R: [Linux-cluster] Performance tuning help sought

[Leandro Dardini]

> -----Messaggio originale-----
> Da: linux-cluster-bounces at redhat.com 
> [mailto:linux-cluster-bounces at redhat.com] Per conto di 
> Kadlecsik Jozsef
> Inviato: martedì 4 marzo 2008 11.35
> A: linux clustering
> Oggetto: [Linux-cluster] Performance tuning help sought
> 
> Hello,
> 
> We have some performace problem in our GFS cluster. It 
> manifests mostly in slow folder access in mail clients (pine, 
> mutt) running on the nodes. When switching from one 
> mailfolder to another, sometimes a large folder is loaded 
> quickly - but usually a small folder, or even paging in the 
> same folder takes a noticeable time. Looking at the 
> performance tips of the GFS FAQ:
> 
> - we can neither preallocate files nor disable quota
> - the cluster is alreay connected over 1Gb Ethernet
> - statfs_slots is increased to 128
> - dunno how to increase or decrease the cached GFS locks:
> 	- there is no /proc/cluster/ directory
> 	- how can one actually monitor the number of cached locks?
> - the size of the filesystem is 1TB and was created with the default
>   resource groups size. So recreating the filesystem with increased
>   RGs size is one possible tuning option, as far as I see.
> - breaking up the filesystem into smaller ones would make life much
>   harder for us, so if there are other possibilities we'd 
> better avoid it
> 
> What can we do, besides increasing the RGs size? 
> Is there any kernel parameter, which could be tuned (say vm 
> parameters)?
> Would it help to set higher MTU values for openais? (But how 
> could one set
> it?)
> 
> Best regards,
> Jozsef

To speed up things a little, try mounting your partition without atime (noatime). This will prevent contention locking on directory.

Leandro