Goals| Partners: |
|---|
GoalsA common diagnostic method in human brain research is Magnetoenzephalography (MEG). It can be used to localize electric current dipoles which represent areas of activity in the brain. The algorithm which is used for that contains two different parts which can be solved most efficiently on a massively parallel and a vector computer respectively.
In order to allow interactive response of the experimentalist, all this should be done during the measurement. In this case efficiency becomes an issue and the problem a candidate for heterogeneous metacomputing.
MagnetoenzephalographyIn the Institute of Medicine in the FZJ, several experiments on brain activity are performed. One of them is the Analysis of magnetoezephalography data. The magnetic field around a human head is measured with an array of superconducting quantum interference devices (SQUIDs). From these data, the distribution of electric currents in the brain can be reconstructed by solving an inverse problem. In Jülich, this is done with the 'Multiple Signal Classification' (MUSIC) algorithm. With MUSIC parameters of a finite number of current dipoles are obtained in three phases. The number of current dipoles is estimated in a preprocessing step (phase 0), the positions of the dipoles in phase 1, orientation and strength in phase 2. Phase 1 involves a nonlinear optimization, which is done best on a massively parallel system (MPP). Phase 2 is a least-squares problem, which is well-suited for a vector computer. Therefore, this application is a good candidate for heterogeneous metacomputing, because it can benefit from running distributed on different supercomputer architectures. On the other hand, it does not suffer from latency problems, since data has to be transfered in one direction only - from the MEG-device to the MPP running phase 1 and from there to the vector computer running phase 2.
StatusThree separate programs for the three phases of the reconstruction process have been implemented and tested on workstations. The parallelization of the nonlinear optimization problem (phase 1) is currently under way. It turned out that this phase is the most CPU-time consuming part of the algorithm. From the nature of the algorithm we expect a sufficiently good speedup on an MPP system to compensate that for.
