DISTORT

High-Performance Disaggregated Storage over RDMA natively orchestrated in Kubernetes

Near-Local Performance

Decouples storage capacity from compute nodes using NVMe-over-Fabrics (NVMe-oF) with RDMA, delivering microseconds of latency and extreme high throughput.

Kubernetes-Native Schedulers

Driven by a declarative Custom Resource Definition (CRD) state machine to coordinate device discovery, node health, partition claims, and volume bindings.

SPDK User-Space Polling

Supports fully uncompromised storage pathways. Bypasses the traditional Linux kernel block layer to eliminate context switches and interrupt overhead completely.

Welcome to DISTORT

DISTORT (DISaggregated STorage Over Rdma Transport) is a high-performance, Kubernetes-native storage engine that bridges the gap between physical disaggregated storage fabrics and dynamic container environments.

As cloud-native architectures become the standard for deploying scalable applications, the demand for high-performance, low-latency storage has surged. Disaggregated storage, particularly NVMe-over-Fabrics (NVMe-oF) using Remote Direct Memory Access (RDMA), offers a promising solution by decoupling storage capacity from compute nodes, maintaining near-local access speeds.

However, integrating physical RDMA targets with the dynamic, containerized environments of Kubernetes presents significant orchestration challenges.

DISTORT is a Kubernetes-native storage engine designed to bridge this gap. It leverages a Custom Resource Definition (CRD)-driven state machine to manage the lifecycle of physical NVMe devices, partition them asynchronously, and expose them as NVMe-oF RDMA targets directly to Kubernetes Pods via a Container Storage Interface (CSI) driver. By offloading the control plane to Kubernetes while maintaining a lightweight data path, DISTORT achieves high-performance, dynamic storage provisioning suitable for data-intensive cloud applications.

Core Pillars

graph TD classDef k8s fill:#326ce5,stroke:#fff,stroke-width:2px,color:#fff; classDef engine fill:#1f2937,stroke:#3b82f6,stroke-width:2px,color:#fff; classDef hardware fill:#059669,stroke:#fff,stroke-width:2px,color:#fff; subgraph Orchestration [Kubernetes Control Plane] PVC["Persistent Volume Claim"] CRD["Declarative CRDs
(NVMePartition, NVMeDevice)"] MGMT["Management Controller"] PVC -.-> CRD MGMT --> CRD end subgraph DataPath [DISTORT High Performance Data Path] SPDK["SPDK User-space Polling
(VFIO-PCI / Lvol)"] RDMA["SoftRoCE / Physical RDMA"] CSI["DISTORT CSI Driver
(Staging & Mounting)"] end subgraph Media [Physical Hardware Layer] NVMe["Physical NVMe Disks"] end CRD --> CSI CSI --> SPDK SPDK --> RDMA RDMA --> NVMe class PVC,CRD,MGMT k8s; class SPDK,RDMA,CSI engine; class NVMe hardware;

Getting Started

Explore the different sections of the documentation to understand and consume DISTORT:

  • Architecture: Read our high-level design, component roles, interactive control sequence, and implementation layout.
  • Using DISTORT: Learn how to discover underlying storage controllers, claim hardware drives using NVMeDeviceClaim specs, and request StorageClasses.
  • Contributing: Setup a local multi-node virtual K3s cluster with virtual NVMe drives and SoftRoCE loopback to build code and run E2E Ginkgo validation tests.