[Linux-cluster] problems with 8 node production gfs cluster

[nemanja at yu.net]

We are using sendmail/procmail and popa3d for pop3.


Quoting Rainer Duffner <rainer at ultra-secure.de>:

> Nemanja Miletic wrote:
>> Hello,
>>
>> I am working for major ISP and we have gfs cluster deployed for our mail
>> system. Cluster includes six smtp and two pop3 nodes. Nodes are on blade
>> servers, and are accessing FC HP EVA 5000 storage. Whole cluster is on
>> the same gigabit subnet. Partition that holds mailboxes is shared. We
>> are using gentoo linux distribution with 2.6.16-gentoo-r1 kernel and
>> 1.02 linux-cluster.  We have a problem that in busy hours reading 
>> and writing from gfs
>> partition gets very slow. This causes many processes that need to use
>> the disk to go in D state. This further causes great load (over 200) on
>> the machines in the cluster. When we cut the pop3 and smtp access on the
>> firewall load slowly decreases. At the moment we have a limit on syn
>> connection on 110 port on our load balancers (LVS based) in order to
>> control the load.
>>
>>
>>
>> Thank You,
>>
>
>
>
> Out of personal interest: what MTA/MDA are you running?
>
>
>
> cheers,
> Rainer
>