Improving scalability of large-scale distributed Spiking Neural Network simulations on High Performance Computing systems using novel architecture-aware streaming hypergraph partitioning

After theory and experimentation, modelling and simulation is regarded as the third pillar of science, helping scientists to further their understanding of a complex system. In recent years there has been a growing scientific focus on computational neuroscience as a means to understand the brain and its functions, with large international projects (Human Brain Project, Brain Activity Map, MindScope and \textit{China Brain Project}) aiming to further our knowledge of high level cognitive functions. They are a testament to the enormous interest, difficulty and importance of solving the mysteries of the brain. Spiking Neural Network (SNN) simulations are widely used in the domain to facilitate experimentation. Scaling SNN simulations to large networks usually results in more-than-linear increase in computational complexity. The computing resources required at the brain scale simulation far surpass the capabilities of personal computers today. If those demands are to be met, distributed computation models need to be adopted, since there is a slow down of improvements in individual processors speed due to physical limitations on heat dissipation. This is a significant change that requires careful management of the workload in many levels: partition of work, communication and workload balancing, efficient inter-process communication and efficient use of available memory. If large scale neuronal network models are to be run successfully, simulators must consider these, and offer a viable solution to the challenges they pose. Large scale SNN simulations evidence most of the issues of general HPC systems evident in large distributed computation. Commonly used distribution of workload algorithms (round robin, random and manual allocation) do not take into consideration connectivity locality, which is natural in biological networks, which can lead to increased communication requirements when distributing the simulation in multiple computing nodes. State-of-the-art SNN simulations use dense communication collectives to distribute spike data. The common method of point to point communication in distributed computation is through dense patterns. Sparse communication collectives have been suggested to incur in lower overheads when the application's pattern of communication is sparse. In this work we characterise the bottlenecks on communication-bound SNN simulations and identify communication balance and sparsity as the main contributors to scalability. We propose hypergraph partitioning to distribute neurons along computing nodes to minimise communication (increasing sparsity). A hypergraph is a generalisation of graphs, where a (hyper)edge can link 2 or more vertices at once. Coupled with a novel use of sparse-aware communication collective, computational efficiency increases by up to 40.8 percent points and simulation time reduces by up to 73\%, compared to the common round-robin allocation in neuronal simulators. HPC systems have, by design, highly hierarchical communication network links, with qualitative differences in communication speed and latency between computing nodes. This can create a mismatch between the distributed simulation communication patterns and the physical capabilities of the hardware. If large distributed simulations are to take full advantage of these systems, the communication properties of the HPC need to be taken into consideration when allocating workload to route frequent, heavy communication through fast network links. Strategies that consider the heterogeneous physical communication capabilities are called architecture-aware. After demonstrating that hypergraph partitioning leads to more efficient workload allocation in SNN simulations, this thesis proposes a novel sequential hypergraph partitioning algorithm that incorporates network bandwidth via profiling. This leads to a significant reduction in execution time (up to 14x speedup in synthetic benchmark simulations compared to architecture-agnostic partitioners). The motivating context of this work is large scale brain simulations, however in the era of social media, large graphs and hypergraphs are increasingly relevant in many other scientific applications. A common feature of such graphs is that they are too big for a single machine to cope, both in terms of performance and memory requirements. State-of-the-art multilevel partitioning has been shown to struggle to scale to large graphs in distributed memory, not just because they take a long time to process, but also because they require full knowledge of the graph (not possible in dynamic graphs) and to fit the graph entirely in memory (not possible for very large graphs). To address those limitations we propose a parallel implementation of our architecture-aware streaming hypergraph partitioning algorithm (HyperPRAW) to model distributed applications. Results demonstrate that HyperPRAW produces consistent speedup over previous streaming approaches that only consider hyperedge overlap (up to 5.2x speedup). Compared to multilevel global partitioner in dense hypergraphs (those with high average cardinality), HyperPRAW is able to produce workload allocations that result in speeding up runtime in a synthetic simulation benchmark (up to 4.3x). HyperPRAW has the potential to scale to very large hypergraphs as it only requires local information to make allocation decisions, with an order of magnitude less memory footprint than global partitioners. The combined contributions of this thesis lead to a novel, parallel, scalable, streaming hypergraph partitioning algorithm (HyperPRAW) that can be used to help scale large distributed simulations in HPC systems. HyperPRAW helps tackle three of the main scalability challenges: it produces highly balanced distributed computation and communication, minimising idle time between computing nodes; it reduces the communication overhead by placing frequently communicating simulation elements close to each other (where the communication cost is minimal); and it provides a solution with a reasonable memory footprint that allows tackling larger problems than state-of-the-art alternatives such as global multilevel partitioning