[Linux-cluster] CLVM and AoE

[David Golden]

On 2006-04-12 16:59:26 -0400, Bowie Bailey wrote:
> > - What kind of pitfalls should we be aware of?
> 

> Some people have complained about throughput issues with GFS.  Our
> application doesn't require high throughput, so I can't comment on
> this.  I haven't found any issues in my testing so far.
>

Well, a thing I _think_ we've seen a few times is that the case of many
simultaneous  writes to different files in different directories is MUCH faster
than many simultaneous writes to different files in the same directory. 
I think this may have been mentioned before on-list, IIRC it's a design trade-off,
something to do  with GFS's efforts to preserve strict unix-like
consistency (generally regarded as a major advantage of GFS over
the horrors of NFS), directory metadata about the files needs to be 
updated an awful lot in the same  directory case, and the
directory therefore needs to be locked for update an awful lot, 
which can lead to much slowdown.

I don't have hard numbers, nor available facilities to generate them
right now, so feel free to regard this as FUD.