High performance communication on reconfigurable clusters

High Performance Computing (HPC) has matured to where it is an essential third pillar, along with theory and experiment, in most domains of science and engineering. Communication latency is a key factor that is limiting the performance of HPC, but can be addressed by integrating communication into accelerators. This integration allows accelerators to communicate with each other without CPU interactions, and even bypassing the network stack. Field Programmable Gate Arrays (FPGAs) are the accelerators that currently best integrate communication with computation. The large number of Multi-gigabit Transceivers (MGTs) on most high-end FPGAs can provide high-bandwidth and low-latency inter-FPGA connections. Additionally, the reconfigurable FPGA fabric enables tight coupling between computation kernel and network interface. Our thesis is that an application-aware communication infrastructure for a multi-FPGA system makes substantial progress in solving the HPC communication bottleneck. This dissertation aims to provide an application-aware solution for communication infrastructure for FPGA-centric clusters. Specifically, our solution demonstrates application-awareness across multiple levels in the network stack, including low-level link protocols, router microarchitectures, routing algorithms, and applications. We start by investigating the low-level link protocol and the impact of its latency variance on performance. Our results demonstrate that, although some link jitter is always present, we can still assume near-synchronous communication on an FPGA-cluster. This provides the necessary condition for statically-scheduled routing. We then propose two novel router microarchitectures for two different kinds of workloads: a wormhole Virtual Channel (VC)-based router for workloads with dynamic communication, and a statically-scheduled Virtual Output Queueing (VOQ)-based router for workloads with static communication. For the first (VC-based) router, we propose a framework that generates application-aware router configurations. Our results show that, by adding application-awareness into router configuration, the network performance of FPGA clusters can be substantially improved. For the second (VOQ-based) router, we propose a novel offline collective routing algorithm. This shows a significant advantage over a state-of-the-art collective routing algorithm. We apply our communication infrastructure to a critical strong-scaling HPC kernel, the 3D FFT. The experimental results demonstrate that the performance of our design is faster than that on CPUs and GPUs by at least one order of magnitude (achieving strong scaling for the target applications). Surprisingly, the FPGA cluster performance is similar to that of an ASIC-cluster. We also implement the 3D FFT on another multi-FPGA platform: the Microsoft Catapult II cloud. Its performance is also comparable or superior to CPU and GPU HPC clusters. The second application we investigate is Molecular Dynamics Simulation (MD). We model MD on both FPGA clouds and clusters. We find that combining processing and general communication in the same device leads to extremely promising performance and the prospect of MD simulations well into the us/day range with a commodity cloud