Definitions

• NUMA (Non-Uniform Memory Access): Non-uniform memory access (NUMA) is a computer memory design where the memory access time depends on the memory location relative to the CPU. Under NUMA, a processor can access its own local memory faster than non-local memory and thus improves VM performance.
• pNUMA Node: A pNUMA node (physical NUMA node) is a host NUMA node predefined based on the host NUMA architecture. It is used to manage the CPUs and memory of the host. A host can have one or more pNUMA nodes. A pNUMA node mainly consists of one or more physical CPU cores (pCPUs) and local memory.
• vNUMA Node: A vNUMA node (virtual NUMA node) is generated by passing-through associated pNUMA nodes via CPU pinning. It is used to manage the CPUs and memory of a virtual machine. A vNUMA node mainly consists of one or more virtual CPU cores (vCPUs) and local memory.
• vNUMA Topology: A vNUMA topology (virtual NUMA topology) is the topology of VM NUMA nodes generated by passing-through associated pNUMA nodes via CPU pinning.
• Local Memory: Local memory is the memory that a CPU (pCPU or vCPU) accesses through the Uncore iMC (Integrated Memory Controller) of the same NUMA (pNUMA or vNUMA) node. Compared with accessing non-local memory, accessing local memory has lower latencies.

How It Works

After a host is added to ZStack Cube Virtualization Edition, you can view the host pNUMA topology and configure vNUMA for virtual machines running on the host based on the topology.

The virtual machine vNUMA configuration is implemented based on CPU pinning. CPU pinning strictly associates virtual machine vCPUs with host pCPUs and allocates specific pCPUs to the virtual machine. During vNUMA configuration, all vCPUs of the virtual machine are pinned to pCPUs, and each pCPU pinned to a single vCPU is in the same pNUMA node.

After vNUMA is configured, the associated host pNUMA node topology is exposed to the virtual machine, generating one or more vNUMA nodes that form the virtual machine vNUMA topology. Based on the vNUMA topology, the virtual machine vCPUs preferentially access local memory of their nodes.

pNUMA Topology

Open the Overview tab of the target host, and click View pNUMA Topology in the hardware overview information box. Host pNUMA topology information includes:

• All host pNUMA nodes and the virtual machines associated with each node.
• Total memory, which is the total local memory in the pNUMA node that the pCPU can directly access.
• Free memory, which is the free local memory in the pNUMA node that the pCPU can directly access.
• Total memory and free memory are both obtained from the real physical memory capacity of the pNUMA node.

You can also open the vNUMA topology of associated virtual machines.

Physical NIC General Configuration

You can perform the following operations on host physical NICs:

Physical NICs SR-IOV

Based on the SR-IOV specification, you can virtualize a physical NIC into multiple VF NICs and assign them directly to virtual machines. This provides I/O performance close to that of physical devices and reduces host CPU resource consumption.

Prerequisites

• Ensure that the physical NIC supports SR-IOV virtualization.
• Ensure that Intel VT-d/AMD IOMMU and SR-IOV are enabled in the BIOS of the host on which the physical NIC is installed.
• Ensure that the IOMMU status of the host on which the physical NIC is installed is Available.

Procedure

1. Navigate to Host Details Page > Hardware Device > Physical NIC.
2. Select the target physical NIC, and then click Actions > Configure SR-IOV.
3. In the Configure SR-IOV dialog, enable SR-IOV Status and specify the number of VF NICs to be partitioned.

Notes

• If the physical NIC is already configured in a bond, continuing to use SR-IOV may affect communication between VF NICs and vNICs. You are advised to partition the physical NIC through SR-IOV first and then configure the bond on the distributed switch. Select only one virtualized physical NIC on each host.
• If VF NICs are being used by virtual machines, disabling SR-IOV Status also detaches the related NICs from the virtual machines.
• Virtual machines that are powered on and have attached VF NICs do not support migration. Power off the virtual machine or detach the VF NICs before migration.

View and Maintain NIC LLDP Information

Link Layer Discovery Protocol (LLDP) identifies the physical switch port connected to a physical NIC of a specified distributed switch. Before you view peer device information, ensure that the NIC supports LLDP and that LLDP is enabled on the peer switch device.

Recommended Hardware Specifications

• Switches: Huawei switches, H3C switches, and Shengke switches.
• NICs: Intel 82599ES, Intel x710, Intel x722, and Mellanox CX4.

Supported OS Types

• x86: H84r and x86_KylinV10P3.
• ARM: arm_KylinV10P3 and H22e.

View peer device information:

1. Click the physical NIC name to open the details page.
2. Click LLDP to view peer device information. The following table describes the LLDP TLV units.

Modify the LLDP mode:

1. Click the physical NIC name to open the details page.
2. On the NIC details page, click LLDP.
3. On the LLDP tab, click Modify.
4. In the Modify LLDP Mode dialog, select the NIC LLDP mode from the drop-down list.
5. Click OK.

View and Configure Bonds

If physical NICs were associated with a distributed switch for aggregation when the distributed switch was created, you can view the bond on the Bond tab of the corresponding host physical NIC.

Click the refresh button to view the latest information, including bond mode, bond status, rate, distributed switch association, IPv4 address, and creation time. For more information about creating and modifying bonds, see Create a Distributed Switch and Modify Uplink Configuration.

Physical GPU Devices

You can enable, disable, partition, or restore physical GPU devices as needed.

Before virtualization partitioning, ensure that the physical GPU device meets the following requirements:

• The physical GPU model supports virtualization partitioning.
• The physical GPU is not passed through to a virtual machine.
• Intel VT-d / AMD IOMMU is enabled in the host BIOS, and IOMMU support is enabled in the host kernel.
• The IOMMU readiness status of the host is Available.

The virtualization partitioning method varies slightly by physical GPU vendor.

• NVIDIA: Supports partitioning a single NVIDIA physical GPU based on the selected partitioning specification.
• AMD: Supports partitioning all AMD physical GPUs on the current host based on the selected partitioning quantity.

Before restoring virtualization, ensure that all vGPUs partitioned from the physical GPU are detached from virtual machines.

• NVIDIA: Before restoring NVIDIA vGPUs, ensure that all vGPUs created from this physical GPU are detached from virtual machines.
• AMD: Before restoring AMD vGPUs, ensure that all AMD vGPUs on the current host are detached from virtual machines.

vGPU Devices

You can enable or disable vGPU devices as needed.

Attach Modes

• Direct: Attaches a USB device from the host running the virtual machine to that virtual machine. Before migrating the virtual machine, detach the USB device first.
• Forward: Attaches a USB device from any host in the same data center as the virtual machine. You do not need to detach the USB device before migrating the virtual machine.

Attach Limits

• The same USB device can be passed through to only one virtual machine.
• A single virtual machine supports up to one USB 1.0 device, six USB 2.0 devices, and four USB 3.0 devices.
• A running virtual machine, or a stopped virtual machine that uses local storage, can use only available USB devices on the host running the virtual machine. Cross-host USB device attachment is not supported.
• A stopped virtual machine that uses shared storage can use multiple USB devices from a host in the same cluster as the virtual machine.

For more information about using USB devices on virtual machines, see Create a Virtual Machine.