paging question (dont make fun of me!)

[WiccaChic]

This is probably a dumb question, but is there a quick way to determine what the top 10 processes that are paging out are? Had a scarry moment today when my system response reduced to a crawl and when I looked at vmstat I saw a lot of numbers in the po column that I dont normally see...but before I could figure out who was doing what the situation cleared up.

 

[hfaix]

This isn't the best way, however, often times I run topas and once in topas do a shift-P

then use the pid in a "ps -ef | grep PID"

This is a quick look at the running processes.

 

[MoshiachNow]

Could you please post the result of "vmstat 2 10" at your peack time.

"Long live king Moshiach !"

http://www.7for70.com/

 

[raztaboule]

svmon -P should help you next time.

 

[WiccaChic]

Hey, I did not even know about that topas <P> option! Thats kind of nice. Thanks for the svmon tip, that looks like the best way to go for me.

 

[screwloose]

ps gv | head -n1; ps gv | egrep -v "RSS" | sort +6b -7 -n -r | head -n 5

or

ps au | head -n 1; ps au | egrep -v "RSS" | sort +3 -r | head -n 5

or

svmon -P -n -w -v

 

[sbix]

You can have a good idea of processes causing paging activity doing:
ps avx|head -1;ps avx|grep -v PGIN|sort -nr +4

This will give you the 10 top processes causing more I/O (PGIN field)

 

[RodKnowlton]

The PGIN field is a bit of a red herring because AIX uses the same mechanism to page in from paging space as it does to read files. Both sorts of pages are counted in the PGIN figure.

That said, if you see a process with a rapidly rising PGIN field during a slowdown with lots of 'po' and not as much 'pi', it is a fair indicator that the vmtune parameter maxperm is set too high. For many loads, the default setting of 80 is too high.

Let's say your computational pages take up 4 of your 8GB. Everything will be fine until 4GB of file pages have been read. After that, the VMM has to free pages for each new page, whether the new page is a file page or computational. It will first free file pages only, but if enough freed file pages are immediately paged back in (repage) because of an access, the VMM will start paging computational pages out, as well. This will continue until the computational page repage count exceeds the file repage, or the percentage of real memory occupied by file pages exceeds maxperm. Only in the second situation (file pages % > maxperm) are computational pages off limits to the page stealer.

Optimally, as long as your computational working set is smaller than real memory, you should be able to adjust maxperm so that paging space is rarely used. Off course, it's not really that simple. It could be that you'd get better performance overall by letting a certain amount of underutilized computational pages page out, increasing the real memory available to file caching.

Clear as mud, huh?


Rod Knowlton
IBM Certified Advanced Technical Expert pSeries and AIX 5L