Federated cloud for Swedish National facility of Magnetoencephalography

The idea behind the integration with Karolinska NATMEG cloud and PDC Cloud is to offload the computation and storage in an Elastic way. For doing this we are planning to use OpenStack on both ends and storage will be done using IRODs for archival purposes.

First of all, the scientific analysis of MEG data is very interactive. In two ways; first of all in a literal sense of dealing with different visualizations of our data during analysis, with the researcher interacting with the data using e.g. mouse clicks or some kind of visual interface (may it be Matlab or a python interface). In fact, most of the different stages of an analysis require visual inspection of the data at some point.

natmag2

Secondly, the interaction involves going back and forth in the analysis pipeline, adjusting parameters and trying out difference branches of analysis. A standard analysis pipeline does not really exist, although there are parts that might be standardized. There are computations that are repeated many times (e.g. frequency analysis over time-points and channels), but they are almost always embedded in the above-mentioned interactive, or exploratory analysis.
Lastly, the real bottleneck, or challenge in analyzing MEG data is memory rather than CPU. Although I do not underestimate the benefit of high computing power, a limitation in CPU can always be resolved through time (waiting longer for a process to finish) and although it might reduce the output of a project somewhat, it is surmountable. A lack of memory, however, can really limit the analysis in a way that cannot be otherwise resolved. Note, that the dataset we work with are often a minimum of 2 Gig, and during analysis can easily increase many times over.
FieldTrip on Matlab platform. For Fieldtrip’s current and previous efforts on distributed computing (which does not use Matlab’s parallel computing toolbox), see:
Freesurfer
MNE (uses Python)
Possibly SPM