Receive Side Scaling
The phrase “Receive Side Scaling” (RSS) refers to the idea that all receive data processing is shared (scaled)
across multiple processors or processor cores. Without RSS, all receive data processing is performed by a
single processor, resulting in less efficient system cache utilization.
RSS can be enabled for a LAN or for FCoE. In the first case, it is called “LAN RSS”. In the second, it is called
“FCoE RSS”.
LAN RSS
LAN RSS applies to a particular TCP connection. Note the following:
l
RSS has no effect if your system has only one processing unit.
l
RSS must be enabled for Intel® I/O Acceleration Technology to function.
l
RSS is not supported on some adapters configured to use Virtual Machine Queues (VMQ). On these
adapters VMQ will take precedence over RSS. RSS will appear disabled.
Configuration
LAN RSS is enabled on the Advanced tab of the adapter property sheet. If your adapter or operating system
does not support it, the RSS settings will not be displayed. If your adapter does support it, the following
settings will be displayed:
l
Port NUMA Node
: the NUMA node number of a device.
l
Receive Side Scaling Queues
: allocates queue space to buffer transactions between the network
adapter and CPU(s). Range:
1 queue is used when low CPU utilization is required.
2 queues are used when good throughput and low CPU utilization are required.
4 or more queues are used for applications that demand maximum throughput and
l
Starting RSS CPU
: allows you to set the preferred starting LAN RSS processor. Change this setting
if the current processor is dedicated to other processes. The setting range is from 0 to the number of
logical CPUs - 1. In Microsoft* Windows Server* 2008 R2, LAN RSS will only use CPUs in group 0
(CPUs 0 through 63).
l
Preferred NUMA Node
: allows you to choose the preferred NUMA (Non-Uniform Memory Access)
node to be used for memory allocations made by the network adapter. In addition the system will
attempt to use the CPUs from the preferred NUMA node first for the purposes of LAN RSS. On NUMA
platforms, memory access latency is dependent on the memory location. Allocation of memory from
the closest node helps improve performance. The Windows Task Manager shows the NUMA Node ID
for each processor. Note that this setting only affects NUMA systems; it will have no effect on non-
NUMA systems.
Teaming
If RSS is not enabled for all adapters in a team, RSS will automatically be disabled for the team. If an adapter
that does not support RSS is added to a team, RSS will automatically be disabled for the team. If you create a
Multi-Vendor Team, you must manually verify that the RSS settings for all adapters in the team are the same.
FCoE RSS
If FCoE is installed, FCoE RSS is enabled and applies to FCoE receive processing that is shared across
processor cores.
