Problem:

 

Since upgrading AIX to 6100-07-04 or 6100-0705, I noticed that paging space usage is very high.

 

Solution:

You may have hit an APAR.

IV26272: REDUCE EARLY WORKING STORAGE PAGING


http://www-01.ibm.com/support/docview.wss?uid=isg1IV26272
An explanation and possible resolution 

The kernel parameter, numperm_global , was implemented and enabled with with AIX 6.1 TL7 SP4/7.1 TL1 SP4 to be able to look at the paging from a global perspective. That means that before AIX 6.1 TL7 SP4/7.1 TL1 SP4 the number_global tunable was not available. Unfortunately the number_global might cause in some environments early paging due to failed pincheck on 64K pages related to different size mempools. There are two possible ways to prevent the problem from happening:             

1) Disable global numperm (numperm_global=0)
2) Increase the number of unpinned pages for the page size that is close to maxpin - In the previous case, the 64K page pool is close to maxpin.  Most of the pinned pages are kernel heap and there is a large number of 4K computational pages used by the customers application. Forcing the application to use 64K pages (LDR_CNTRL) will reduce the percentage pinned count for the 64K pool and therefore prevent the problem from happening.

Customers who are running workloads like DB2 or Oracle and follow the best practices (using 64K page size) are unlikely to experience this early paging problem.

Problem conclusion

  • LRU processing changed to perform file only 4K page stealing
    when 64K page size is at maxpin condition and there are 4K file
    pages available to steal in the mempool.