[Systems] Migrating a.sl.o to beamrider.

[Bernie Innocenti]

El Sun, 01-11-2009 a las 00:14 -0700, Ivan Krstić escribió:
> On Oct 31, 2009, at 5:13 AM, Bernie Innocenti wrote:
> > 4) (maybe) share a single memcached instance
> 
> Why not have instances on both machines?

Right... the overhead of network communication probably out-weights the
benefits of cache sharing.


> > - Instead of NFS, we could experiment with some of the fancy
> >   parallel networked filesystems.b Pohmelfs comes to mind.
> >   I'm currently evaluating DRBD and GFS2 for the FSF, but I'm
> >   not yet so confident about it.
> 
> Don't. These systems are a massive pain in the ass in production, and  
> almost never worth the added administration complexity and overhead.

I've always suspected it wasn't worth the pain in most cases, and now
I'm starting to see it in practice.

To be fair, DRBD looks really solid and more manageable than anything
else I've tried before. However, performance is a disappointment. I
haven't ran a proper benchmark suite yet, but merely running mkfs in the
drbd partition was painfully slow. And it's certainly not a networking
problem, because performance sucks even when no secondary nodes :-)

I think I'd agree with you: this sort of complexity is worth it only for
mission-critical applications that couldn't tolerate a couple of hours
of downtime in the rare case of major hardware failure.


> > - We should not use memcached
> 
> Why?

I meant to say "a shared memcached instance", which we don't need to
use.

Well, on second thought: the whole memcache thing seems designed to
enable cache sharing among multiple application servers, which may be
useful in some scenarios. If we don't care about this, then an
in-process cache like APC would be faster and easier to manage. Does
Remora support APC too?

-- 
   // Bernie Innocenti - http://codewiz.org/
 \X/  Sugar Labs       - http://sugarlabs.org/