Analyzing the Energy Consumption of Synchronous and Asynchronous Checkpointing Strategies

At the start of the exascale computing age, the number of components installed in high performance computing (HPC) systems has increased by more than 70 percent, leading to a shorter mean time between failure (MTBF) and large power budgets approaching and exceeding 20 MWs. In this context, resilience strategies such as checkpoint/restart (C/R), which are traditionally adopted by tightly coupled HPC applications, face the need to become energy efficient, in addition to the need to minimize the runtime overheads at scale. With respect to the latter aspect, C/R has evolved to employ asychronous multi-level resilience techniques that leverage heterogeneous storage. Thus, there are many factors that affect the performance and scalability of C/R. The impact of these factors on energy efficiency and the resulting trade-offs have been studied only to a limited degree so far in the literature. To address this gap, we propose to study two state-of-the-art C/R libraries, VELOC (local checkpoints to DRAM or SSD followed by asynchronous flushes to PFS) and GenericIO (optimized synchronous flushes to PFS). We perform weak and strong scalability experiments and show that DRAM provides
$\approx 1.5\times$
greater throughput while utilizing
$\approx 25\%$
less energy than SSDs. Further, we show that asynchronous C/R provides
$\approx 4\times$
greater throughput while using practically a third of the energy than synchronous C/R. Further we show that data size and throughput are directly correlated to energy consumption, and therefore C/R developers should focus on ways to improve/maintain high throughput in order to effectively reduce energy consumption, rather than throttle hardware, to efficiently address exascale needs.