NVMe-oF exposes storage network latency issues
NVMe-oF is bringing NVMe's latency-lowering capabilities to shared storage. It's an optimal time for this to happen as AI, machine learning and big data analytics applications demand better performance from storage networks.
In the past -- pre-NVMe and NVMe-oF -- no one worried much about storage network latency. Storage managers had bigger latency concerns. Applications, as well as storage systems and devices, were far more likely sources of performance degradation than network switches and adapters. Today, NVMe and NVMe-oF have eliminated many of those sources of storage latency. That leaves formerly overlooked problems with network ports, adapters, switches and cables exposed and shines a light on NVMe-oF storage latency.
Storage professionals are feeling the pressure to attend to network issues and properly configure their storage networks to deliver on NVMe-oF's low-latency promise. To do this, they must make use of performance monitoring capabilities that help identify and solve performance issues.
This handbook takes a deep dive into NVMe-oF storage latency issues and storage network performance in general. First, Brien Posey identifies the three key questions to ask before deploying NVMe-oF to ensure optimal performance: Do you have enough throughput? Does your adapter support remote direct memory access (RDMA)? Are your device drivers and OS kernels up to date?
Then, the article entitled "NVMe-oF performance monitoring best practices that work" explains how to use real-time telemetry data capture, driven by on-switch application-specific integrated circuits, to see what's happening to a storage network at any given moment and identify performance issues.
Finally, Marc Staimer looks at the high-performance interconnects that guarantee low-latency storage. He compares the architecture, bandwidth, congestion and routing capabilities of InfiniBand and RDMA over Converged Ethernet, and identifies the upcoming high-performance interconnect technologies focused on minimizing storage latency.