Kai Schaetzl wrote:
Feizhou wrote on Tue, 06 Jun 2006 10:54:26 +0800:
Yeah but with that patch that changed the vm system behaviour, setting swappiness to 0 had ZERO effect. The kernel happily swapped out what data it could from programs that were running.
But you apparently need some additional environmental condition for this to happen. I see normal behavior as ever with my 4.3. After boot it slowly fills up memory until less than 50 MB is left. If the machine is scarce on memory it even gets almost to zero physical mem left. swap file usage stays at a few 100k. I don't have desktop machines, only servers. I suspect that the problematic behavior occurs only when you close applications. In that case it might swap the diskcache of the closed application out instead of keeping it in RAM. On a server this doesn't happen, you have your applications running continuously and no big memory eaters that don't run all the time - well there are exceptions, of course.
I don't suppose you are running on AMD64? The patch only affects Intel processor-based boxes.
Here chew this:
https://bugzilla.redhat.com/bugzilla/attachment.cgi?id=129064
My previous colleagues all think that RHEL4 or Centos 4 is slow which is true but due to the dumb vm patch Redhat applied to the RHEL4 kernel which is makes the box swap.
I do not access to those boxes now but I can tell you that vmstat will report swapping activity with constant non-zero figures under si and so. That is swapping. Hence, vmstat will also have high numbers under wa and id due to the apps all waiting for i/o on the RHEL4 box. The FC4 has completely different vmstat reports which are basically zero swapping, full utilization (cpu idle = 0) and zero or low wa numbers.
Redhat has admitted as much that the problem was their vm patch. Setting swappiness or whatever has zero bearing on the current RHEL4 kernel's knack of writing pages to swap even though there is plenty of RAM available.