[Gluster-users] dbench & similar - as a valid benchmark

[Jeff Darcy]

On Wed, 25 Apr 2012 17:20:13 +0100
lejeczek <peljasz at yahoo.co.uk> wrote:

> would a tool such as dbench be a valid bechmark for gluster?

Dbench is a pretty good tool.  As long as you use the fileio back end,
your loadfiles should work on GlusterFS just fine.

> and, most importantly, is there any formula to estimate raw 
> fs to gluster performance ratio for different setups?
> for instance:
> having a replicated volume, two bricks, fuse mountpoint to 
> volume via non-congested 1Gbps
> or even
> a volume on single brick with fuse client mountpoing locally
> 
> what percentage/fraction of raw filesystem performance 
> should we expect from gluster? roughly?

As I'm sure you know, deriving system or application performance from
component performance is an exercise in dealing with chaos.  There's
certainly no formula for it, and even sophisticated models usually
can't overcome the fact that very small differences in how the parts
interact - e.g. latency distributions or queuing behavior - can result
in large changes to the result.  *In general* your network is going to
be the main factor affecting bandwidth, and for small numbers of disks
they're going to govern latency.  To know how that affects your
application, you'd have to know whether it's bandwidth- or
latency-bound, and how well it can take advantage of parallelism.