NUMA optimization in XenServer 9

Last published : Oct 01, 2026
Related information
Monitor and manage your deployment

Overview

Non-Uniform Memory Access (NUMA) is a multiprocessor memory architecture in which memory is divided into nodes attached to specific CPU sockets. Accessing memory on the local NUMA node has lower latency than accessing memory on a remote node. When VM vCPUs and their memory are co-located on the same NUMA node, workloads can achieve lower memory-access latency and better performance.

UMA and NUMA compared

  • UMA (Uniform Memory Access) hosts present a single physical memory node to CPUs.
  • NUMA hosts expose multiple memory nodes. CPUs have lower latency when accessing memory from their local node than from remote nodes.

Benefits of optimized NUMA placement

  • Improved application responsiveness and workload throughput.
  • Higher VM density per host, reducing the need for additional hosts and overall infrastructure cost.

Limitations

  • If no single NUMA node on a host has enough free memory for a VM, XS9 starts the VM with memory spread across multiple nodes to allow the VM to start.
  • When memory is spread, the VM uses multiple NUMA nodes (NUMA nodes > 1) and is not NUMA-optimized.
  • Performance improvements vary by workload, benchmark, and CPU type.

Monitor and troubleshoot

Per-VM RRD metrics (disabled by default) help monitor NUMA behavior and vCPU scheduling.

Metrics

  • NUMA nodes (numa_nodes): Number of NUMA nodes used by the VM. Value 1 indicates optimized; values greater than 1 indicate not optimized.
  • Memory NUMA node X (memory_numa_node_X): Amount of VM memory allocated from physical NUMA node X. Value 0 means no memory from that node.
  • vCPUs runnable any (runnable_any): Percentage of time at least one vCPU is ready but no pCPU is available. This is an early indicator of CPU contention.
  • vCPUs runnable sum (runnable_vcpus): Expected percentage slowdown across all vCPUs due to scheduling delays.
  • NUMA non-affine vCPUs (numa_node_nonaffine_vcpus): Percentage of time vCPUs run on pCPUs that are not in the same NUMA node as the VM memory. Higher values indicate increased memory-access latency.
Numa

Debug checklist

  1. Check NUMA nodes to see how many NUMA nodes the VM uses.
  2. Check vCPUs runnable any for domain-level scheduling contention.
  3. Check vCPUs runnable sum for vCPU-level scheduling delays.
  4. Check NUMA non-affine vCPUs for NUMA affinity violations and related latency.

Enable metrics and export samples

Per-VM RRD metrics are disabled by default. To enable a metric (for example, runnable_any) for a VM and export a sample:
# Enable the data source for a VM
xe vm-data-source-record uuid=<vm-uuid> data-source=runnable_any

# Export sample RRD output
rrd2csv VM:AVERAGE::runnable_any

Configure NUMA placement (xe CLI)

NUMA placement optimization is enabled by default on XS9 hosts. No action is required to turn it on.
To disable NUMA placement optimization on a host, set the host NUMA affinity policy to any.
# View current policy
xe host-param-get uuid=<host-uuid> param-name=numa-affinity-policy

# Disable optimization
xe host-param-set uuid=<host-uuid> numa-affinity-policy=any

# Verify
xe host-param-get uuid=<host-uuid> param-name=numa-affinity-policy
When enabled by default, the policy value is default_policy. When disabled, the value is any.