Re: nfs server performance (netbsd-5)

[Matthias Scheler <tron%zhadum.org.uk@localhost>]

On Wed, Apr 07, 2010 at 08:21:56PM +0900, Izumi Tsutsui wrote:
> > UDP mount and rsize/wsize=32768 is ambitious...
> 
> Isn't it default on i386 in sys/arch/i386/include/param.h?
> 
> #ifndef NFS_RSIZE
> #define NFS_RSIZE     32768
> #endif
> #ifndef NFS_WSIZE
> #define NFS_WSIZE     32768
> #endif
> 
> I wonder if using sizes without fragments is better
> for modern NICs with TCP/UDP hardware checksums.

Large writes on TCP connections allow a modern NIC to use TSO (TCP
Segment Offload).

Small write sizes also mean more overhead caused by RPC and more
resources used on the NFS server because it needs to handle
more individual requests.

        Kind regards

-- 
Matthias Scheler                                  http://zhadum.org.uk/