Network Latency Caused by Guest OS Checksum Offloading

Products

• VMware NSX

Issue/Introduction

When Checksum Offloading(tx-checksum-ip-generic) is enabled inside the Guest OS(Linux),

Symptom:

 - A network latency issue could be observed on specific VMs.

 - The latency disappeared immediately after disabling Checksum Offloading inside the Guest OS. No abnormalities were found in ESXi host NIC drivers, NSX (VDR) datapath behavior, or overlay encapsulation settings.

Environment

VMware vSphere 8.0

Cause

the possible cause below is likely responsible for the latency symptoms:

Inner Checksum Handling by vmxnet3 Causing Processing Delay

• When Checksum Offloading is enabled inside the Guest OS, vmxnet3 is responsible for calculating and writing the inner L4 checksum into the packet.

• If the VM is running on an older Virtual Hardware Version, the checksum metadata (csum_start, csum_offset, segmentation parameters) generated by vmxnet3 may not fully align with the capabilities or expectations of downstream components (ESXi, VDR, physical NIC drivers).

• This mismatch can cause ESXi/NSX to reprocess or recalculate checksums during GENEVE encapsulation, potentially leading to increased latency and reduced throughput.

• The fact that disabling Guest OS Checksum Offloading immediately resolved the symptoms indicates that the inner checksum path handled by vmxnet3 was the likely bottleneck.

Resolution

1. Upgrade the Virtual Hardware Version

 - Upgrade the Virtual Hardware Version to the latest recommended level.

2. After upgrading the Hardware Version, re-enable Checksum Offloading inside the Guest OS and verify whether the latency issue remains resolved.