Auto-tuning for neural simulations on different parallel hardware

由于网络活动与总计算时间的依赖性、网络内部连通性的时间演化以及网络的异构性，生物逼真神经模拟对并行计算提出了新的挑战。自动调优技术可以帮助使用代码生成在不同级别（算法、数据结构）选择优化。在这个项目中，我们将研究在线和离线自动调优方法，以提高不同硬件（共享内存，分布式或基于gpu的系统）上神经模拟的并行效率，特别关注性能测量的改进，数据结构的分析，使用图分区的数据分布，启发式学习以限制搜索空间和参数化代码生成在运行时为特定网络。所得到的机制将集成到现有的神经模拟器中。

Biologically-realistic neural simulations pose new challenges to parallel computing because of the dependency between the network activity and the total computation time, the temporal evolution of their internal connectivity and the heterogeneity of the networks. Auto-tuning techniques can help selecting optimizations at various levels (algorithmic, data structures) using code generation. In this project, we will investigate online and offline auto-tuning methods to increase the parallel efficiency of neural simulations on different hardware (shared-memory, distributed or GPU-based systems), focusing particularly on the improvement of performance measurements, the analysis of data structures, distribution of data using graph partitioning, the learning of heuristics to restrict the search space and parameterized code generation at run-time for a particular network. The resulting mechanisms will be integrated in an existing neural simulator.