Why is a cloud model chosen instead of a compute cluster (like the Titan), running maybe GPGPU? Or to put it another way, what kind of workload is benefited by a cloud instead of a compute cluster?
Off the top of my head, I'm thinking that a system like this could simplify carving/sharing resources - elasticity seems to be a defining "cloud" characteristic. Also, adding an additional layer of abstraction could enable/simplify upgrade of the underlying infrastructure.
I arrived at CERN when the decisions about this were made already, so I'm mostly involved with the actual implementation.
I would say that the main reason is that a lot of experiments (not just the higgs boson search) are going on here right now. Somehow you need to distribute resources and let people do stuff with them and the centrally managed Openstack/Puppet stack vastly facilitates the task. The fact these solutions are open source and we can add our stuff (or others) on top of the code is also quite important for us.
Thanks for fielding questions - the article mentioned bare metal provisioning with OpenStack. Are you planning on using this https://wiki.openstack.org/wiki/Baremetal ? Another method I could envision would be using LXC's as the back end for compute instead of KVM, giving users self provisioning closer to the metal - I just wanted to clarify what method was planned.
Sounds like a fun opportunity! And yeah, that's fair - I doubt anybody has baremetal via OpenStack in production yet. Looks like it was added in the Grizzly release. I'll have to check this out.
10 comments
[ 2.8 ms ] story [ 34.3 ms ] threadOff the top of my head, I'm thinking that a system like this could simplify carving/sharing resources - elasticity seems to be a defining "cloud" characteristic. Also, adding an additional layer of abstraction could enable/simplify upgrade of the underlying infrastructure.
Let the downvoting galore begin.