Stefan Hajnoczi

vhost-user-blk is a userspace block I/O interface that has traditionally been used to connect software-defined storage to hypervisors. This talk covers how any application that needs fast userspace block I/O can use vhost-user-blk and its advantages over network protocols. A client library called libblkio is available for C and Rust applications will be introduced. The protocol is also summarized for those wishing to understand how it works or implement it from scratch.

This talk is intended for developers interested in connecting applications to SPDK or qemu-storage-daemon and those who want to know more about software-defined storage interfaces.

• Local block storage interfaces
  • Kernel vs userspace
  • Notifications vs polling
  • Message-passing vs zero-copy
• What is vhost-user-blk?
  • Implemented by qemu-storage-daemon and SPDK
  • virtio-blk and VIRTIO
• How to connect using libblkio (C/Rust)
• How to implement a server using libvhost-user (C) or vhost-user-backend (Rust)
• How to integrate with the Linux kernel block layer using VDUSE

The VIRTIO standard defines I/O devices that are commonly used in virtual machines today. The last version of the standard was released in 2019 and much has changed since then. This presentation covers new devices and features in the upcoming VIRTIO 1.2 standard.

There are 9 new device types: fs, rpmb, iommu, sound, mem, i2c, scmi, gpio, and pmem. We will look at the functionality offered by these devices and their status in Linux.

This presentation is aimed at users of virtualization who may be interested in new virtual devices that are becoming available in Linux, QEMU, etc. It may also be of interest to driver and virtual machine monitor developers who are considering implementing new devices.

File descriptor monitoring is at the core of event-driven applications from graphical applications to web servers. Over the history of Linux, a number of system calls APIs have been introduced to improve upon the performance, features, and interface design. Developers may ask themselves which API they should use and how they differ.

This talk covers select(2), poll(2), epoll(7), as well as the more recent Linux AIO and io_uring APIs. We will look at the classic scalability challenges with these APIs as well as the latest shared kernel memory ring and polling approaches. An understanding of the evolution of file descriptor monitoring in Linux exposes API design topics that have relevance even if you don't need to implement an event loop in your application.

The Linux logo used for this talk was created by Larry Ewing.

This talk covers the new virtio-fs shared file system that allows a host directory tree to be shared with guests. Sharing files with the guest is required by several use cases including container VMs, File-System-as-a-Service, and traditional virtualization. virtio-fs goes further than previous attempts by taking advantage of the co-location of the guest and host using DAX to share the host page cache. This presentation explains how to use virtio-fs, a bit about how it works internally, and the current status.

virtio-fs is a new shared file system giving access to a directory that both the host and guests can access. Traditionally shared file systems have been used to make data available to the guest during installation, boot a guest from a directory tree on the host, or to develop code on the host and test it in-place without copying files into the guest. New use cases including container VMs and File-System-as-a-Service have introduced new requirements that virtio-fs is designed to meet.

Previous attempts at shared file systems have included virtio-9p or simply used network file systems. virtio-fs is unique because it is possible to access files directly from the host page cache. This eliminates data copies and communication, resulting in lower memory footprint and higher performance for many workloads.

This presentation explains how to use virtio-fs and covers its architecture. It includes performance benchmarks showing how various features perform and a comparison with virtio-9p.

The introduction of non-volatile memory changes how applications, databases, and virtual machines will work in the future. NVDIMM is not simply a faster block device. Programs can avoid block I/O entirely and use byte-addressable NVDIMM to benefit from the performance characteristics of RAM. This requires new storage APIs that applications must use instead of traditional block I/O.

These new programs run successfully inside KVM virtual machines thanks to the vNVDIMM support already available in QEMU. Virtualization offers additional options for managing and using NVDIMM beyond what is available on bare metal.

This talk covers the NVDIMM programming model and how KVM virtual machines can use NVDIMM for faster I/O, reduced memory footprint, and faster boot times.

Xiao Guangrong recently gave a talk at KVM Forum covering the ACPI and hardware interface for NVDIMM under KVM. I want to follow that up with a talk for a wider audience of developers and system administrators who care more about the use of NVDIMM than its implementation.

This talk covers recent work on NVDIMM in QEMU and focuses on how applications can take advantage of this new class of storage device. It is suitable for an audience interested in storage or virtualization. Most of the concepts carry over to other hypervisors besides KVM.

KVM is the most popular hypervisor deployed with OpenStack and is also often used with libvirt in non-cloud scenarios. With multiple layers of software between the user and the virtual machine, how does one troubleshoot failures and performance issues?

This presentation covers tools and techniques for observing virtual machines with the KVM hypervisor. It gives you a mental model of KVM's architecture so you'll know how to get to the bottom of questions about virtual machine behavior and performance.

Areas covered include guest CPU activity, RAM, disk I/O, network traffic, and the QEMU monitor. Examples are based on real-life scenarios often encountered by KVM users.

The KVM hypervisor is part of the Linux kernel, making it an easy and popular way to run virtual machines on Linux hosts. KVM is also the most well supported and popular hypervisor deployed with OpenStack, typically through the libvirt library.

When something goes wrong with a virtual machine there are often questions like "how do I see the network packets that the host is receiving for this virtual machine?" or "can I find out which virtual machine is causing excessive disk I/O?".

Luckily KVM is not a black box, it's an open source hypervisor that runs together with the userspace QEMU emulator. There are a wealth of tools for observing virtual machine activity and they can be used to troubleshoot failures and performance issues. From packet sniffing to CPU profilers, tracing to debuggers, this talk covers the arsenal of tools and how to apply them.

This presentation arms you with knowledge of KVM's architecture and how to observe virtual machine activity. Monitoring and debugging virtual machines is actually much like (and sometimes even easier than) on physical machines so you can reuse many familiar tools.

Once you understand how virtual machines interact with the outside world, how CPU, RAM, disk I/O, and network resources are connected, investigating issues with virtual machines becomes a systematic process. This presentation includes examples of common questions you can ask with the help of performance monitoring and debugging tools.

This talk is suitable for both users of KVM virtual machines and developers of management software like libvirt, OpenStack, or oVirt. The aim is share knowledge gained from real-world cases so that virtual machines are not just a joy to use, but also a powerful vehicle for observability.