[Beowulf] diskless cluster nfs

[Ariel Sabiguero]

Josh Kayse wrote:

>Ok, my first post, so please be gentle.
>  
>
bienvenido!

>Basic setup follows:
>
>Master node: system drive is 1 36GB SCSI drive
>                     /home raid5 5x 36GB SCSI drives
>Master node exports /tftpboot/192.168.1.x for the nodes.
>  
>
I would also recomend you to have some sort of fault tolerance on the / 
filesystem.

>The engineers use files over nfs for message passing, and no, they
>will not change their code to mpi even though it would be an
>improvement in terms of manageability and probably performance.
>  
>
I do not know what are your "messages" like, but how "big" are they?
If size is not a problem you might want to create a ramfs drive and 
export it for message-passing.
Maybe an extra GB of ram at the server solves your problem and messages 
never get to the disk.
You might want to split what is message passing and application data 
this way.
Maybe you need to recode "something" to use different directories for 
messages and data (my apologies for those engineers who don't like 
recoding!), but if the concept of "message" that the application uses is 
not the one of a petabyte-message this may help.

>Basically, my question is:  what are some ways of testing the
>performance of nfs ande then, how can I improve the performance?
>  
>
I believe that previous messages on the list gave you great clues on how 
to speed-up the network (MTU+jumboframes), nfs and the underlying raid.
My approach is different: why to persist something volatile?
Any way, MTU, jumboframes and nfs stuff is also worth trying here.

>Thanks for any help in advance.
>  
>
Just remember helping-back in the future ;-)

Ariel

>PS: nfs mount options: async,rsize=8192,wsize=8192,hard
>       file sizes: approx 2MB
>  
>