DPDK applications inside kubernetes pods are failing to come up when CPU and hugepage memory are from different NUMA nodes

http://bugs.dpdk.org/show_bug.cgi?id=1857

Bug ID: 1857
Summary: DPDK applications inside kubernetes pods are failing
to come up when CPU and hugepage memory are from different NUMA nodes
Product: DPDK
Version: 25.11
Hardware: All
OS: All
Status: UNCONFIRMED
Severity: major
Priority: Normal
Component: core
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---

DPDK sample applications like l2fwd, l3fwd are not coming up and failing with below error when they are run with cores assigned from NUMA-1 , but hugepages are allocated on NUMA-0.

===========
EAL: set_mempolicy failed: Invalid argument][N/A][N/A]
EAL: Using IOMMU type 1 (Type 1)][N/A][N/A]
EAL: set_mempolicy failed: Invalid argument][N/A][N/A]
PCI_BUS: Probe PCI driver: net_iavf (8086:154c) device: 0000:31:04.6 (socket
0)][N/A][N/A]
EAL: set_mempolicy failed: Invalid argument][N/A][N/A]
ETHDEV: Cannot allocate ethdev shared data][N/A][N/A]
PCI_BUS: Releasing PCI mapped resource for 0000:31:04.6][N/A][N/A]
PCI_BUS: Calling pci_unmap_resource for 0000:31:04.6 at
0x312000400000][N/A][N/A]

===========

When a kubernetes pod is launched , cores and memory/hugepages are allocated to the pod from containerd and we dont have much control of the numa nodes from where they are allocated (we are using best-effort in the topology manager configuration of the kubelet).
In this scenario, we are allocated cpus from NUMA 1 and hugepages from NUMA 0 . DPDK application is trying to allocated memory from the same numa node on which core is present and it is failing as there are no hugepages available on that particular NUMA node (Node 1).

As of now DPDK application design works only when both CPU and hugepages are from same NUMA node , but with increase in DPDK application usage inside kubernetes pods, this problem should be addressed where users do not have much control of resource reservation from same NUMA nodes.

So DPDK applications should work when cores and hugepages are from different NUMA nodes.

You are receiving this mail because:
You are the assignee for the bug.