Feature List
Feature |
Sub-feature |
Description |
Constraints |
VM Support |
Remarks |
|---|---|---|---|---|---|
Virtualized CPU acceleration |
Interrupt passthrough |
When hardware-assisted virtualization acceleration is enabled on new Kunpeng 920 processor models, the direct interrupt injection of GICv4.1 (including vSGI passthrough) cuts interrupt response times and enhances throughput for demanding network and I/O workloads. |
Supported OSs: openEuler 22.03 LTS SP3 and later Constraints: Only new Kunpeng 920 processor models are supported. |
No |
The virtualization software cannot be used on VMs. |
Virtualized CPU acceleration |
Cluster awareness |
A cluster is a hardware unit of a CPU. Each cluster contains several cores. Cores in a cluster share the same L3 cache tag. By adding the option for cluster task scheduling tuning of the OS kernel, cross-cluster thread scheduling can be prevented and L3 cache tag resources can be reused, improving the CPU scheduling efficiency and memory bandwidth utilization of multi-thread applications. Cluster awareness displays the cluster topology to the guest OS so that VM users can better use the cluster to optimize services. It also leverages hardware resources more efficiently, increases the system throughput, and speeds up responses to requests. For details, see Cluster Tuning Guide. |
Supported OSs: openEuler 22.03 LTS SP2 and later Constraint: The topology information of the physical CPU cluster must be correctly mapped to the VM. |
Yes |
The enabling mode on a VM is the same as that on a physical machine. The OS must be openEuler 22.03 LTS SP2 or later. |
Virtualized CPU acceleration |
NUMA awareness |
NUMA awareness is an optimization technology for the non-uniform memory access (NUMA) architecture. Accessing the memory of other CPUs in the NUMA architecture causes access delay and performance deterioration. NUMA awareness displays the NUMA topology to the guest OS so that VM users can better use NUMA to optimize services. For details, see Configuring Guest NUMA. |
Supported OSs: CentOS 7.6 and openEuler 20.03 LTS Constraints: After guest NUMA is configured, the service software may fail to identify NUMA. If the service software is not optimized for identifying NUMA, cross-NUMA memory access occurs, causing performance deterioration. |
No |
The virtualization software cannot be used on VMs. |
Virtualized CPU acceleration |
VM lock virtual-real synergy optimization |
This feature uses shared memory to relay vCPU preemption status from the hypervisor to VMs, minimizing conflict-induced errors and improving system stability and reliability. |
Supported OSs: openEuler 20.03 LTS SP1 or later libvirt 6.2.0 or later QEMU 6.2.0 or later |
Yes |
The enabling mode on a VM is the same as that on a physical machine. The OS must be openEuler 20.03 LTS SP1 or later. |
Virtualized CPU acceleration |
VM deadlock detection |
By leveraging non-maskable interrupts (NMIs), this feature monitors interrupt responses in real time to detect deadlocks within VMs, ensuring recovery from unresponsive states caused by deadlocks. |
Supported OS: openEuler 22.03 LTS SP2 libvirt 6.2.0 or later QEMU 6.2.0 or later |
Yes |
The enabling mode on a VM is the same as that on a physical machine. The OS must be openEuler 22.03 LTS SP2 or later. |
Virtualized CPU acceleration |
Cache topology awareness |
This feature provides a method of specifying the cache size in the VM XML configuration of libvirt or the QEMU command for starting VMs. By using this feature, the size of the cache used by a VM can be accurately reflected, and more accurate cache structure information is available for further optimizing the application performance in the VM. |
Supported OS: openEuler 22.03 LTS SP4 libvirt 6.2.0 or later QEMU 6.2.0 or later |
No |
Physical machines transparently transmit topology to VMs. |
Virtualized CPU acceleration |
VM cache topology information passthrough |
This feature applies to all AArch64 KVM VM scenarios that require the guest to be aware of the actual cache hierarchy of the host, especially for workloads sensitive to cross-CCL communication and memory access locality (such as large-scale databases, distributed computing frameworks, and high-performance computing tasks). The L3 cache can be manually attached at the CCL level to keep the MC scheduling domain within a single CCL, reducing cross-CCL task scheduling and IPI communication and improving memory access locality. This feature can be used together with the SGI distribution path optimization in IPI Latency Optimization for High-Specification VMs to further reduce the end-to-end IPI latency within a VM. The two features are independent of each other and can be deployed and used separately. |
Supported OS: openEuler 22.03 LTS SP4 libvirt 6.2.0 or later QEMU 6.2.0 or later |
No |
Physical machines transparently transmit topology to VMs. |
Virtualized CPU acceleration |
Interrupt passthrough |
When hardware-assisted virtualization acceleration is enabled on new Kunpeng 920 processor models, the direct interrupt injection of GICv4.1 (including vSGI passthrough) cuts interrupt response times and enhances throughput for demanding network and I/O workloads. |
Supported OSs: openEuler 22.03 LTS SP3 and later Constraints: Only new Kunpeng 920 processor models are supported. |
No |
The virtualization software cannot be used on VMs. |
Virtualized CPU acceleration |
GICv4.1 overcommitment optimization |
This feature allows the GIC to skip VMOVP instructions when vCPUs are migrated between CPUs that share the same vPE table, thus improving VM service performance in overcommitment scenarios. |
Supported OS: openEuler 24.03 LTS SP3 Hardware constraints: Only the new Kunpeng 920 processor model is supported. |
No |
The virtualization software cannot be used on VMs. |
Virtualized CPU acceleration |
IPI latency optimization for high-specification VMs |
By optimizing the scheduling on the SGI distribution path to O(1), the end-to-end IPI latency can be significantly reduced for high-specification VMs, improving service performance in multi-core concurrency scenarios. |
- Hardware environment: Only new Kunpeng 920 processor models and Kunpeng 950 processors are supported. - Software environment: The host OS is openEuler 24.03 LTS SP3, the kernel baseline is OLK-6.6 (6.6.0-135.0.0), the QEMU version is 8.2.0, and the libvirt version is 9.10.0. In addition, the SGI optimization and cacheinfo patches must be applied for QEMU/libvirt. |
No |
The virtualization software cannot be used on VMs. |
Virtualized CPU acceleration |
VF ITS selection |
This feature provides a PF-level sysfs interface, which allows users to configure a group of ITSs that can be used by VFs for a PF (by specifying the ITS raw indexes). Before accepting the configuration, the kernel verifies the configuration based on the ITS currently used by the PF, platform topology rules, and the chip capability trustlist. The configuration is saved and takes effect during subsequent VF enumeration only when all conditions are met. |
Hardware constraint: The hardware platform supports only Kunpeng 950 processors. Software constraints OS: openEuler 24.03 LTS SP3 Kernel source code baseline: OLK-6.6 6.6.0-133.0.0 libvirt: 9.1.0 (Yum repository) QEMU: 8.2.0 (Yum repository) BIOS version: 10.79 |
No |
This feature affects the ITS selection for VFs. VMs do not have ITS hardware. |
Virtualized I/O acceleration |
vKAE |
The Kunpeng Accelerator Engine (KAE) encryption and decryption module uses the KAE to implement the RSA, SM3, SM4, DH, MD5, and AES algorithms. It provides high-performance symmetric and asymmetric encryption/decryption algorithms based on the lossless user-mode driver framework. Compatible with OpenSSL 1.1.1a and later versions, it supports the synchronous and asynchronous mechanisms. This module can accelerate applications in VMs and containers. For details about how to use KAE in VMs, see Using KAE on a KVM. For details about how to use KAE in containers, see Using KAE on Docker. For details about how to use KAE-accelerated live migration in VMs, see Virtualization Scenario KAE Accelerated Live Migration Feature Guide. |
Supported OSs: openEuler 20.03 LTS SP1 and later, and openEuler 22.03 LTS SP1 and later Constraint: You need to apply for a license on a physical machine. For details, see Obtaining a License. |
Yes |
The VM OS must be openEuler 20.03 LTS SP1 (or later) or openEuler 22.03 LTS SP1 (or later). |
Virtualized I/O acceleration |
Virtualization scenario KAE passthrough live migration |
VM live migration is a critical operational capability that enables seamless transfer of running VMs between physical hosts without service interruption. KAE passthrough live migration specifically addresses the scenario where VMs are configured with KAE passthrough devices, offering enhanced operational flexibility and continuous service availability. |
Supported OSs: openEuler 22.03 LTS SP4 and openEuler 24.03 LTS SP2 and later Constraints:
|
No |
- |
Virtualized I/O acceleration |
OVS flow table NIC acceleration (based on Mellanox) |
Supports OVS flow table NIC acceleration (based on the Mellanox ConnectX-5 NIC) on the Kunpeng platform. For details, see OVS Flow Table NIC Acceleration Feature Guide. |
Supported OSs: CentOS 7.6 and openEuler 20.03 LTS Constraints:
|
No |
The virtualization software cannot be used on VMs. |
Virtualization management optimization |
vCPU hotplug |
This feature increases or decreases the number of vCPUs of a running VM without interrupting services. |
Supported OS: openEuler 24.03 LTS Constraints: The CPU specifications of the VM applying the hotplug feature cannot exceed the maximum CPU specifications supported by the hypervisor and the maximum CPU specifications supported by the guest OS. |
No |
The virtualization software cannot be used on VMs. |
Virtualization management optimization |
QEMU VM memory hot add |
This feature allows for starting a VM of a NUMA node whose initial memory configuration is 0 in the XML configuration file and dynamically adding memory to the NUMA node. |
Supported OS: openEuler 22.03 LTS SP4 Constraints:
|
Yes |
The enabling mode on a VM is the same as that on a physical machine. The OS must be openEuler 22.03 LTS SP4 or later. |
Virtualization management optimization |
Virtualization scenario KAE-accelerated live migration |
During the VM live migration, compression technologies (for example, zlib library) are usually used on the source physical host to compress memory pages before data transmission, and then the memory pages are decompressed on the target physical host, thereby speeding up the VM live migration. KAEzlib is a standard zlib interface provided by the KAE compression module. It uses the Kunpeng hardware acceleration module to implement the Deflate algorithm and works with the lossless user-mode driver framework. Therefore, KAE can replace the open-source compression library zlib to accelerate VM live migration. |
Supported OSs: openEuler 22.03 LTS SP1/SP2/SP3/SP4 Constraints:
|
No |
- |
Virtualization management optimization |
PMU virtualization |
PMU virtualization enables PMU events to be collected on a VM, which helps analyze and optimize the performance of the VM OS and service software. |
Supported OSs: openEuler 22.03 LTS SP4 and later |
Yes |
The enabling method on the VM is the same as that on the physical machine. |
Virtualization management optimization |
MPAM-enabled libvirt |
MPAM is used to allocate VMs to different bandwidth/cache partitions, implementing rate limiting or priority control. |
Supported OS: openEuler 24.03 LTS SP2 Constraints: Arm-based KVM and QEMU virtualization platforms are required. Only libvirt 9.10.0 and QEMU 8.2.0 are supported. |
No |
The virtualization software cannot be used on VMs. |
Virtualization management optimization |
Cross-generation VM live migration |
VM live migration allows a VM to be migrated from one physical host to another without interrupting the VM running. The key benefit of cross-generation migration lies in its ability to maintain VM operation and service continuity during hardware upgrades by supporting migration between different hardware generations. Successful implementation typically requires sufficient compatibility between source and target hardware platforms, and hardware discrepancy tolerance at the VM OS and application layers. While technically challenging, this capability significantly enhances operational flexibility for data centers and cloud service providers by ensuring continuous service availability. |
Supported OSs: openEuler 24.03 LTS SP1 and later For details about constraints, see General Constraints. |
No |
The virtualization software cannot be used on VMs. |
Virtualization management optimization |
VM single-core and single-page exception handling |
To ensure long-term stable running of Internet systems, Reliability, Availability, and Serviceability (RAS) capabilities are required to obtain the faulty hardware of the current physical machine and process the corresponding hardware information, minimizing the impact scope. After the single-core and single-page exception handling feature is enabled, single-core corrected errors (CEs) can be isolated online on Kunpeng servers without affecting service running; uncorrected errors (UEs) in a single page of memory affect only one process in a VM, preventing the VM from going offline. |
Supported OS: openEuler 24.03 LTS SP3 libvirt 9.10.0-26.oe2403sp3 or later QEMU 8.2.0 or later |
No |
The virtualization software cannot be used on VMs. |
Virtualization management optimization |
VM huge-page memory configuration reduction |
The memory price in the current market keeps increasing. To reduce the VM memory overhead and improve the physical memory utilization, technical measures need to be taken to optimize the memory configuration. This feature uses a PV mechanism to reclaim idle page cache and perform memory defragmentation within VMs, the ZRAM module to compress and store the cold-page memory of VMs, and the Kunpeng Accelerator Engine (KAE) to improve the compression efficiency. In addition, this feature allows huge-page memory to be swappable. This implements efficient utilization of memory resources, so that more VMs can be deployed with limited physical memory. |
Supported OS: openEuler 24.03 LTS SP3 (kernel baseline: OLK-6.6) QEMU 8.2.0 or later libvirt 9.10.0 or later For details about constraints, see Constraints. |
No |
The virtualization software cannot be used on VMs. |
Feature check tool |
Hardware virtualization feature check tool |
The hardware virtualization feature check tool (VirtCheck) is designed for new Kunpeng 920 processor models. It automatically scans and analyzes customer environments and scenarios to determine whether the virtualization features provided by Kunpeng BoostKit are supported or enabled. |
Supported OS kernel versions:
|
Yes |
Install the RPM package of the check tool on the VM and run the check command virtcheck --guest. |
UnifiedBus SuperPoD optimization |
Capacity balancing optimization for UB memory borrowing in virtualization scenarios |
In cross-node memory borrowing scenarios, the UBSE scheduler needs to select a lending node from multiple candidate nodes. However, in high-bandwidth service scenarios such as databases, multiple nodes may borrow memory from the same lending node. As a result, that node's bandwidth load becomes too high, creating a performance bottleneck and failing to meet the requirement for smooth bandwidth distribution under heavy traffic. To address this, the capacity balancing feature uses performance-priority mode to keep differences in lending bandwidth load across nodes within a controllable range and prevent any single node from being overloaded. Once bandwidth balancing is achieved, existing borrowing relationships are reused first to reduce link setup overhead; next, the node with the most free memory is selected to improve memory utilization. This mitigates the memory access performance fluctuations caused by uneven bandwidth load in high-bandwidth service scenarios. |
Supported OS: openEuler 24.03 LTS SP3 (kernel 6.6.0-145.3.27.158 or later) QEMU 8.2.0 or later libvirt 9.10.0 or later UBSE version to be determined. |
No |
The virtualization software cannot be used on VMs. |
UnifiedBus SuperPoD optimization |
Page cache and fragment reclamation optimization in virtualization scenarios |
In a VM, the host machine cannot detect the hot and cold status of pages in the page cache. As a result, the page cache in the VM accumulates over time, and the cache that is not accessed for a long time cannot be reclaimed. Reclaiming the cache directly through the balloon mechanism can cause VM performance to fluctuate. In addition, the balloon mechanism reports free pages at a 4 KB granularity, while free subpages are scattered across 2 MB memory blocks, making them impossible to reclaim. The page cache and fragment reclamation feature is a solution to the preceding two problems. It adds a feature bit to virtio-balloon to support memory compaction, page-cache reclamation, and fragment reclamation. When the detected fragmentation rate reaches a threshold, defragmentation and reclamation are triggered when the CPU usage is low. |
Supported OS: openEuler 24.03 LTS SP3 Guest kernel 6.6.0-145.3.27.158 or later QEMU 8.2.0-88 or later libvirt 9.10.0-35 or later memlinkd 1.0.0-214 or later |
Yes |
The guest kernel needs to be replaced with the corresponding version. |
Open-source enablement |
CubeSandbox deployment |
CubeSandbox can be deployed and optimized on Kunpeng servers to implement fast sandbox startup. |
OS: openEuler 24.03 LTS SP3 CubeSandbox: v0.5.1 AArch64 release Python SDK: cubesandbox==0.5.0 Docker: 25.0.3 |
No |
The virtualization software cannot be used on VMs. |
Open-source enablement |
E2B sandbox deployment |
E2B can be deployed on Kunpeng servers to quickly create and destroy lightweight sandbox instances. |
OS: openEuler 24.03 LTS SP3 e2b-infra: 2026.09-3.oe2403sp3.aarch64 |
No |
The virtualization software cannot be used on VMs. |
Open-source enablement |
Kata sandbox deployment |
Kata Containers can be deployed on Kunpeng servers to start AI sandboxes. In addition, the Nydus image acceleration solution can be used together to efficiently distribute container images. |
OS: openEuler 24.03 LTS SP3 Kubernetes: 1.28.14 containerd: 1.7.27 Kata Containers: 3.27.0 Nydus: 2.4.0 Nydus snapshotter: 0.15.12 nerdctl: 1.7.5 |
No |
The virtualization software cannot be used on VMs. |
Open-source enablement |
Hybrid deployment on OpenStack |
Supports hybrid deployment for BMSs and KVMs in availability zones (AZs). Each AZ must contain only x86 or Kunpeng servers. |
Supported OSs: CentOS 7.6 and openEuler 20.03 LTS Constraints:
|
No |
The virtualization software cannot be used on VMs. |
Open-source enablement |
Hybrid deployment on Kubernetes |
Supports hybrid deployment of Kunpeng and x86 servers for Docker containers. For details, see Hybrid Container Deployment Guide. |
Supported OSs: CentOS 7.6 and openEuler 20.03 LTS Constraints: Kubernetes management nodes do not support hybrid deployment. |
No |
The virtualization software cannot be used on VMs. |
Open-source enablement |
oVirt for VM cluster management |
Supports oVirt use on the Kunpeng platform. For details, see Kunpeng oVirt Lightweight Virtualization Management Platform Deployment Guide. |
Supported OSs: openEuler 20.03 LTS SP1 and openEuler 22.03 LTS SP2 |
No |
The virtualization software cannot be used on VMs. |
Open-source enablement |
Large-scale Docker container networking |
For details about the large-scale Docker container networking scenario, see Large-Scale Docker Container Networking Feature Guide. |
Supported OSs: CentOS 7.6 and openEuler 20.03 LTS |
Yes |
The Docker container technology can be used on VMs. |
Open-source enablement |
Fixed IP address of the Docker container even after the porting |
||||
Cloud-native infrastructure |
KAE device plugin |
The Kunpeng Accelerator Engine (KAE) is a hardware-based acceleration solution built on Kunpeng 920 series processors. It supports encryption, decryption, and decompression. The KAE device plugin automatically manages all KAE devices on the server and streamlines the KAE device passthrough operations. Using this plugin, you can pass through KAE devices to containers through simple statements to accelerate encryption, decryption, and data compression, and set QoS for KAE devices. |
Supported OSs: openEuler 22.03 LTS SP4 and later Constraints: Before using the KAE device plugin, ensure that the VF of the KAE device has been created. |
No |
This feature can be used only in container scenarios. |
Cloud-native infrastructure |
KAE-enabled Envoy acceleration |
In microservice scenarios, Envoy needs to process a large number of TLS requests, regardless of whether it functions as an ingress gateway or a microservice proxy. Especially in the handshake phase, asymmetric encryption and decryption consume a large number of CPU resources. If microservices are deployed on a large scale, this overhead may become a system performance bottleneck. To address this, the KAE private key provider of Envoy offloads time-consuming encryption and decryption operations from the CPU to KAE. This accelerates encryption and decryption while releasing CPU computing power for other service workloads. |
Supported OSs: openEuler 22.03 LTS SP4 and later Constraints: Before using the KAE device plugin, ensure that the VF of the KAE device has been created. |
No |
This feature can be used in container and physical machine scenarios. |
Cloud-native high-performance networking |
Kubernetes SR-IOV device plugin |
The SR-IOV technology can virtualize a single physical PCIe device into multiple virtual PCIe devices, and then pass through these virtual devices to each VM, so as to be applied in scenarios where a single physical PCIe device supports running of multiple VMs. The Kubernetes SR-IOV device plugin automatically identifies and manages SR-IOV devices on nodes, and automatically completes the mounting process based on the SR-IOV device type specified in the configuration file, streamlining the usage process. For details, see Cloud Native Scenario Kubernetes SR-IOV Device Plugin User Guide. |
Supported OSs: openEuler 22.03 LTS SP4 and later Constraints: Before using the SR-IOV device plugin, ensure that the VF of the SR-IOV device has been created. |
No |
This feature can be used only in container scenarios. |
Cloud-native resource affinity and isolation |
Kunpeng Topology Affinity Plugin |
In the Kubernetes container deployment scenario, the CPU scheduling range of containers can be automatically adjusted based on the CPU load of compute nodes to ensure NUMA affinity. |
Supported OSs: openEuler 20.03 LTS SP3 and openEuler 22.03 LTS SP4 Kubernetes versions: 1.23.x and 1.28.4 containerd version: 1.7.14 Constraint: Use containerd as the container runtime. Docker version: 20.10.14 Constraint: Use Dockershim as the runtime communication component. |
No |
This feature can be used only in container scenarios. |
Cloud-native resource affinity and isolation |
Kunpeng QoS Controller |
MPAM is a new feature of Armv8.4-A. The MPAM mechanism adds different labels (PARTID/PMG) to all requests of different service flows at the CPU/IO source so that the hardware can sense the service flows. Based on the information, resources, such as the cache capacity and DMC bandwidth, can be dynamically allocated to each component (such as the cache and DDR) of system resources to isolate different service flows and reduce interference. SMT QoS limits the BE task instruction throughput by using the WFI instruction or other methods to ensure that LS tasks obtain more microarchitecture resources. It also allows BE tasks to use the idle SMT threads of LS tasks, improving overall resource utilization. Kunpeng QoS Controller can dynamically isolate the resource usage of offline services in hybrid deployment scenarios. When resources are idle, offline services can use more resources. When resources are insufficient, the resource usage of offline services is limited to ensure the performance of online services and improve the overall resource utilization of the system. |
Supported OSs: openEuler 22.03 LTS SP4 and later Constraints: Before using the MPAM plugin, ensure that MPAM has been enabled. |
No |
This feature can be used only in container scenarios. |