NFS scalability

[steve]

On Mon, Sep 10, 2007 at 07:49:51AM -0700, Matthew Dillon wrote:
> 
> :Just out of interest (without the time/boxes to try it out myself), do 
> :you guys think this is a "good" solution? I mean, does NFS handle a lot 
> :concurrent accesses well? I never used NFS...
> :
> :Hehe, I know, using HAMMERFS would be a terrible good solution ;-)
> :
> :Regards,
> :
> :   Michael
> 
>     NFSv3 has no real cache coherency between clients.  Theoretically
>     NFSv4 can but I think the concept of using a centralized filesystem
>     in a cluster is a bad idea generally even if it is the easiest 
>     solution at the moment.

	Easiest, possibly the only for large flexible systems right
now.
 
	In theory, NFS has always problematic.

	In practice, it is very possible to develop workable solutions
in which hundreds of NFS clients mount dozens of NFS servers serving
terabytes of data (the servers serving data from disk, are "diskless"
also, ie. PXE boot/nfs). Our system uses FreeBSD 4.10 servers
with 6.2 client front ends. UDP/NFS. 900 days uptime on NFS servers.

	Yes there are issues that need to be solved. 
And yes architecture does matter.

	Its not perfect, nothing is; but nothing that stops it from
"working" well enough to serve lots of data and lots of websites.

	Looking forward to "real" clustering though.

	-steve

> 
> 					-Matt
> 					Matthew Dillon 
> 					<dillon at backplane.com>
>