Virtual NUMA

In addition to its expanded processor and memory support on hosts and for VMs, Hyper-V in
Windows Server 2012 also expands support for Non-Uniform Memory Access (NUMA) from
the host into the VM. NUMA allows the use of memory by processors to be optimized based on the location of the memory with respect to the processor. High-performance applications like Microsoft SQL Server have built-in optimizations that can take advantage of the NUMA topology of a system to improve how processor threads are scheduled and memory is allocated.
In previous versions of Hyper-V, VMs were not NUMA-aware, which meant that when applications like SQL Server were run in VMs, these applications were unable to take advantage of such optimizations. Because NUMA was not used in previous versions, it was possible for a VM’s RAM to span NUMA nodes and access non-local memory. There is a performance impact when using non-local memory due to the fact that another memory controller (CPU) has to be contacted.

But with VMs now being NUMA-aware in Windows Server 2012, the performance of
applications like SQL Server can be significantly better. Note, however, that NUMA support
in VMs works in Hyper-V in Windows Server 2012 only when Dynamic Memory has not been
configured on the host.

 

How it works

Virtual NUMA presents a NUMA topology within a VM so that the guest operating system and applications can make intelligent decisions about thread and memory allocation that are reflected in the physical NUMA topology of the host. For example, Figure 2-9 shows a
NUMA-capable four-socket host machine with four physical NUMA nodes labeled 1 through 4. Two VMs are running on this host, and two virtual NUMA nodes are presented within each VM, and these virtual NUMA nodes align with physical NUMA nodes on the host based on policy. The result is that NUMA-aware applications like SQL Server installed on the guest operating system of one of these VMs would be able to allocate its thread and memory resources as if it was running directly upon a physical server that had two NUMA nodes.

 

Virtual NUMA and failover clustering

Virtual NUMA support also extends into high-availability solutions built using failover clustering in Windows Server 2012. This enables the failover cluster to place VMs more appropriately by evaluating the NUMA configuration of a node before moving a VM to the node to ensure the node is able to support the workload of the VM. This NUMA-awareness for VMs in failover clustering environments helps reduce the number of failover attempts which results in increased uptime for your VMs. See Chapter 3 for more information concerning failover clustering enhancements in Windows Server 2012.

 

Network adapter hardware acceleration

Besides the increased processor and memory support available for both hosts and VMs,
Windows Server 2012 also supports various hardware acceleration features of high-end network adapter hardware to ensure maximum scalability and performance in cloud scenarios. As Figure 2-10 shows, most of these features can be enabled in the Hyper-V Settings of Hyper-V Manager, provided that your network adapter hardware supports these functionalities.

 

Virtual Machine Queue (VMQ)

Virtual Machine Queue (VMQ) was first available for the Hyper-V role in Windows Server
2008 R2 for host machines that had VMQ-capable network adapter hardware. VMQ employs hardware packet filtering to deliver packets from an external VM network directly to VMs using Direct Memory Access (DMA) transfers. This helps reduce the overhead of routing packets from the host to the VM, which helps improve the performance of the host operating