Resumen de MPI layer techniques to improve network energy efficiency
Branimir Dickov
Interconnection networks represent the backbone of large-scale parallel systems. In order to build ultra-scale supercomputers larger interconnection networks are being designed and deployed. As compute nodes become more energy-efficient, the interconnect is accounting for an increasing proportion of the total system energy consumption. The interconnect's energy consumption is, however, only starting to receive serious attention. Most of this power consumption is due to the interconnection links. The problem, in terms of power, of an interconnect link is that its power consumption is almost constant, whether or not it is actively exchanging data, since both ends stay active to mantain synchronization. This thesis complements ongoing efforts related to power reduction and energy proportionality of the interconnection network. The thesis contemplates two directions for power savings in the interconnection network; one is the possibility to use lower bandwidth links during the communication phases and thus save energy, while the second one addresses shifting links to low-power mode during computation phases when they are unused. To address the first one we investigate the potential benefits from MPI data compression. When compression of MPI data is possible, the reduction in link bandwidth is enabled without incurring any performance penalty. Consecutively, lower bandwidth leads to lower link energy consumption. In the past, several compression techniques have been proposed as a way to improve the performance and scalability of parallel applications. Those works have shown significant speed-ups when applying compressors to the MPI transfers of certain algorithmic kernels. However, these techniques have not seen widespread adoptation in current supercomputers. In this thesis we will show that although data compression naturally leads to improved performance, the benefit is small, for modern high-performance networks, and it varies greatly between applications. In contrast, combining data compression with switching to low-power mode preserves performance while delivering effective and consistent energy savings, in proportion with the reduction in data rate. In general, application developers view time spent in a communication as an overhead, and therefore strive to keep it at minimum. This leads to high peak bandwidth demand and latency sensitivity, but low average utilization, which provides significant opportunities for energy savings. It is therefore possible to save energy using low-power modes, but link wake-up latencies must not lead to a loss in performance. Thus, we propose a mechanism that can accurately predict when links are idle, allowing them to be switched to more power efficient mode. Our runtime system called the Pattern Prediction System (PPS) can accurately predict not only when a link will become unused but also when it will become active again, allowing links to be switched off during the idle periods and switched back on again in time to avoid incurring a significant performance degradation. Many HPC application benefit from prediction, since they have repetitive computation and communication phases. By implementing the energy-saving mechanisms inside the MPI library, existing MPI programs do not need to be modified. We also develop more advanced version of the prediction system, Self-Tuned Pattern Prediction System (SPPS) which is capable of automatically tuning to the current application communication characteristic and shaping the switching on/off of the links in the most appropriate way. The proposed compression and prediction techniques are evaluated using an event-driven simulator, which is able to replay the traces from real execution of MPI applications. Experimental results show significant energy savings in the IB links while the performance overhead due to wake-up latencies and additional computation time have negligible effects on the final application performance.
