Accelerating Network Communication and I/O in Scientific High Performance Computing Environments

High performance computing has become one of the major drivers behind technology inventions and science discoveries. Originally driven through the increase of operating frequencies and technology scaling, a recent slowdown in this evolution has led to the development of multi-core architectures, which are supported by accelerator devices such as graphics processing units (GPUs). With the upcoming exascale era, the overall power consumption and the gap between compute capabilities and I/O bandwidth have become major challenges. Nowadays, the system performance is dominated by the time spent in communication and I/O, which highly depends on the capabilities of the network interface. In order to cope with the extreme concurrency and heterogeneity of future systems, the software ecosystem of the interconnect needs to be carefully tuned to excel in reliability, programmability, and usability. This work identifies and addresses three major gaps in today's interconnect software systems. The I/O gap describes the disparity in operating speeds between the computing capabilities and second storage tiers. The communication gap is introduced through the communication overhead needed to synchronize distributed large-scale applications and the mixed workload. The last gap is the so called concurrency gap, which is introduced through the extreme concurrency and the inflicted learning curve posed to scientific application developers to exploit the hardware capabilities. The first contribution is the introduction of the network-attached accelerator approach, which moves accelerators into a "stand-alone" cluster connected through the Extoll interconnect. The novel communication architecture enables the direct accelerators communication without any host interactions and an optimal application-to-compute-resources mapping. The effectiveness of this approach is evaluated for two classes of accelerators: Intel Xeon Phi coprocessors and NVIDIA GPUs. The next contribution comprises the design, implementation, and evaluation of the support of legacy codes and protocols over the Extoll interconnect technology. By providing TCP/IP protocol support over Extoll, it is shown that the performance benefits of the interconnect can be fully leveraged by a broader range of applications, including the seamless support of legacy codes. The third contribution is twofold. First, a comprehensive analysis of the Lustre networking protocol semantics and interfaces is presented. Afterwards, these insights are utilized to map the LNET protocol semantics onto the Extoll networking technology. The result is a fully functional Lustre network driver for Extoll. An initial performance evaluation demonstrates promising bandwidth and message rate results. The last contribution comprises the design, implementation, and evaluation of two easy-to-use load balancing frameworks, which transparently distribute the I/O workload across all available storage system components. The solutions maximize the parallelization and throughput of file I/O. The frameworks are evaluated on the Titan supercomputing systems for three I/O interfaces. For example for large-scale application runs, POSIX I/O and MPI-IO can be improved by up to 50% on a per job basis, while HDF5 shows performance improvements of up to 32%