On the Non-Suitability of Non-Volatility

For many emerging and existing architectures, NAND flash is the storage media used to fill the cost-performance gap between DRAM and spinning disk. However, while NAND flash is the best of the available options, for many workloads its specific design choices and trade-offs are not wholly suitable. One such workload is long-running scientific applications which use checkpoint-restart for failure recovery. For these workloads, HPC data centers are deploying NAND flash as a storage acceleration tier, commonly called burst buffers, to provide high levels of write bandwidth for checkpoint storage. In this paper, we compare the costs of adding reliability to such a layer versus the benefits of not doing so. We find that, even though NAND flash is non-volatile, HPC burst buffers should not be reliable when the performance overhead of adding reliability is greater than 2%.