high number of page faults

[Paul Gold]

we have recently been experiencing hangs and application crashes on our K class 10.20 system

swap is 3 x memory
dbc_max is 20%

we can find nothing particularly wrong using dmesg or syslog - nothing is logged

when looking at memrpt screen in glance we have an average of 300+ Page Faults per second

this is about 3 to 4 times as many page faults as there are page requests

a couple of questions

1/ is there another command or tool other than glance that will confirm these page faults are actually happening and not a problem with glance, are there any knwon problems with glance and page fault reporting that are patched (wish I still ahd access to kdb/kmine)

I have looked at sar and vmstat but am not as familiar with them as with glance

glance version C.02.15.00
sar revision 76.3.1.7

2/
Any ideas on how I can start tracing page faults or performance problems like those stated above

thanks in advance

Paul Gold
CAST

[CHRIS_ANORUO]

Hi Paul,

Use this command (echo "bufpages"/D|adb -k /stand/vmunix /dev/kmem), to know the exact system bufpage requirement at the peak period. This will help you set the value of dbc_max_pct. Also set nbuf and bufpages to zero and swapmem_on to 1, for you buffer caching to be dynamic.

You can check these documents S3100002312A,B,C(DESCRIPTION : "Sys Adm: determining the cause of system performance problems")

[CHRIS_ANORUO]

Hi Paul,

Use this command (echo "bufpages"/D|adb -k /stand/vmunix /dev/kmem), to know the exact system bufpage requirement at the peak period. This will help you set the value of dbc_max_pct. Also set nbuf and bufpages to zero and swapmem_on to 1, for you buffer caching to be dynamic.

You can check these documents S3100002312A,B,C(DESCRIPTION : "Sys Adm: determining the cause of system performance problems")

[Paul Gold]

Chris

thanks for the reply

the command as typed doesnt give anything other than a newline ??

nbuf and bufpages are set to 0 already and swapmem_on is 1

I will have a look at the documents

thanks

Paul Gold
CAST

[John Palmer]

Hi Paul,

3 times the number of page faults compared to page requests seems pretty typical on some of the systems that I manage.

'vmstat -s' will give you all sorts of (cumulative) stats about paging, and 'vmstat 5' will give you page in (pi) and page out (po) rates over 5 second intervals.

There are no problems with glance that I know of and you seem to have a reasonably new version.

You say that your Buffer cache is limited to 20% max and presumably 5% min. What in this range does the buffer cache normally occupy? I find that the cache size gives you an initial feel for how much memory pressure that the system is under as yours seems to be. Another is how much cpu time that the process 'vhand' consumes.

Regards,
John

[paul courry]

I might also suggest SOS from Lund Performance Solutions as a better alternative to Glance.