Improving Bridging Throughput

You can improve bridging throughput with Receive Side Scaling. Starting in NSX 6.4.2, you can also use Software Receive Side Scaling to improve bridging throughput.

With Receive Side Scaling (RSS) technology, you can spread incoming traffic across different receive descriptor queues. If you assign each queue to a different CPU core, the incoming traffic can be load balanced, improving performance.

However, RSS does not work well with unknown unicast and multicast traffic. These packets end up in the default queue processed by a single CPU core, which leads to low throughput. Most of the packets received by the ESXi host performing VLAN-VXLAN bridging belong to this category, so bridging throughput is low.

Some physical NIC vendors support a feature called Default Queue Receive Side Scaling (DRSS). Using DRSS, you can configure multiple hardware queues backing up the default RX queue, spreading VLAN-VXLAN flows across multiple CPU cores.

For physical NICs that do not support DRSS (for example, ixgbe, ixgben), you can use Software Receive Side Scaling (SoftRSS) to improve bridging network throughput.

SoftRSS offloads the handling of individual flows to one of the multiple kernel worlds, so the thread which pulls packets from the NIC can process more packets. Similar to RSS, network throughput improvement when using SoftRSS has a linear correlation with CPU utilization.

For more information, see Enable Software Receive Side Scaling.