[Gluster-users] DHT vs LVM for multiple bricks on a server

[Jeff Darcy]

On 01/11/2013 12:00 PM, Gaurav P wrote:
> Couldn't find the discussion you referenced in the irc logs going even a month
> back :-(

My usual answer is: it depends.  I've seen cases where using each disk as a 
separate brick performed better, and I've seen cases where combining them via 
LVM performed better.  There doesn't even seem to be a simple pattern to which 
will be faster for which workloads, though I'd say brick-per-disk probably wins 
slightly more often than not.  It will also have better failure characteristics 
than RAID0 - a point also brought up recently by the HDFS folks.

http://hortonworks.com/blog/why-not-raid-0-its-about-time-and-snowflakes/

They're characteristically wrong about having to wait for the slowest (might be 
true for their specific workload but not for most others as would be the case 
for RAID1) but make some other good points.

You also bring up the issue of bricks needing to be the same size.  This is 
kind of true right now.  It won't fail completely, but it also won't distribute 
files properly and that can lead to premature ENOSPC.  However, I expect that 
to be addressed fairly soon so you might want to consider that as you make your 
longer-term plans.