High Runq/load

[Roger Baptiste]

Hi,


N class server/8MB/4 Way CPU/hpux11.00
Shows high "load" in the output of top/glance/measurware/sar, even though the CPU seems to be having idle time. The run queue is too large.

top output:
CPU LOAD USER NICE SYS IDLE BLOCK SWAIT INTR SSYS
avg 71.69 27.2% 0.0% 38.0% 34.8% 0.0% 0.0% 0.0% 0.0%

sar output:
22:01:50 runq-sz %runocc swpq-sz %swpocc
22:01:55 67.3 90 0.0 0
22:02:01 73.1 95 0.0 0
22:02:06 74.4 90 0.0 0
22:02:11 51.7 95 0.0 0
22:02:16 51.4 80 0.0 0
22:02:20 52.4 85 0.0 0
22:02:25 55.8 80 0.0 0
22:02:30 55.4 85 0.0 0
22:02:35 68.2 85 0.0 0
22:02:40 60.7 90 0.0 0

Average 61.3 87 0.0 0

Measurware extract:
| | |System| User |Memory|Active | Pri | Run |
Date |Time |CPU % |CPU % |CPU % | % | Proc |Queue |Queue

04/29/02|20:55| 63.15| 54.85| 8.30| 57.55| 402| 0.39| 70.92|
04/29/02|21:00| 77.06| 59.61| 17.45| 56.90| 398| 0.52| 71.25|
04/29/02|21:05| 72.73| 56.90| 15.83| 56.53| 383| 0.30| 70.45|
04/29/02|21:10| 72.81| 57.61| 15.20| 56.40| 381| 0.62| 71.25|
04/29/02|21:15| 73.92| 58.90| 15.02| 56.19| 385| 0.34| 70.82|
04/29/02|21:20| 64.33| 55.40| 8.93| 55.98| 381| 0.38| 71.12|
04/29/02|21:25| 65.84| 57.54| 8.30| 56.02| 374| 0.38| 70.89|
04/29/02|21:30| 79.17| 59.79| 19.38| 56.03| 383| 0.69| 71.50|
04/29/02|21:35| 83.00| 61.01| 22.00| 55.90| 374| 0.84| 71.78|
04/29/02|21:40| 84.90| 61.52| 23.38| 55.61| 380| 1.04| 71.54|
04/29/02|21:45| 79.63| 60.36| 19.27| 55.56| 383| 0.73| 71.92|
04/29/02|21:50| 85.85| 57.26| 28.59| 54.83| 380| 0.36| 71.18|

****


Swap is set correctly (8Mb), all kernel params seem good (including usual suspects like timeslice, dbcmax etc).
System is a database server, but the top cpu consuming process is midaemon.

I understand loadaverage/runq in isolation is not meaningful in measuring system performance. I am trying to figure out why in the first place is the value soo high? This is a newly configured system.
Any clues?
thanks

[James R. Ferguson]

Hi:

One thing to check is the kernal 'timeslice' value. If it is set low (perhaps to one when you used a DB template) then high context switching may result in more switching and waiting for a processor than doing real work. Generally, a timeslice value of <10> is appropriate.

Regards!

...JRF...

[Michael Tully]

Hi RajMan,

Have a look at your disk IO. This will produce this type of problem. I had this very problem after disconnecting a disk array from the host. Also have a look for any SCSI errors in the system log, root's mail and the EMS event monitor log. /var/opt/resmon/log/event.log

~Michael~

[Wodisch]

Hi RajMan,

hopefully your system does NOT have 8MB - that's the amount of RAM my Linux PDA has got ;-)
In that case I would recommend the "McObjectsDB" for embedded systems...

Else I guess you are running Oracle, and then have a look into your "init*ora" for the parameter (or absence of) "cpu_count" - Oracle uses only as many CPUs as permitted by that parameter, and the default is ONE.
Which means that 3 out of 4 CPUs are usesless for database processing, and hence your runQ will be high (but only for that single CPU).

Just my $0.02,
Wodisch

[Roger Baptiste]

Thanks for the responses.
Timeslice is 10 (as i already mentioned in my question post).
There are no disk-related errors.
Cpu_count is not set, but it is not set on many other systems too and none of them have a high runq.

So, the question remains. Any more clues ?

thanks

[Helen French]

Hi Raj:

I would defenitley check the system patch level and will update if necessary. There is patch - PHKL_26059, which has some fix for high CPU loads, issues regarding midaemon etc.

If you have other systems with the same setup, then I will defenitely compare the system parameters, kernel parameters, networks, disks and file systems.

Not sure whether this helps ...
Shiju