Unveiling the performance of fast NVM storage with the uDepot KV-store

Storage devices based on novel NVM technologies, such as 3D XPoint, have recently become available and drastically improve performance compared to conventional SSDs. On one hand, these devices present opportunities to accelerate systems by complementing conventional (slower) storage and to reduce the costs of systems that rely on DRAM in order to meet their performance demands. On the other hand, their performance exposes many inefficiencies in applications and the IO stack, rendering them unable to fully exploit these new devices. In this paper, we present uDepot, a key-value store built bottom-up to match the performance of fast NVM storage devices. uDepot is carefully crafted to avoid inefficiencies, and employs a novel, task-based IO runtime system to maximize IO performance, enabling applications to use fast NVM devices at their full potential. We show that uDepot’s performance nearly matches the raw capabilities of the underlying devices, both in terms of throughput and latency, and is highly scalable across multiple devices. As an example, uDepot delivers more than 6 million GET operations per second on a single server. Moreover, we demonstrate a memcache service built on top of uDepot that offers similar performance to DRAM-based implementations at a much lower cost.