BlinkFS: A distributed file system for intermittent power

Abstract

The ability to use intermittent power in data centers introduces numerous new opportunities for optimization, including (i) using real-time markets to buy more power when it is cheap, (ii) increasing the fraction of clean, but intermittent, renewable power, (iii) capping power for long periods to extend UPS lifetime during blackouts, and (iv) fully utilizing a data center's power delivery infrastructure. The capability to run off intermittent power also moves us closer to the vision of a net-zero data center that consumes no net energy from the electric grid and has a small carbon footprint. However, designing systems to operate under intermittent power is challenging, since applications often access persistent distributed state, where power fluctuations can impact data availability and I/O performance. To address the problem, we design and implement BlinkFS, which combines blinking with a power-balanced data layout and popularity-based replication/reclamation to optimize I/O throughput and latency as power varies. Our experiments show that BlinkFS outperforms approaches that co-opt existing energy-proportional distributed file systems (DFSs) for intermittent power, particularly at low steady power levels and high levels of intermittency. For example, we show that BlinkFS reduces completion time for MapReduce-style jobs by 42% at 50% full power compared to existing energy-proportional DFSs.