Load average recommended thresholds

[luissma]

Hi all,

I need to establish a policy for monitoring my HP-UX systems and I’m trying to figure out which would be the recommended thresholds for warning and critical load average alerts.

After having read a lot of documentation it’s not very clear for me what would be recommended values from top’s load averages.

Some people recommend maximum load averages of 3 or 4 and some other say there is a problem on the system only from values of 14-15 on.

I have, for example, a HP-UX 11.31 running on a ia64 hp server BL870c with following characteristics:

CPU info:
4 Intel(R) Itanium 2 9100 series processors (1.59 GHz, 18 MB)
532 MT/s bus, CPU version A1
16 logical processors (4 per socket)

Memory: 131041 MB (127.97 GB)


Top’s output usually shows values like following ones:

Load averages: 0.35, 0.39, 0.42
826 processes: 737 sleeping, 87 running, 2 zombies
Cpu states:
CPU LOAD USER NICE SYS IDLE BLOCK SWAIT INTR SSYS
0 0.46 29.7% 0.0% 6.9% 63.4% 0.0% 0.0% 0.0% 0.0%
1 0.25 41.6% 0.0% 8.9% 49.5% 0.0% 0.0% 0.0% 0.0%
2 0.47 23.8% 0.0% 5.0% 71.3% 0.0% 0.0% 0.0% 0.0%
3 0.26 47.5% 0.0% 6.9% 45.5% 0.0% 0.0% 0.0% 0.0%
4 0.50 25.7% 0.0% 9.9% 64.4% 0.0% 0.0% 0.0% 0.0%
5 0.24 21.8% 0.0% 3.0% 75.2% 0.0% 0.0% 0.0% 0.0%
6 0.42 15.8% 0.0% 7.9% 76.2% 0.0% 0.0% 0.0% 0.0%
7 0.26 15.0% 0.0% 6.0% 79.0% 0.0% 0.0% 0.0% 0.0%
8 0.44 29.0% 0.0% 6.0% 65.0% 0.0% 0.0% 0.0% 0.0%
9 0.27 23.0% 0.0% 2.0% 75.0% 0.0% 0.0% 0.0% 0.0%
10 0.42 74.0% 0.0% 2.0% 24.0% 0.0% 0.0% 0.0% 0.0%
11 0.25 21.0% 0.0% 3.0% 76.0% 0.0% 0.0% 0.0% 0.0%
12 0.40 20.0% 0.0% 7.0% 73.0% 0.0% 0.0% 0.0% 0.0%
13 0.26 17.0% 0.0% 5.0% 78.0% 0.0% 0.0% 0.0% 0.0%
14 0.42 97.0% 0.0% 0.0% 3.0% 0.0% 0.0% 0.0% 0.0%
15 0.26 5.0% 0.0% 1.0% 94.0% 0.0% 0.0% 0.0% 0.0%
--- ---- ----- ----- ----- ----- ----- ----- ----- -----
avg 0.35 31.7% 0.0% 5.0% 63.4% 0.0% 0.0% 0.0% 0.0%


I’d thank if somebody could put some light on all this and let me know when could I run into trouble with load average. Of course, I should consider some other facts like slowness of system, etc, etc, but for the moment I need to take a specific value to set my alerts.

Thanks in advance.

[Hakki Aydin Ucar]

Your current status ,load average is good (very low)

If you use (if you have) glance tool ,you can use its default, to analyze load average.
By default cpu bottleneck is around 50% ,alarm will be triggered as warning in glance.

additionaly, On multi-processor systems, the values represent the average over all CPUs.
Thus the percentage columns never exceed 100. For individual processor detail,
use the 'a' (CPU By Processor) screen.

[TTr · Accepted solution]

The CPU load is a very misleading metric including its name.

You can have a load below 1 and have a serious bottleneck and vice-versa you can have a load of 20 and have a very responsive system.

Look at Bill's responses in these threads

http://forums11.itrc.hp.com/service/forums/questionanswer.do?threadId=750840

http://forums11.itrc.hp.com/service/forums/questionanswer.do?threadId=876751

http://forums11.itrc.hp.com/service/forums/questionanswer.do?threadId=452648

http://forums11.itrc.hp.com/service/forums/questionanswer.do?threadId=1071636

[Court Campbell]

I have always hated the load average output. It may be because I have known old school admins that place to much emphasis on it. From my understanding you are possibly hitting a bottleneck if your load average is nearing the same amount you have in cpus/cores. So if your load average starts getting close to 16, then you might have a problem. I say might because it is not necessarily bad that your processors are doing work. Tuning is a science and an art. You really how the system is supposed to function and what your users are expecting. I would suggest getting some baseline data and making comparisons along the way.

It is hard to give an exact number since every server/application is a little different. There are no hard set numbers to give anyone. That is why performance/tuning guys make lots o' cash. In the words of any good tuning guru, the answer is "It depends".

[luissma]

Understood. Thanks to all, including Bill Hassell :)

Regards,

Luisma Arranz