**5. Conclusions**

In this work, we presented an implementation of an MPI-based DNA sequence matching algorithm for evaluating two critical aspects of using one of the more promising neuromorphic emerging technology. As the first point, we benchmarking the SpiNNaker many-core neuromorphic platform and its MPI support, showing that the scaling performances are kept linear when an increasing number of cores is used during the computation. As the second point, we demonstrated that by using the spinMPI library, which provides MPI support for SpiNNaker, we could easily port algorithm implemented for standard computers on the many-core neuromorphic platform.

The MPI standard exposes a programming model for the development of parallel applications in a distributed memory environment without knowledge of the interconnections between the computing units of the underlying architecture. The implementation of MPI for a specific architecture is therefore expected to implement the most suitable features in order to exploit the available resources and to synchronise the computing flow.

In the case of SpiNNaker, the implementation of MPI must deal with a resource limit both in terms of memory and computing power. However, it can take advantage of the technology offered by on-chip routers, obtaining efficient communication. SpinMPI is also in charge of managing communication between the *MPI Runtime* running on the host computer and the SpiNNaker cores; this is done by using the ACP protocol and memory entities. This software stack creates a simple working framework offering a universally known programming model capable of making the SpiNNaker architecture available for a wide range of applications.

We have succeeded in performing a benchmark of the SpiNNaker board by using a highly-parallel implementation of a DNA matching algorithm. Results show that the scalability of the SpiNNaker board reaches an ideal profile (98% of efficiency) when using more than 100 processors, a 90% efficiency using 600 processors, reaching 88% efficiency when all 767 application processors are used.

**Author Contributions:** Conceptualization, G.U. and F.B.; methodology, G.U., F.B. and A.A.; software, E.F., E.P. and F.B.; validation, E.F., E.P. and F.B.; formal analysis, E.F., E.P. and F.B.; investigation, G.U., E.F., E.P.; resources, G.U., A.A., and E.M.; data curation, E.P.; writing–original draft preparation, G.U., E.F., E.P. and F.B.; visualization, G.U., E.F. and E.P.; supervision, G.U.; project administration, G.U., A.A., and E.M.; funding acquisition, G.U. and E.M.

**Funding:** This research was funded by European Union Horizon 2020 Programme [H2020/2014-20] grant number 785907.

**Acknowledgments:** The research leading to these results has received funding from European Union Horizon 2020 Programme [H2020/2014-20] under grant agreement no. 785907 [HBP-SGA2].

**Conflicts of Interest:** The authors declare no conflict of interest.
