Intel® Ethernet 800 Series
Linux Performance Tuning Guide
Resource Limitations
Using E810 in 8x10G Mode
When E810 adapters are used in 8x10 Gb/s mode, the device can run out of hardware resources (MSI-X vectors) by the time the driver gets to the 8th port during driver initialization. No errors appear in the system logs, but performance on the 8th port can be affected, and the ethtool statistics for the device show all RX traffic for the 8th port going to RX Queue 0. The first 7 ports see the traffic evenly distributed among all the RX queues as expected.
Mitigation: This issue is fixed in ice v1.12.X and later.
Using E835 in 8x25Gb Mode
When operating Intel Ethernet 800 Series adapters based on the E835 controller (e.g., E835-CQDA2 / CQDA2 OCP3.0) in 8x25Gb breakout mode , attempting to instantiate SR-IOV Virtual Functions (VFs) across all 8 Physical Functions (PFs) may fail or result in Flow Director initialization warnings in dmesg.
Unnecessary queue pre-allocations during PF initialization-combined with default Flow Director and LAN queue reservations across all 8 physical ports-exhaust the shared global hardware queue, Virtual Station Interface (VSI), and MSI-X interrupt vector pools.
Mitigation: This issue is fixed in ice v2.7.X and later.
To provision the maximum supported number of VFs (up to 32 VFs per PF) in 8x25G mode, perform the following manual steps to reallocate MSI-X interrupt vectors from host PF queues to the VF vector pool.
Understanding the MSI-X Vector Allocation Limit (Default vs. Tuned).
Each Physical Function has a maximum hardware limit of 256 MSI-X vectors. Under default driver settings, host PF queues and RDMA reserve the majority of vectors, leaving insufficient vectors for VF allocation. By reducing host Ethernet queues (msix_eth) to 32 and RDMA vectors (msix_rdma) to 33, 123 MSI-X vectors are freed for msix_vf, allowing full SR-IOV scale (up to 32 VFs with multi-queue support
Step-by-Step Manual Configuration Procedure:
Execute the following steps on each interface
- Reduce the active combined host channels from default 64 down to 32 to free up MSI-X vectors
#ethtool -L <pf_interface> combined 32 - Reallocate vectors from host LAN/RDMA queues to the VF resource path
#devlink resource set pci/<pci_address> path msix/msix_vf size 123 - Reduce RDMA MSI-X vector allocation to 33
#devlink resource set pci/<pci_address> path msix/msix_rdma size 33 - Reload the device configuration via devlink to apply changes
#timeout 60 devlink dev reload pci/<pci_address> - Provision up to 32 VFs on the physical function
#echo 32 > /sys/class/net/<pf_interface>/device/sriov_numvfs - Verify vector allocations
#devlink resource show pci/<pci_address> - Verify created total VFs
#cat /sys/bus/pci/devices/<pci_address>/sriov_totalvfs