Table of Contents
The USRP X3x0 is capable of delivering very fast sample rates to the host PC, and even high-powered desktops can have trouble keeping up at the higher rates. You can improve the performance of your host by configuring a number of settings that affect the performance of your computer.
These are:
- Kernel Version
- Network Configuration
- Power Management Configuration
- Real-Time & Priority Scheduling
- Building with ORC & Volk
These items are covered in more detail, below.
Performance issues may be encountered with Linux kernels earlier than 3.11. Ettus Research strongly recommends using kernel version 3.11 or higher for high sample rates.
When using Ethernet interfaces to communicate with the device, it is necessary to configure a number of facets regarding your network connection.
Fedora and Ubuntu both use NetworkManager to manage network connections. Unfortunately, NetworkManager often tries to take control of a connection and will disconnect the interface.
You should open your NetworkManager configuration and tell it to ignore the network interface you are using. This is not the same as simply setting a static IP address. You must tell NetworkManager to ignore the interface.
It is necessary to increase the maximum size of the socket buffers to avoid potential overflows and underruns at high sample rates. Add the following entries into /etc/sysctl.conf (root privileges required):
net.core.rmem_max=33554432 net.core.wmem_max=33554432
Either restart the system or issue the following commands:
sudo sysctl -w net.core.rmem_max=33554432 sudo sysctl -w net.core.wmem_max=33554432
In order to achieve maximum performance, we recommend setting the MTU size to 9000 for 10 GigE and 1500 for 1 GigE. It is possible to use smaller MTUs, but this can affect performance. With some NICs, setting the MTU too high can also cause issues, though. To set the MTU to 9000, you can use the following command:
sudo ifconfig <interface> mtu 9000 # For 10 GigE
Using these MTUs will set the frame sizes for UHD communication to 8000 and 1472, respectively.
In some cases, specifying the frame size manually by adding the argument "<send/recv>_frame_size=1472" can solve issues. Note that a frame size of 1472 will limit the available sampling rate, although this is likely not a problem issue on 1 GigE.
Many Linux distributions come installed with a Firewall, by default. The Firewall will often interfere with your ability to communicate with your USRP. You should configure your firewall to "trust" the interface you are using. Setting this properly depends on your OS and firewall configuration method.
On Fedora systems, you can configure the network interface mostly from one place (with the exception of the socket buffers). Each interface on your system should have a file in:
/etc/sysconfig/network-scripts/
As an example, if your 1GigE interface is "em1", your "ifcfg-em1" configuration file should look something like this, when configured for use with a USRP X3xx:
TYPE="Ethernet" BOOTPROTO="none" IPADDR0="192.168.10.1" DEFROUTE="yes" IPV4_FAILURE_FATAL="no" IPV6INIT="no" IPV6_FAILURE_FATAL="no" NAME="em1" UUID="<specific to your device>" ONBOOT="no" HWADDR"<specific to your device>" PEERDNS="yes" PEERROUTES="yes" ZONE="trusted" MTU="9000" NM_MANAGED="no"
The above file was generated and modified on a "Fedora 20" system.
Power management on the host system attempts to save power by reducing clock frequencies or even powering off devices while not in use. This can lead to significant performance issues when trying to operate at high sample rates. Ettus Research strongly recommends disabling all power management.
In Linux, the CPU governors dictate the frequency at which the CPU operates and attempt to reduce the CPU frequencies at certain times to save power. When running at high sample rates, reduction of CPU frequencies can cause significant performance issues. To prevent those issues, set the governor to "performance".
Ubuntu: 1. Install cpufrequtils:
sudo apt-get install cpufrequtils
sed s/^GOVERNOR=.*$/GOVERNOR=\"performance\"/g /etc/init.d/cpufrequtils > /etc/init.d/cpufrequtils
sudo /etc/init.d/cpufrequtils restart
Fedora:
sudo cpupower frequency-set -g performance
Enabling real-time and priority scheduling can improve the total processing throughput of your application. Priority scheduling should be enabled for UHD, and real-time scheduling can be enabled by your application.
For information regarding how to enable priority scheduling for UHD on your system, please see the General UHD Notes.
Please note that turning on real-time scheduling in your application may lock up your computer if the processor cannot keep up with the application. You should generally avoid using real-time scheduling unless you need to.
Real-time scheduling is enabled via different methods depending on your application and operating system. In GNU Radio Companion, it can be turned on in each individual flowgraph.
Especially when running high-performance applications, processing performance can be dramatically improved by SIMD instructions. UHD uses ORC to provide SIMD capability, and GNU Radio includes a SIMD library called "Volk". These should both be used to guarantee optimum performance.
ORC, the Oil Runtime Compiler, is a third-party compiler that UHD uses to create efficient SIMD code for your particular computer. ORC is generally easily installed from your OS's package manager.
On Fedora:
$ sudo yum update; sudo yum install orc-compiler orc-devel
On Ubuntu:
$ sudo apt-get update; sudo apt-get install liborc-<version> liborc-<version>-dev
After installing ORC, when building UHD from source, you should see "ORC" as one of the configured UHD components.
-- ######################################################
-- # UHD enabled components
-- ######################################################
-- * LibUHD
<cut for brevity>
-- * ORC
If you are using GNURadio to build applications, you should compile GNURadio with Volk. For instructions on how to do this, refer to the GNURadio wiki.
Testing has shown that some motherboards do not provide enough PCIe bus bandwidth to support higher sample rates. Motherboards with PCIe 3.0 are required and the PCIe architecture of the motherboard should be carefully considered. Slots with dedicated PCIe lanes should be used for PCIe or 10GbE cards that will be connected to the X3x0 device.
Intel or Myricom 10GbE NICs are recommended. Mellanox, SolarFlare, and Chelsio 10GbE NICs are not currently recommended. The Ethernet card should be plugged into the slot that has the most direct connection with the CPU (PCIe lanes are not shared with another slot). Refer to the motherboard manual for more information on PCIe architecture.
The output on the host console provides indicators of performance issues in the form of single upper-case letters. The following table lists the letters, their meanings, and possible causes:
| Indicator | Meaning | Possible Causes |
|---|---|---|
| O | Overflow on RX |
|
| D | Dropped packet on RX |
|
| U | Underflow on TX |
|
| L | Late packet (usually on MIMO TX) |
|
| S | Sequence error on TX |
|