Re: [Gluster-devel] Client memory footprint

[Gordan Bobic]

On Wed, 24 Aug 2011 11:50:45 +0200, address@hidden (Emmanuel Dreyfus) 
wrote:
> Pavan T C <address@hidden> wrote:
>
> > The glusterfs server process uses iomem pools. Initially, it starts 
> > with
> > one pool. Under memory pressure, it allocates more, but does not 
> > give
> > back the pool when the pressure decreases. It can be debated whether
> > this is a bug, but one can come up with heuristics to free the pool.
>
> I was concerned about machines with little memory. It would be nice 
> to
> have the ability to tell glusterfsd to avoid eating too much memory 
> for
> performances: if you do not have that memory, you clearly prefer to 
> have
> poor performance than to have a crash.

Indeed. One of the (to me at least) obvious use-cases for a distributed 
file system (GlusterFS with DHT) is micro-servers, which might have very 
small memory (512MB on a typical ARM server, 1GB if you're lucky), and a 
small amount of storage space (e.g. <= 32GB SD card) that would be handy 
to pool together, but for this to work a very tight memory footprint is 
needed. I'm not talking about just freeing memory as soon as it isn't 
used, I'm talking about revising what actually needs to be kept in 
memory at all.

Emmanuel, in the absence of a low-memory solution, you may want to look 
into zram (ramzswap before 2.6.32-ish) for swapping as a workaround.

Gordan