The sonet algorithm [1] is used to generate a connectivity matrix enriched in second order connectivity motifs, which is subsequently used for simulation of brain activity during pattern separation or pattern completion [2]. For a model with 1 million neurons (approximately the number of dentate gyrus granule cells in the hippocampus), the current algorithm requires a shared memory computer. The algorithm has been already adapted in order to reduce the required amount of memory by a factor of four and improve the speed by optimizing memory access patterns [3, 4]. Despite these improvements, 4 Terabyte of RAM are needed to compute the connectivity matrix. Currently, in Austria no other computer provides a sufficient amount of memory for these types of computations.
Making the algorithm work on a distributed computer cluster is a non-trivial task, and its effort can not - within the current research project - be justified by the possible gains on speed and flexibility of use. On contrary, the approach of running the software on a big shared memory machine like Mach2 provides the fastest "time-to-result".
References