Re: 6GB free memory, but paging


Sorry for the delay in replying from my side, I'll try and add a bit to
what Laurenz said...

On 02/19/2009 03:58 PM, Falk Grunert wrote:
Laurenz Albe schrieb:
The kernel might write a memory page that is not used to the paging
space.
Free memory is not unused memory: it will be used for file system
buffering.

Now I'm really confused. I thought always that the file system cache is
included in the *used* memory an will be shown as numperm and is
(normally) between min- and maxperm%.


The size of the file system cache is more or less determined by minperm%
and maxperm%, while numperm% shows the usage of this cache.


But on this LPAR there are
"really" free memory and it is not be used for caching or something else.

Two things:

Firstly, AIX tries to maintain some small amount of "really" free
memory, controlled by minfree and maxfree.

Secondly, AIX may still page out seldomly used application pages in
favour of your file system cache pages even if there is plenty of real
memory available for your application.

On the O/S side this obviously depends a great deal on minperm/maxperm
again, but it could also depend on other settings.

You application settings (like PGA/SGA and Java settings) also play a
role in how much memory is used, since they place bounds on what you are
allowed to allocate.


In my simple mind I think it may be so:
Java says: I need more memory! My programmer have given me only
laughable 1GB! But damn, I mustn't. So I must fiddle with the paging
space. One 4k-frame in and two frames out, all in this tiny limit he
give to me :-/

Not correct. The portion of paging space you see in your case, is likely
unused application pages. See, the VMM "thinks" it is fine to use up to
80% of your real memory for file caching, since your maxperm% = 80. So
at the moment, the O/S has paged out the least recently used application
pages in favour for some anticipated future file access in the file cache.

To check your setup you do not need to first worry all that much about
paging space usage, but rather your paging _rate_, as seen from PgspIn
and PgspOut in topas, or from the 'sar -q' output.

If you notice a high paging rate you have two options - 1) decrease
maxperm% and maxclient% to a lower value a few percent at a time or 2)
leave maxperm% and maxclient% at a high value (90%) and turn
lru_file_repage = 0, in which cases file pages will always be first to
be stolen.

The first option makes sense on a hybrid environment, the second option
makes perfect sense on a dedicated database server. You should maybe
google for "Protecting computational memory" if you want more info.

Hope this helps!
Niel
.

Relevant Pages

  • Windows Vista Memory Tweak Guide
    ... Windows Vista is by large no exception to this rule. ... There are several relevant components to the Windows memory subsystem. ... A value of 0 specifies that inactive kernel-mode drivers ... When set to 0 a standard sized file system cache is ...
    (misc.news.internet.discuss)
  • Re: 6GB free memory, but paging
    ... memory, ... paging space usage, but rather your paging _rate_, as seen from PgspIn ... maxperm% and maxclient% to a lower value a few percent at a time or 2) ...
    (comp.unix.aix)
  • seg_map and cachelist
    ... While trying to understand Solaris file system caching process(Solaris ... Is 12% of physical memory, the max possible for file system cache ... pages getting relased are dirty, they are not moved into the cache ...
    (comp.unix.solaris)
  • Re: System Cache
    ... Yes, there is a cache in the memory too, and I think this works together ... much faster memory cache instead of the disk. ... This is the "file system cache". ... 256MB RAM. ...
    (microsoft.public.windowsxp.general)
  • Re: System Cache
    ... I thought that the File System Cache was an area of memory which held files recently accessed on the hard drive? ... If the application then needs to use the same file again or write to it, it is retrieved from or written to the much faster memory cache instead of the disk. ... 256MB RAM. ...
    (microsoft.public.windowsxp.general)