monitoring major page faults cause

jack malkovick · 06-04-2018, 12:17 PM

Hello

How could I monitor the cause of the major page faults in my app?
At least the virtual address in process space where they happen.
I know they are happening, but I'm using a couple of memory mapped files and I don't know which one of them are producing the corresponding disk activity.

Thank you

X-LFS-2010 · 06-05-2018, 12:48 PM

sounds to me like you have bad libraries (your version of libc that supports "memory mapped files", broken)

it's not memory mapped if the red led disk light is on is it? nope.

HOWEVER: remember that if you "sync" or if the file changes that the kernel may "do maintenance". memory mapped file does not mean loaded into memory. it's still subject to kernel maintenance which could mean writing cached data because ... "there's nothing else to do right now".

Habitual · 06-05-2018, 07:32 PM

collectl may be able to help.
sysstat utilities been utilized?
snmp ?

syg00 · 06-05-2018, 08:39 PM

I would be a bit surprised if this sort of info was exposed to userland.
I see there is a kernel trace event for "page_fault_user" that has a field for address - probably the guy you're looking for, but needs kernel tracing set up for it and filtering for PID. Not to mention a kernel with the appropriate option enabled.

Habitual · 06-06-2018, 12:01 PM

http://collectl.sourceforge.net/Documentation.html

syg00 · 06-06-2018, 05:42 PM

Last I looked collectl gets its info (primarily) from /proc and sysfs - with luck Mark will drop by and comment authoritatively.

Habitual · 06-07-2018, 11:35 AM

Quote:

Originally Posted by syg00

Last I looked collectl gets its info (primarily) from /proc and sysfs - with luck Mark will drop by and comment authoritatively.

He has the best answers!.

jack malkovick · 06-10-2018, 04:10 PM

Thanks for your replies. I used trace-cmd, but unfortunately not all kernel versions have support for page-fault events...