The numa block is used to configure how Nomad will assign CPU cores for a task while taking the NUMA hardware topology of a node into consideration. Workloads that are sensitive to memory latency can perform significantly better when pinned to CPU cores on the same NUMA node.

job "example" {
  group "group" {
    task "task" {
      resources {
        cores = 8
        numa {
          affinity = "require"
        }
      }
    }
  }
}

In the example above, the task is requesting Nomad to reserve 8 CPU cores, all of which must be colocated on the same hardware NUMA node. The Nomad scheduler will ensure the task is placed on a Nomad node with sufficient unused CPU cores in a compatible configuration.

Configuring the numa block requires the task specifies CPU resources using the cores parameter.

numa Parameters

• affinity (string: "none") - Specifies the strategy Nomad will use when selecting CPU cores to assign to a task. Possible values are "none", "prefer", or "require".

  • none - Nomad is free to allocate CPU cores using any strategy. Nomad uses this freedom to allocate cores in such a way that minimizes the amount of fragmentation of core availability per NUMA node. It does so by bin-packing the chosen cores onto the NUMA nodes with the fewest number of cores available.
  • prefer - Nomad will select the set of CPU cores on a node that minimizes the total distance between those cores, but does not limit those CPU core selections to come from a single NUMA node.
  • require - Nomad will select a set of CPU cores that are strictly colocated on the same hardware NUMA node. If there are no Nomad nodes with a sufficient number of available cores in a compatible configuration, task placement will fail due to exhausted resources.

• devices (list<string>: nil) - Specifies which devices must be colocated on the name NUMA node, along with allocated CPU cores. Must be a subset of the devices listed in the resources block. May only be used with affinity set to require.

numa Examples

This example will allocate a 1080ti GPU and ensure it is on the same NUMA node as the 4 CPU cores reserved for the task.

resources {
  cores  = 4
  memory = 2048

  device "nvidia/gpu/1080ti" {
    count = 1
  }

  numa {
    affinity = "require"
    devices = [
      "nvidia/gpu/1080ti"
    ]
  }
}