[CentOS] kswapd taking 100% cpu with no swap on system

[Peter Kjellström]

Please don't top-post it's hard to follow.

On Monday, May 09, 2011 03:07:57 PM Ali Ahsan wrote:
> Thanks for reply,I dont know it has some thing to do with linux
> scheduler  but some on suggested about linux scheduler cat
> /sys/block/sda/queue/scheduler there are quite few of them noop
> [anticipatory] deadline cfq.I have changed from cfq default to
> anticipatory,Now problem is less for me kswapd0  take 10-50% of cpu in
> very short bust less then milliseconds.Also i have changed
> /proc/sys/vm/pagecache from 100 to 10.

This means that you're limiting the amount of RAM that can be used as 
pagecache (basically caching I/O) to 10%. This may not be an issue for you but 
may also severely limit your performance (all depending on work load).

An alternative may be to simple set swappiness to 0.

> Any more suggestion will help me
> make server better,As for i have read the best scheduler for databases
> are deadline like for mysql it worked great.

You'll really have to try them to know what works for you. The two main 
variables that'll affect the choice is your I/O load and the device under 
(type of raid for example). For example some raid controllers really prefer 
noop.

Lastly, you didn't tell us if you're on i386-pae or x86_64 which may make a 
big difference.

/Peter
 
> On 05/09/2011 06:01 PM, Peter Kjellström wrote:
> > On Saturday, May 07, 2011 09:35:48 PM Ali Ahsan wrote:
> >> Hi All
> >> 
> >> I have xeon server with 16 Gb Ram and no Swap memory.I am running
> >> cassandra server on two node in cluster.When there is high load on
> >> server kswapd0 kicks inn and take 100% cpu and make machine very slow
> >> and we need to restart out cassandra server.I have  latest kernel
> >> 2.6.18-238.9.1.el5.Please let me know how can i fix this issue .Its
> >> hurting us badly this our production server any quick help will be
> >> appreciated.
> > 
> > There is more than one bug that causes this behaviour. A few related
> > memory managent situations (possibly responsible) may actually be
> > avoided if you add some swap (even if it's not used). My suggestion
> > would be to add some swap, set swappiness to 0 and see what happens.
> > 
> > /Peter
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 198 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.centos.org/pipermail/centos/attachments/20110509/fabce364/attachment.sig>