On 7 March 2011 14:56, Peter Grandi <[log in to unmask]> wrote:
>
> * Any experiences of using VServer (or OpenVZ) for Grid deployment?
> * What about Atom based server nodes? Except perhaps for CEs where X3400s
> might be required?
>
The only work I'm aware of on Atom-based servers for HEP is
http://indico.cern.ch/contributionDisplay.py?contribId=435&confId=35523
, which is somewhat negative (although it's also more concerned about
using them for WNs, rather than front-end servers).
Note, however, that CEs are certainly not going to be a good idea on
Atom, just from looking at our server loads. So, what other services
do you have? SE head nodes need to be able to handle a high-ish CPU
load in principle (although not generally very often, now that VOs
have generally improved their data models). (Batch servers like PBS
also shouldn't go on an Atom service, for obvious reasons). I don't
think Site-BDIIs or the APEL publishing service use much CPU at all,
though, so they'd be potential candidates for any kind of low-CPU use
solution (which is why our APEL service is sitting on a VM right now).
On the other hand, this breakdown looks like you'd be buying two Atom
servers for two services, which doesn't necessarily look like you're
making much of a saving (especially since you need at least one CE and
one SE in addition to them).
I also feel I should note that Tier-2 sites are not expected to have
Tier-1 levels of resiliency or redundancy, which appears to be what
you're trying to approach. ;)
Sam
|