Previous: Results on IBM
A Car-Parrinello ab-initio molecular dynamics method used hitherto on workstations and vector-computers has been parallelized using a master-slave model with message-passing. A fairly simple parallel algorithm based on farming a modest number of k-points out to slave processors has been used, complementary to other parallel Car-Parrinello algorithms. The memory savings are significant in our algorithm, since each processor only holds the wavefunction array for a single (or a few) k-points .
We find that k-point -parallel algorithms are non-trivial because of the changed convergence properties, owing to changes in the way the potential is screened. Updating the charge density after each band has been treated (in parallel) makes the algorithm stable.
The speedups measured for a test problem show satisfactory results for up to about 6 processors (depending on the problem at hand), which is more than adequate for our large-scale production jobs.
The timings obtained on an IBM SP2 show that the present parallel algorithm is bound by the communication bandwidth between the master processor and its slaves. Two options are identified to alleviate this bottleneck: 1) the investigation of less communication intensive algorithms, and 2) the efficient implementation of global reduction and broadcast operations within the message passing library.