Analyzing, Modeling, and Provisioning QoS for NVMe SSDs

NVMe-based SSDs are in huge demand for Big Data analytics owing to their extremely low latency and high throughput for both read and write operations. Their inherent parallelism in request processing makes them ideal to be used in virtualized environments, where sharing of resources is a given. Given the shared resource-driven ideology of cloud environments, it is imperative to design middleware which can provide some guarantee of service to applications. In this paper, we show how such QoS can be provided for NVMe SSDs in virtualized environments. Our contributions are threefold: (1) design of accurate NVMe emulation mechanisms in QEMU to provide QoS schemes, (2) theoretical modeling of arbitration mechanisms for assisting in SLA provisioning, and (3) proposing designs in Intel SPDK to seamlessly use the hardware-based QoS provided by NVMe. We provide a complete case for our designs and validate them through thorough experimental evaluation. We show that Deficit Round Robin (DRR) as a hardware-based arbitration scheme is more suited for providing bandwidth guarantees for NVMe SSDs. Our evaluations show that by combining our proposed QoS-aware NVMe emulator in QEMU and enhanced SPDK runtime, we can achieve I/O bandwidth SLA guarantees in an application oblivious manner.