sec:
acl_mode = disable

ipc:
sem_msl = 100
sem_mni = 64
shm_mni=256
shm_seg=128
shm_max=2139095040
proc:
max_proc_per_user = 512
per_proc_stack_size=33554432
per_proc_data_size=201326592

I find this values in Oracle-Tru64's documentations and forum but i'm not sure that it was a good choice ..

Oups !!, I forgot to say that I had also 25 Progress 91D database in the same Oracle server ...

Didier.

[Hein van den Heuvel]

That's to 'typical' to have stats enabled to the max and then not look at the data. :-).
Try basic. If you can, collect (with the collect tool!) a before/after picture with a similar load? Mayb over a batch job?).

ubc:

sysconfig -q vm | grep ubc
ubc_minpercent = 10
ubc_maxpercent = 100

gh:

Equally valid for all Alpha's. But the GS series tend to have more memory and more user processes sharing that memory so the effect will be more dramatic.

big-pages: similar improvements to gh, more uses (code and data), easier to configure (just switch on). Be sure to get the latests patchkit (pk3) though because there have been problems with big-pages and earlier versions. Myself I used them find with no incidents for Oracle/Sap benchmarks, but I suppose you should give it a trial run if you can and doublecheck your backup situtation.

Hein.

[BRAGEUL]

Hi hein,

this is the value for
Myriam_Root >sysconfig -q vm | grep ubc
ubc_minpercent = 10
ubc_maxpercent = 100
ubc_borrowpercent = 20
vm_ubcpagesteal = 24
vm_ubcfilemaxdirtypages = 4294967295
vm_ubcdirtypercent = 10
ubc_maxdirtywrites = 5
ubc_maxdirtymetadata_pcnt = 70
ubc_kluster_cnt = 32
vm_ubcseqstartpercent = 50
vm_ubcseqpercent = 10
vm_ubcbuffers = 256
ubc_ffl = 1

If i understood, i must set
ubc_maxpercent to 20 instead of 100 ?

I got Tru64 5.1B PK2 in the first (DS25)
Tru64 5.1A PK5 on the second (DS20)

The kernel parameters and applications are the same in both servers.

I suppose that i can not use bigpage parameter without applyed the the lastest dupatch kit ?

Didier.

[Hein van den Heuvel]

> If i understood, i must set
ubc_maxpercent to 20 instead of 100 ?

I would recommend that as a first attempt.
This is based on our experience with Oracle testing. But as with all performance tuning 'it depends'. We actually go down to 2 or 4% for systems that clearly only do Oracle work, minimal fiel system activity.
The suggested 20% is 'middle-of-the-road'.
It is dynamic, so just try!


big-pages is a V5.1B feature, and should only be used in production with PK3 or later.

Hein.

[Antoine Morpain]

Before enabling bigpages you may apply the following early release patches:

9-V51BB22-E-20040427 - HP Tru64 UNIX V5.1B PK2 (BL22) ERP Kit:
VM-Related Fixes for V
Content type: Tru64 5.X patch document (ITRC Login Required)
OS: Tru64
Release date: Sunday May 9, 2004
http://your.hp.com/m/S.asp?HB13644154639X3500131X373163X

9-V51BB24-E-20040423 - HP Tru64 UNIX V5.1B PK3 (BL24) ERP Kit:
VM-Related Fixes for V
Content type: Tru64 5.X patch document (ITRC Login Required)
OS: Tru64
Release date: Sunday May 9, 2004
http://your.hp.com/m/S.asp?HB13644154639X3500133X373163X

[Ralf Puchner]

Also check with the Oracle tuning guide what values must be set. Ask Oracle for the documentation containing also formulas how to calculate each value and what type of memory model is best for your database.

Oracle can also give you the list of supported OS-versions and necessary patch kits tested (e.g. certified).