Shulou(Shulou.com)06/02 Report--

Summary: Ali Yun in KVM Forum 2018

KVM Forum is a high-end technology forum organized by the Linux Foundation, which provides a forum for community maintainers, developers, and users to discuss Linux virtualization technology trends and challenges. Attendees are concentrated in areas related to KVM virtualization and are the most important and authoritative conference for the KVM community.

Overview

The 2018 KVM Forum will be held in Edinburgh, UK, on October 24-26. This meeting attracted many technical experts to participate in the meeting. At the same time, the participating companies gathered various cloud manufacturers as well as Red Hat, Intel, IBM, AMD and other system and hardware manufacturers.

Let's first take a look at the super popularity of KVM Forum.

Cdn.com/83f1008859a9c9703c3fe2f92691708c194b7209.png ">

On the first day, keynote was packed.

A full dining area

There are a total of 51 keynote speeches covering all areas of virtualization, such as hot migration, nested virtualization, ARM virtualization, device pass-through, lightweight virtualization, security and performance optimization.

Hot topic

In terms of the number of issues and on-site effects, the following three areas have attracted more attention:

Nested Virtualization

Virtualization performance optimization

Heat transfer

Nested Virtualization

To put it simply, nested virtualization is to solve the problem of how to nest and run virtual machines in virtual machines. As a new hot topic recently, there are as many as 4 lectures about nested virtualization on this KVM Forum.

The meeting introduced the technologies of para-virtualization VMCS, paravirtualization MSR-bitmap and clock optimization, which can improve the handover performance between vCPU when nesting and help KVM run better in virtual machines.

In addition, the meeting also covered interrupt virtualization performance optimization for nested virtualization and support for PowerPC processors.

It can be seen that nested virtualization technology has achieved some results in commercial applications, but the challenges brought by security are still the focus of questions.

Virtualization performance optimization

In fact, the performance optimization of virtualization has attracted more attention in KVM Forum in recent years. At last year's conference, Zhang Yang, a senior technical expert from Aliyun, proposed effective solutions to the problems of poor performance of messaging services on KVM, virtual clock overhead, and the performance degradation seen by messaging middleware in KVM.

This time, the related lectures covered vCPU optimization, memory optimization, and Icano optimization. VCPU optimization includes paravirtualization (para-virtualization) optimization, such as PV TLB shootdown,PV IPI performance optimization; memory optimization involves the introduction of paravirtualization interface by virtio-mem, support for memory hot-plug and so on; Igamo optimization introduces zero-page copy (zero copy) of vHost Virtio-net network receiving queue.

Heat transfer

Flexible scheduling and dynamic scaling are the core features of cloud technology. However, they all rely very much on the technical performance of the underlying heat transfer. At the same time, the impact of physical failure on customers can be greatly reduced through thermal migration. At this Congress, all the speeches related to hot transfer are very popular.

In the past year, Aliyun Virtualization team has invested heavily in the field of hot migration and gained a lot of valuable experience. The three topics of this meeting are all about heat transfer. Here, we will give you a detailed introduction.

Ali Yun in KVM Forum 2018

In this year's KVM Forum, there are a total of five lectures on hot migration, three of which are Ali Yun.

"Cloudatlas: Ways to Make Live Migration Easy and Expectable"

Speaker: Zhang Chao, senior technical expert of Aliyun; Xie Feng, technical expert of Aliyun

Zhang Chao, a senior technical expert at Aliyun

Xie Feng, a technical expert at Aliyun.

How to balance hot migration resource consumption and performance overhead while minimizing the business impact of hot migration on customers is a challenge for all public cloud vendors. Aliyun creatively uses machine learning methods to predict the cost of hot migration and the load of virtual machines, so as to determine a reasonable timing and strategy for hot migration, and ultimately improve the success rate of migration and reduce the impact of hot migration on users.

Aliyun Virtualization team designed a complete system to perform hot migration tasks. In the speech, mentioned the machine learning algorithm part, Aliyun uses FFT to determine whether the periodic load VM, and calculates the cycle of the periodic load VM and then forecasts it. For the VM without obvious periodic load, the team uses two time series models, ARIMA and LSTM, to do regression prediction. The combination of the above three algorithms can quickly and accurately predict the load trend of more than 95% VM in the next 24 hours. In order to find an optimal migration time.

This lecture involved a lot of algorithm research, and the audience asked further questions about the realization of very details, such as the use case, the principle of the scheme, the prediction effect and so on. An industry expert said on the spot: this is really addressing the concerns of public cloud vendors.

"A Perfect Solution for Live Migration with Pass-through Devices"

Speaker: Xu Quan, a technical expert from Aliyun

The presentation addressed the hot migration problem of existing pass-through devices: how to transfer the memory of the device DMA and how to save and restore the state of the device.

Xu Quan, a technical expert at Aliyun

The speech made an in-depth analysis of some of the existing methods and problems. For example, on the 82599 network card, Intel introduces self emulation layer into the device driver inside the virtual machine, simulates the device DMA write operation, and restores the state of the device. However, due to the specific device driver to work, and the defects in the design of the device itself (registers are read-only and cannot be recovered at 100% in recovery), it does not solve the problem very well. From the perspective of new hardware and device drivers, the cost is huge, and the maintenance cost of the software is increased.

Aliyun extends the functions of virtio hardware, such as sensing virtual machine hot migration, providing device bitmap recording device DMA access memory, and dynamically saving and restoring all device registers, while modifying the existing hot migration and VFIO software framework. This can maximize the use of existing device drivers, well support Windows and Linux virtual machines, and do not need to maintain each version of the device driver separately, which greatly reduces the cost.

After the speech, there were technical questions from more than a dozen experts at the scene. After the meeting, Xu Quan also discussed with the Linux community VFIO maintainer Alex Wiliamson several times how to improve the existing VFIO to support the general framework for hot migration of cut-through equipment.

"Live Migration Support for GPU with SRIOV: Challenges and Solution"

Ali Yun and AMD joint speech, Ali Yun speaker: Zheng Xiao, senior technical expert of Ali Yun

The hot migration support of GPU is a difficult point in the industry. First of all, the scheduling and context switching of GPU hardware is based on millisecond, which is several orders of magnitude of CPU, which leads to the problem of preemption of GPU tasks. In addition, the expansion of GPU in the field of cloud computing is a hot spot in recent years, while the support of GPU hardware itself for virtual hot migration is not perfect, such as GPU's dirty track for local memory, hardware support for non local memory's dirty track and so on. All these bring many challenges to the thermal migration of GPU.

Zheng Xiao, Senior Technical expert of Aliyun

The speech mentioned the timeliness and preemption of GPU tasks during the migration, such as how to make up for it through the measures of hypervisor and system software when the GPU hardware itself does not support framebuffer dirty track, as well as the details of the context switching of GPU itself, and so on.

All the issues mentioned above were considered and designed on the first day of Aliyun's joint development with AMD. There are many unique innovations. For example, when some models of GPU hardware do not support dirty track, how to optimize the data from 6 seconds to about 0.35 seconds when the key performance indicators such as Service downtime and other key performance indicators do not meet expectations; in terms of function, from single machine migration to multi-machine multi-card migration; in terms of stability, from downtime at the beginning to thousands of consecutive stress tests.

In addition, the speech also joined the live Live Demo, from the use of the experience above, has been able to achieve the smooth migration of GPU rendering tasks.

The speech received a lot of attention. After the meeting, some top hardware manufacturers also discussed more details of thermal migration and performance indicators with the speaker.

Summary

After participating in this KVM Forum, we deeply feel that how to solve the practical problems encountered in the actual production environment is getting more and more attention from the industry.

Through participating in the meeting, it not only promotes our communication with the virtualization industry, but also allows the team to obtain a lot of information and valuable experience, which can help us to better improve and constantly sharpen the underlying capabilities of flexible computing products.

Author: wild Goose Eagle

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.