Development for LHC@home

Over the past few months we have been re-engineering our internal infrastructure so that we can send real CMS simulations jobs to the CMS@home project and are nearly ready to try this out for real. From the VM side of things, we only need to update the CMSJobAgent.py script which will be done via the magic of CVMFS [1], so no new application release will be needed.

Although this should be transparent, there is a high chance we will temporary break something, so please bear with us during this potential period of instability. I would estimate that we will be ready within the next two weeks and I will send an announcement before we make any changes.

Many tanks to all of you who have been supporting with the testing of this project.

[1] http://iopscience.iop.org/1742-6596/219/4/042003

Since we will be crunching real tasks soon, will we also be getting graphics similar to the ones that Atlas@home provides? Also, is there Nvidia Cuda support planned in the future? What about Android support? I have six android devices and four graphics cards just begging to compute for CMS.

Since we will be crunching real tasks soon, will we also be getting graphics similar to the ones that Atlas@home provides? Also, is there Nvidia Cuda support planned in the future? What about Android support? I have six android devices and four graphics cards just begging to compute for CMS.

I can't speak for the graphics, though naturally CMS has a range of visualisation techniques. I have heard of some work being done with CUDA, and ARM (though not necessarily Android), but it's not yet mainstream for our software, which is called CMSSW. Now, it's not hard to get a look at the CMSSW repository, even if you're not a Collaboration member; if you ever do, you'll see the enormous task necessary to port to other operating systems (e.g. OSX), let alone other paradigms such as GPUs or Xeon Phi. As well as our own software, we also have to carry along lots of other externals in specific versions to maintain our release integrity, e.g. gcc, ROOT,geant4, pythia, herwig, python, curl, and fftw3 to name but a few of the more important ones -- I count 160 externals in a recent release area.
Then, of course, the software has to run in a VM which supports the host (and guest) system, so maybe you should be asking Oracle what plans they have to release GPU and Android versions of VirtualBox.

---

I am ready (of course) but the only difference I would use is more than one task at a time like I do with vLHC or Atlas since all of my GPU's are taken (Einstein)

---

Mad Scientist For Life

As Ivan mentioned, efforts are already underway to run HEP code on GPUs. However, it is not clear whether you can do this from within a VM. This is something that we may consider investigating in the future but first we have to get this project up and running! The advantage of using a VM is that we can easily run anywhere that support virtualization so if a suitable hypervisor exists for Android, we in theory could run there.

As Ivan mentioned, efforts are already underway to run HEP code on GPUs. However, it is not clear whether you can do this from within a VM. This is something that we may consider investigating in the future but first we have to get this project up and running! The advantage of using a VM is that we can easily run anywhere that support virtualization so if a suitable hypervisor exists for Android, we in theory could run there.

Using GPUs in a VM is problematic. You need a virtualisation product which supports a feature called "GPU passthrough". NVidia has a page on the subject: http://www.nvidia.com/object/dedicated-gpus.html, which says that NVidia GPUs can be used with Citrix XenDesktop or VMware Horizon View. I expect AMD will have a similar list for their GPUs.

Unfortunately, the BOINC virtual machine interface - originally developed at CERN, I believe - only supports Oracle VirtualBox, which doesn't currently support GPU passthrough. So, unless either Oracle upgrades VB, or BOINC interfaces with a second VM technology, this doesn't look like a starter.

Exactly. We either will need to re-write the job wrapper for another hypervisor or work with Oracle so that VirtualBox provides the feature.

Still going to be CMS Simulation v46.15 (vbox64) ?

Or a change soon?