[Linux-cluster] GFS

[Jonathan E Brassow]

On Oct 28, 2004, at 11:11 AM, Mohamed Magdi Abbas wrote:

> David Teigland wrote:
>> On Tue, Oct 26, 2004 at 04:41:29PM -0400, Dascalu Dragos wrote:
>>> We are working on a similar scenario but adding mailman into the 
>>> mix. The ideal outcome would be for multiple mailman/postfix servers 
>>> to write archives, etc to the same centralized location on a SAN. 
>>> After doing some tests this setup does not appear to be trivial. We 
>>> ran into a similar problem when using NFS; if multiple machines 
>>> write to the same file at the same time the file gets mangled as the 
>>> machines cut each other off. With GFS we noticed that each machine 
>>> has a 4k buffer window in which it writes its data. If a second 
>>> process decides to start writing to the same file we noticed 
>>> alternating writes to the file after 4k of data.
>> Note that this sounds like perfectly correct behavior on the part of 
>> gfs.
>> The application is responsible for the necessary file locking, of 
>> course,
>> while gfs is responsible for keeping the fs uncorrupted.
>
> I thought the idea of GFS is that it would handle locking to enable 
> shared filesystems among different nodes with simultaneous r/w access 
> to the filesystem.
>

Think of it as you would two processes on the same node writing to the 
same file.  The local file system (e.g. ext[23]) will not guarantee the 
outcome.  Similarly, two processes on different machines using the same 
file system (e.g. GFS) will not guarantee the outcome.  It has always 
been up to processes concurrently writing to the same file to ensure 
the desired result is achieved.

  brassow