Performance Optimization

This section presents some guidelines to get optimal performance from the fast path with FPN-SDK Add-on for DPDK.

Set up your BIOS

To get optimal performance, you may have to edit your BIOS settings.

For instance, the following parameters can enhance performance on an Intel Ivy-Bridge platform:

BIOS CPU Setting	Value
Intel(R) QPI Frequency Select	Auto Max
Intel(R) Turbo Boost Technology	Disabled
Enhanced Intel SpeedStep(R) Tech	Disabled
Processor C3	Disabled
Processor C6	Disabled
Intel(R) Hyper-Threading Tech	Enabled
Active Processor Cores	All
Execute Disable Bit	Enabled
Intel(R) Virtualization Technology	Enabled
Intel(R) VT for Directed I/O	Enabled
Interrupt Remapping	Enabled
Coherency Support	Disabled
ATS Support	Enabled
Pass-through DMA Support	Enabled
Intel(R) TXT	Disabled
Enhanced Error Containment Mode	Disabled
MLC Streamer	Enabled
MLC Spatial Prefetcher	Enabled
DCU Data Prefetcher	Enabled
DCU Instruction Prefetcher	Enabled
Direct Cache Access (DCA)	Enabled
Extended ATR	0x01

See also

DPDK documentation

Allocate logical cores to the fast path

NICs, PCI slots, ports and cores

NICs are attached to PCI slots, which are attached to sockets. NICs are recognized as fast path ports.

Core - NIC relation

Cores allocation

When starting the fast path via the fast-path.sh script, some logical cores are allocated to the fast path, and other logical cores, to the other Linux processes.

Cores allocation to the fast path

You can specify which logical cores are allocated to the fast path via the /etc/fast-path.env fast path configuration file FP_MASK parameter.

Important

Never allocate core 0 to the fast path.

Example

Allocate logical cores 2, 3, 18 and 19 to the fast path:

: ${FP_MASK:=2,3,18,19}

Note

Logical core numbering can change between platforms and should be determined beforehand.

Make sure logical cores poll only ports located on the same socket

As the fast path constantly polls the NICs, it is critical that logical cores allocated to the fast path do not poll ports located on a different socket. Otherwise, memory accesses constantly traverse the processor interconnect, which may lead to a bottleneck.

Cores managing ports on another socket

Map ports and logical cores

If you set CORE_PORT_MAPPING manually in the fast path configuration file, you must map ports and logical cores to fit your network architecture, as illustrated in the topologies below. By default (when CORE_PORT_MAPPING is set to auto), the core/port mapping already matches these constraints.

To specify which fast path logical cores poll which ports, edit the fast path fast-path.env configuration file CORE_PORT_MAPPING parameter.

Topology 1: as many logical cores as ports on the same socket

When there are as many fast path logical cores as ports, each core should poll one different port:

As many cores as ports

Topology 2: not as many logical cores as ports on the same socket

When there are not as many fast path logical cores as ports, the easiest topology to implement is one of the following, where all fast path logical cores poll all ports. These topologies often offer a good support of borderline cases.

Different number of cores and ports

You can also test other configurations to better suit your needs.

Prevent packets from being processed on different sockets

For even better performance, try to prevent too many packets from being processed on different sockets. This avoids overloading the processor interconnect.

RX and TX on different sockets