Brain simulation is making giant leaps towards a better understanding of the inner workings of the human brain.

Brain simulation is making giant leaps towards a better understanding of the inner workings of the human brain. The state-of-the-art simulations in this field are still far away from simulating the whole brain, which is their ultimate goal. Reaching that target without Exascale, or even higher computational power, is not possible.

In our first project DEEP, the Swiss Federal Institute of Technology in Lausanne (EPFL), adapted CoreNeuron – an advanced brain simulation application – to run efficiently on the DEEP prototype.

Multiple optimisations were implemented. For instance, regarding threading, an elaborated load balancing strategy at the thread level was implemented. This strategy takes into account the complexity of simulating different kinds of neurons, and allows for linear scalability to a very large number of threads. Regarding vectorisation, data layout changes and extensive refactoring of the main loops yield an improvement of 5x over the original version.

The above changes allow achieving extremely good scalability and performance for simulations with a large number of neurons. This makes the bulk of the application run efficiently on the DEEP Booster. Given that I/O was done more efficiently on the DEEP Cluster, EPFL decided to start the simulation on the Booster, and offload the I/O to the Cluster. In this way, I/O was improved by more than an order of magnitude (over doing it directly from the Booster). More importantly, this lays out the foundation for interactive supercomputing in the future, showing that the DEEP architecture is suitable and aligned with coming changes needed in the brain simulation community.