Nydus: Dragonfly Container Image Service

Introduction

The nydus project implements a content-addressable filesystem on top of a RAFS format that improves the current OCI image specification, in terms of container launching speed, image space, and network bandwidth efficiency, as well as data integrity.

The following benchmarking result shows the performance improvement compared with the OCI image for the container cold startup elapsed time on containerd. As the OCI image size increases, the container startup time of using Nydus image remains very short.

Nydus' key features include:

• Container images can be downloaded on demand in chunks for lazy pulling to boost container startup
• Chunk-based content-addressable data de-duplication to minimize storage, transmission and memory footprints
• Merged filesystem tree in order to remove all intermediate layers as an option
• in-kernel EROFS or FUSE filesystem together with overlayfs to provide full POSIX compatibility
• E2E image data integrity check. So security issues like "Supply Chain Attach" can be avoided and detected at runtime
• Compatible with the OCI artifacts spec and distribution spec, so nydus image can be stored in a regular container registry
• Native eStargz image support with remote snapshotter plugin nydus-snapshotter for containerd runtime.
• Various container image storage backends are supported. For example, Registry, NAS, Aliyun/OSS, S3.
• Integrated with CNCF incubating project Dragonfly to distribute container images in P2P fashion and mitigate the pressure on container registries
• Capable to prefetch data block before user IO hits the block thus to reduce read latency
• Record files access pattern during runtime gathering access trace/log, by which user abnormal behaviors are easily caught
• Access trace based prefetch table
• User I/O amplification to reduce the amount of small requests to storage backend.

Currently Nydus includes following tools:

Currently Nydus is supporting the following platforms in container ecosystem:

To try nydus image service:

1. Convert an original OCI image to nydus image and store it somewhere like Docker/Registry, NAS, Aliyun/OSS or S3. This can be directly done by nydusify. Normal users don't have to get involved with nydus-image.
2. Get nydus-snapshotter(containerd-nydus-grpc) installed locally and configured properly. Or install nydus-docker-graphdriver plugin.
3. Operate container in legacy approaches. For example, docker, nerdctl, crictl and ctr.

Build Binary

# build debug binary
make
# build release binary
make release
# build static binary with docker
make docker-static

Quick Start with Kubernetes and Containerd

For more details on how to lazily start a container with nydus-snapshotter and nydus image on Kubernetes nodes or locally use nerdctl rather than CRI, please refer to Nydus Setup

Build Nydus Image

Build Nydus image from directory source: Nydus Image Builder.

Convert OCIv1 image to Nydus image: Nydusify, Acceld or Nerdctl.

Nydus Snapshotter

Nydus-snapshotter is a non-core sub-project of containerd.

Check out its code and tutorial from Nydus-snapshotter repository. It works as a containerd remote snapshotter to help setup container rootfs with nydus images, which handles nydus image format when necessary. When running without nydus images, it is identical to the containerd's builtin overlayfs snapshotter.

Run Nydusd Daemon

Normally, users do not need to start nydusd by hand. It is started by nydus-snapshotter when a container rootfs is prepared.

Run Nydusd Daemon to serve Nydus image: Nydusd.

Run Nydus with in-kernel EROFS filesystem

In-kernel EROFS has been fully compatible with RAFS v6 image format since Linux 5.16. In other words, uncompressed RAFS v6 images can be mounted over block devices since then.

Since Linux 5.19, EROFS has added a new file-based caching (fscache) backend. In this way, compressed RAFS v6 images can be mounted directly with fscache subsystem, even such images are partially available. estargz can be converted on the fly and mounted in this way too.

Guide to running Nydus with fscache: Nydus-fscache

Run Nydus with Dragonfly P2P system

Nydus is deeply integrated with Dragonfly P2P system, which can greatly reduce the network latency and the single point of network pressure for registry server, testing in the production environment shows that using Dragonfly can reduce network latency by more than 80%, to understand the performance test data and how to configure Nydus to use Dragonfly, please refer to the doc.

Accelerate OCI image directly with Nydus

Nydus is able to generate a tiny artifact called a nydus zran from an existing OCI image in the short time. This artifact can be used to accelerate the container boot time without the need for a full image conversion. For more information, please see the documentation.

Build Images via Harbor

Nydus cooperates with Harbor community to develop acceleration-service which provides a general service for Harbor to support image acceleration based on kinds of accelerators like Nydus, eStargz, etc.

Run with Docker

A experimental plugin helps to start Docker container from nydus image. For more particular instructions, please refer to Docker Nydus Graph Driver

Run with macOS

Nydus can also run with macfuse(a.k.a osxfuse).For more details please read nydus with macOS.

Run eStargz image (with lazy pulling)

The containerd remote snapshotter plugin nydus-snapshotter can be used to run nydus images, or to run eStargz images directly by appending --enable-stargz command line option.

In the future, zstd::chunked can work in this way as well.

Documentation

Browse the documentation to learn more. Here are some topics you may be interested in:

• A Nydus Tutorial for Beginners
• Nydus Design Doc
• Our talk on Open Infra Summit 2020: Toward Next Generation Container Image
• EROFS, What Are We Doing Now For Containers?
• The Evolution of the Nydus Image Acceleration (Video)
• Introduction to Nydus Image Service on In-kernel EROFS (Video)

Community

Nydus aims to form a vendor-neutral opensource image distribution solution to all communities. Questions, bug reports, technical discussion, feature requests and contribution are always welcomed!

We're very pleased to hear your use cases any time. Feel free to reach/join us via Slack and/or Dingtalk.

• Slack: Nydus Workspace

• Twitter: @dragonfly_oss

• Dingtalk: 34971767

• Technical Meeting: Every Wednesday at 06:00 UTC (Beijing, Shanghai 14:00), please see our HackMD page for more information.