Difference between revisions of "Getting Started with DPDK and UHD"
(→References: ad 23.11) |
(→Dependencies: UHD 4.7 DPDK support) |
||
(12 intermediate revisions by the same user not shown) | |||
Line 59: | Line 59: | ||
* Intel X710-DA4 (4x10Gb) | * Intel X710-DA4 (4x10Gb) | ||
* Intel XL710-QDA2 (2x40Gb breakout to 4x10Gb) | * Intel XL710-QDA2 (2x40Gb breakout to 4x10Gb) | ||
− | * Intel E810-CQDA1 | + | * Intel E810-CQDA1 (1x100Gb and 1x4x10Gb) |
+ | * Intel E810-CQDA2 (1x100Gb and 2x4x10Gb; [https://edc.intel.com/content/www/us/en/design/products/ethernet/perf-tuning-guide-800-series-linux/%E2%80%8Bcheck-system-hardware-capabilities/ note that this NIC does not support 2x100Gb]) | ||
+ | * Intel E810-2CQDA2 (1x100Gb and 2x4x10Gb; 2x100Gb with some work) | ||
* Mellanox MCX4121A-ACAT ConnectX-4 Lx (2x10Gb) | * Mellanox MCX4121A-ACAT ConnectX-4 Lx (2x10Gb) | ||
* Mellanox MCX515A-CCAT ConnectX-5 EN (2x100Gb or 2x10Gb) | * Mellanox MCX515A-CCAT ConnectX-5 EN (2x100Gb or 2x10Gb) | ||
Line 85: | Line 87: | ||
* UHD 4.0 and 4.1 require DPDK 18.11 | * UHD 4.0 and 4.1 require DPDK 18.11 | ||
− | * UHD 4.2 through 4. | + | * UHD 4.2 through 4.7 can use any version of DPDK from 18.11 through and including 21.11 |
== Installing DPDK == | == Installing DPDK == | ||
Line 108: | Line 110: | ||
Once the <code>dpdk</code> and <code>dpdk-dev</code> packages are installed, UHD will locate them during a build and you should see DPDK in the enabled components lists when running <code>cmake</code>. | Once the <code>dpdk</code> and <code>dpdk-dev</code> packages are installed, UHD will locate them during a build and you should see DPDK in the enabled components lists when running <code>cmake</code>. | ||
+ | |||
+ | NOTE that in general UHD installed from PPA or system packages ''does not'' include support for DPDK, and even if DPDK is installed alongside these UHD it will not be used. In order to get UHD with DPDK support, [https://kb.ettus.com/Building_and_Installing_the_USRP_Open-Source_Toolchain_(UHD_and_GNU_Radio)_on_Linux UHD generally has to be built from source]. | ||
== Enable hugepages == | == Enable hugepages == | ||
Line 133: | Line 137: | ||
ip a | ip a | ||
− | You must then create a UHD configuration file | + | You must then create a UHD configuration file. |
+ | |||
+ | For UHD 3 the location is <code>/root/.uhd/uhd.conf</code>. | ||
sudo su | sudo su | ||
mkdir -p /root/.uhd | mkdir -p /root/.uhd | ||
nano /root/.uhd/uhd.conf | nano /root/.uhd/uhd.conf | ||
+ | |||
+ | For UHD 4 the location is <code>/root/.config/uhd.conf</code>. | ||
+ | |||
+ | sudo su | ||
+ | mkdir -p /root/.config | ||
+ | nano /root/.config/uhd.conf | ||
=== UHD 3.x === | === UHD 3.x === | ||
Line 179: | Line 191: | ||
You must verify and/or update the following fields for your configuration from this example: | You must verify and/or update the following fields for your configuration from this example: | ||
− | * The MAC address variables, <code>dpdk_mac</code>, to match your NIC(s) link(s) | + | * The MAC address variables, <code>dpdk_mac</code>, to match your NIC(s) link(s); note that the MAC address info ''must be lowercase'' |
* The <code>dpdk_driver</code> if the location is different on your system. <code>/usr/local/lib/</code> is the default location on when DPDK is built and installed from source. | * The <code>dpdk_driver</code> if the location is different on your system. <code>/usr/local/lib/</code> is the default location on when DPDK is built and installed from source. | ||
Line 201: | Line 213: | ||
dpdk_lcore = 3 | dpdk_lcore = 3 | ||
dpdk_ipv4 = 192.168.10.1/24 | dpdk_ipv4 = 192.168.10.1/24 | ||
+ | dpdk_num_desc = 4096 | ||
[dpdk_mac=aa:bb:cc:dd:ee:f2] | [dpdk_mac=aa:bb:cc:dd:ee:f2] | ||
dpdk_lcore = 4 | dpdk_lcore = 4 | ||
dpdk_ipv4 = 192.168.20.1/24 | dpdk_ipv4 = 192.168.20.1/24 | ||
+ | dpdk_num_desc = 4096 | ||
'''Notes:''' | '''Notes:''' | ||
Line 388: | Line 402: | ||
isolcpus=2-4 | isolcpus=2-4 | ||
− | NOTE: After saving the <code>GRUB_CONFIG</code> file, execute <code>sudo update-grub</code> and then reboot the computer for the changes to take effect. | + | NOTE: After saving the <code>GRUB_CONFIG</code> file, execute <code>sudo update-grub</code> and then reboot the computer for the changes to take effect. It is OK to isolate core used by DPDK via this method. |
=== Disable System Interrupts on Cores/CPUs === | === Disable System Interrupts on Cores/CPUs === | ||
Line 396: | Line 410: | ||
nohz_full=2-4 rcu_nocbs=2-4 | nohz_full=2-4 rcu_nocbs=2-4 | ||
− | NOTE: After saving the <code>GRUB_CONFIG</code> file, execute <code>sudo update-grub</code> and then reboot the computer for the changes to take effect. | + | NOTE: After saving the <code>GRUB_CONFIG</code> file, execute <code>sudo update-grub</code> and then reboot the computer for the changes to take effect. It is OK to isolate core used by DPDK via this method. |
=== Streaming on Multiple Channels using 1 Thread Per Stream === | === Streaming on Multiple Channels using 1 Thread Per Stream === | ||
Line 417: | Line 431: | ||
taskset -c "2-4" | taskset -c "2-4" | ||
+ | |||
+ | NOTE: For best performance do ''not'' include the cores used by DPDK in the <code>dpdk_corelist</code> argument in the <code>taskset</code>. | ||
=== Stopping Extraneous Processes === | === Stopping Extraneous Processes === | ||
Line 422: | Line 438: | ||
The Linux kernel spawns a number of processes and threads that spend most of their time sleeping; one cannot stop these processes/threads. That said, in a typical Linux install (for example Ubuntu) there will be a graphical desktop (e.g., <code>gdm</code>) and various daemons started up by user request (e.g., <code>containerd</code>, <code>docker</code>, <code>snapd</code>). Most of these daemons can be controlled by <code>systemctl</code>; some must be manually quit. Any process that runs regularly stands a chance of interrupting the UHD and DPDK threads, and thus stopping, quitting, or disabling those processes can increase performance. For example, the following commands stop various daemons that execute regularly on Ubuntu; these need to be executed with <code>sudo</code> and some might not exist on your specific system and for these the command will do nothing so it's OK to run it: | The Linux kernel spawns a number of processes and threads that spend most of their time sleeping; one cannot stop these processes/threads. That said, in a typical Linux install (for example Ubuntu) there will be a graphical desktop (e.g., <code>gdm</code>) and various daemons started up by user request (e.g., <code>containerd</code>, <code>docker</code>, <code>snapd</code>). Most of these daemons can be controlled by <code>systemctl</code>; some must be manually quit. Any process that runs regularly stands a chance of interrupting the UHD and DPDK threads, and thus stopping, quitting, or disabling those processes can increase performance. For example, the following commands stop various daemons that execute regularly on Ubuntu; these need to be executed with <code>sudo</code> and some might not exist on your specific system and for these the command will do nothing so it's OK to run it: | ||
− | systemctl stop containerd | + | systemctl stop containerd containerd.socket |
− | systemctl stop docker | + | systemctl stop docker docker.socket |
− | systemctl stop | + | systemctl stop dbus dbus.socket |
− | + | systemctl stop snapd snapd.socket | |
− | systemctl stop | + | systemctl stop udev systemd-udevd-control.socket systemd-udevd-kernel.socket |
− | systemctl stop | + | |
The following command stops the main Ubuntu desktop GUI, so running it will render the system accessible only via networking (e.g., <code>ssh</code>), so make sure network access is enabled before executing this command: | The following command stops the main Ubuntu desktop GUI, so running it will render the system accessible only via networking (e.g., <code>ssh</code>), so make sure network access is enabled before executing this command: |
Latest revision as of 08:27, 31 July 2024
Contents
- 1 Application Note Number and Authors
- 2 Overview
- 3 Abstract
- 4 Supported Devices
- 5 References
- 6 Dependencies
- 7 Installing DPDK
- 8 Installing UHD
- 9 Enable hugepages
- 10 Preparing your UHD Configuration File
- 11 Additional Host Configuration for NIC Vendors
- 12 Running UHD Applications with DPDK
- 13 Tuning Notes
- 13.1 General Host Performance Tuning App Note
- 13.2 Increasing num_recv_frames
- 13.3 Full Rate Streaming (UHD 3.x only)
- 13.4 Full Rate on X3xx
- 13.5 Isolate Cores/CPUs
- 13.6 Disable System Interrupts on Cores/CPUs
- 13.7 Streaming on Multiple Channels using 1 Thread Per Stream
- 13.8 Elevated Streaming Thread Priority
- 13.9 Extra nice Priority
- 13.10 Limit Execution CPUs/Cores
- 13.11 Stopping Extraneous Processes
- 13.12 Disable Hyper-threading
- 13.13 Additional Tuning Notes from Intel
- 14 Known Issues / Troubleshooting
Application Note Number and Authors
AN-500 by Nate Temple, Alex Williams, Wade Fife, Matt Prost, and Michael Dickens
Overview
This application note walks through the process to get started with the Data Plane Development Kit (DPDK) driver within UHD.
Abstract
Up until now, UHD's only support for networked devices was backed by the kernel's sockets implementation. Every call to send()
or recv()
would cause a context switch and invite the kernel's scheduler to replace our thread with something else. Because the typical scheduler is optimized to distribute CPU time fairly across multiple loads, the timing-critical threads might sporadically be hit with sleeping time, and the thread might be migrated off its current CPU and forced to run on another. The overhead and random latency spikes make it difficult to enable reliable real-time streaming at higher rates.
DPDK is a high-speed packet processing framework that enables a kernel bypass for network drivers. By putting the entire driver in user space, avoiding context switches, and pinning I/O threads to cores, UHD and DPDK combine to largely prevent the latency spikes induced by the scheduler. In addition, the overall overhead for packet processing lowers.
Supported Devices
USRPs
DPDK is supported on the following USRP devices:
- E320
- N300 / N310
- N320 / N321
- X300 / X310
- X410
- X440
Host Network Cards
DPDK is supported on many Intel and Mellanox based 10Gb and 100Gb NICs and lots of other NICs. Below is a list of NICs Ettus Research has tested. For a full list of NICs supported by DPDK, please see the DPDK manual.
- Intel X520-DA1 (1x10Gb)
- Intel X520-DA2 (2x10Gb)
- Intel X710-DA2 (2x10Gb)
- Intel X710-DA4 (4x10Gb)
- Intel XL710-QDA2 (2x40Gb breakout to 4x10Gb)
- Intel E810-CQDA1 (1x100Gb and 1x4x10Gb)
- Intel E810-CQDA2 (1x100Gb and 2x4x10Gb; note that this NIC does not support 2x100Gb)
- Intel E810-2CQDA2 (1x100Gb and 2x4x10Gb; 2x100Gb with some work)
- Mellanox MCX4121A-ACAT ConnectX-4 Lx (2x10Gb)
- Mellanox MCX515A-CCAT ConnectX-5 EN (2x100Gb or 2x10Gb)
- Mellanox MCX516A-CCAT ConnectX-5 EN (2x100Gb or 2x10Gb)
- Mellanox MCX516A-CDAT ConnectX-5 Ex EN (2x100Gb or 2x10Gb)
- Mellanox MCX623106AN-CDAT ConnectX-6 Dx EN (2x100Gb or 2x10Gb)
- NI Dual 100 Gigabit Ethernet PCIe Interface Kit (PN 788216-01)
References
- DPDK: https://www.dpdk.org/
- UHD Manual: https://files.ettus.com/manual/page_dpdk.html
Dependencies
- UHD 3.x requires DPDK 17.11
- UHD 4.0 and 4.1 require DPDK 18.11
- UHD 4.2 through 4.7 can use any version of DPDK from 18.11 through and including 21.11
Installing DPDK
We recommend installing DPDK via the system-provided installer; for example with Ubuntu:
sudo apt install dpdk dpdk-dev
While it is possible to install DPDK from source, we recommend using the system-provided install unless there is a very good reason to not do so. DPDK can be challenging to correctly build from source. If you require installing DPDK from source, the install guide for various versions is noted below:
DPDK releases come twice per year in April (version X.04) and November (version X.11). The first release is more of a beta while the second is more of a formal release. While either can be used with UHD, we recommend just the formal releases with version ending in 11.
Installing UHD
Once the dpdk
and dpdk-dev
packages are installed, UHD will locate them during a build and you should see DPDK in the enabled components lists when running cmake
.
NOTE that in general UHD installed from PPA or system packages does not include support for DPDK, and even if DPDK is installed alongside these UHD it will not be used. In order to get UHD with DPDK support, UHD generally has to be built from source.
Enable hugepages
Edit your grub configuration file, /etc/default/grub
and add the follow parameters to GRUB_CMDLINE_LINUX_DEFAULT
:
iommu=pt intel_iommu=on hugepages=2048
On a vanilla Ubuntu system it should look like this:
GRUB_CMDLINE_LINUX_DEFAULT="quiet splash iommu=pt intel_iommu=on hugepages=2048"
Close /etc/default/grub
and at the command prompt, update your grub configuration with the command:
sudo update-grub
For these settings to take effect, reboot your host machine.
Preparing your UHD Configuration File
You must note the MAC addresses for your NICs before proceeding.
The MAC addresses for your NICs can be found by running the command:
ip a
You must then create a UHD configuration file.
For UHD 3 the location is /root/.uhd/uhd.conf
.
sudo su mkdir -p /root/.uhd nano /root/.uhd/uhd.conf
For UHD 4 the location is /root/.config/uhd.conf
.
sudo su mkdir -p /root/.config nano /root/.config/uhd.conf
UHD 3.x
An example uhd.conf
file is listed below. Note that field names in UHD 3.x are slightly different from UHD 4.0.
You should update the following fields for your configuration from this example:
- Update the MAC address variables,
dpdk-mac
, to match your NIC
- Update the
dpdk-driver
if the location is different on your system./usr/lib/x86_64-linux-gnu/dpdk-17.11-drivers/
is the default location on Ubuntu 18.04.x whendpdk
is installed viaapt
.
- Update the
dpdk-corelist
anddpdk-io-cpu
fields. In this example, a two port NIC is used. There should be one core for the maindpdk
thread (in this examplecore #2
), and then separate cores assigned to each NIC (in this examplecore #3
for the first port on the NIC,core #4
for the second port on the NIC)
- Update the
dpdk-ipv4
fields to your desired IP range.-
192.168.30.2
,192.168.40.2
on a default X3xx system -
192.168.10.2
,192.168.20.2
on a default N3xx system -
192.168.10.2
on a default E320 system
-
[use_dpdk=1] dpdk-mtu=9000 dpdk-driver=/usr/lib/x86_64-linux-gnu/dpdk-17.11-drivers/ dpdk-corelist=2,3,4 dpdk-num-mbufs=4095 dpdk-mbufs-cache-size=315 [dpdk-mac=aa:bb:cc:dd:ee:f1] dpdk-io-cpu = 3 dpdk-ipv4 = 192.168.10.1/24 [dpdk-mac=aa:bb:cc:dd:ee:f2] dpdk-io-cpu = 4 dpdk-ipv4 = 192.168.20.1/24
Note: Additional information on the UHD configuration file can be found here: https://files.ettus.com/manual_archive/v3.15.0.0/html/page_dpdk.html#dpdk_nic_config
UHD 4.x
An example uhd.conf
file is listed below. Note that the field names in UHD 4.x are slightly different from UHD 3.x.
You must verify and/or update the following fields for your configuration from this example:
- The MAC address variables,
dpdk_mac
, to match your NIC(s) link(s); note that the MAC address info must be lowercase
- The
dpdk_driver
if the location is different on your system./usr/local/lib/
is the default location on when DPDK is built and installed from source.
- The
dpdk_corelist
anddpdk_lcore
fields. In this example, a two port NIC and both links are used. There must be one core for the maindpdk
thread (in this examplecore #2
because it is not otherwise used in the file) and then separate cores assigned to each NIC link (in this examplecore #3
for the one of the NIC links andcore #4
for the other NIC link).
- The
dpdk_ipv4
fields to your desired IP range(s).-
192.168.30.2
,192.168.40.2
on a default X3xx system -
192.168.10.2
,192.168.20.2
on a default N3xx system -
192.168.10.2
on a default E320 system -
192.168.10.2
and192.168.20.2
on a default X4xx systems
-
[use_dpdk=1] dpdk_mtu=9000 dpdk_driver=/usr/lib/x86_64-linux-gnu/dpdk/pmds-20.0/ dpdk_corelist=2,3,4 dpdk_num_mbufs=4095 dpdk_mbuf_cache_size=315 [dpdk_mac=aa:bb:cc:dd:ee:f1] dpdk_lcore = 3 dpdk_ipv4 = 192.168.10.1/24 dpdk_num_desc = 4096 [dpdk_mac=aa:bb:cc:dd:ee:f2] dpdk_lcore = 4 dpdk_ipv4 = 192.168.20.1/24 dpdk_num_desc = 4096
Notes:
- Additional information on the UHD DPDK configuration file is in the UHD manual.
- The number of cores listed must be used exactly within the file.
- The number of NIC link(s) listed must exactly match those specified in the UHD instantiation args.
- A simple way to move between different DPDK configurations is to create different files and then symlink from the desired file to
uhd.conf
Additional Host Configuration for NIC Vendors
The process for this step is different for Intel and Mellanox NICs and is detailed in individual sections below.
Intel X520 / X710
The Intel based NICs will use the vfio-pci
driver which must be loaded:
sudo modprobe vfio-pci
Next, you will need to rebind the NIC to the vfio-pci
drivers.
First, identify the PCI address your NIC is at:
dpdk-devbind --status
Note the PCI address that your NIC is connected to for the next step.
Before the next step, you will need to turn off the NIC first before doing the rebind.
In Ubuntu under System
-> Network
-> click the switches to off
for the 10Gb ports, then run the dpdk-devbind
commands:
Note: Your PCI address will likely be different than 02:00.0
as shown in the example below.
sudo dpdk-devbind --bind=vfio-pci 02:00.0 sudo dpdk-devbind --bind=vfio-pci 02:00.1
Now if you run dpdk-devbind --status
again, you should see the NICs listed under DPDK devices
# dpdk-devbind --status Network devices using DPDK-compatible driver ============================================ 0000:02:00.0 '82599ES 10-Gigabit SFI/SFP+ Network Connection 10fb' drv=vfio-pci unused=ixgbe 0000:02:00.1 '82599ES 10-Gigabit SFI/SFP+ Network Connection 10fb' drv=vfio-pci unused=ixgbe
Note: More info can be found here on the rebinding process: https://doc.dpdk.org/guides-17.11/linux_gsg/linux_drivers.html#binding-and-unbinding-network-ports-to-from-the-kernel-modules
Mellanox NICs
The Mellanox NICs do not require rebinding using the vfio-pci
driver. Mellanox provides additional drivers for DPDK.
Install and activate the Mellanox drivers:
sudo apt install librte-pmd-mlx5-17.11 sudo modprobe -a ib_uverbs mlx5_core mlx5_ib
For 18.11 you can download and install the latest Mellanox drivers from the mellanox website (https://www.mellanox.com/products/infiniband-drivers/linux/mlnx_ofed).
The MLX5 poll mode driver library (librte_pmd_mlx5) in DPDK provides support for Mellanox ConnextX-4 and ConnectX-5 cards. This driver must be enabled manually with the build option CONFIG_RTE_LIBRTE_MLX5_PMD=y
when building DPDK.
Running UHD Applications with DPDK
UHD based application (including GNU Radio flowgraphs) can now be ran using a DPDK transport by passing in the Device Argument: use_dpdk=1
.
Important Note: In order for UHD to use DPDK, the UHD application *must* be ran as the root
user. Using sudo
will not work, you should switch to the root
user by running sudo su
.
For example, running the benchmark_rate
utility:
# cd /usr/local/lib/uhd/examples # ./benchmark_rate --rx_rate 125e6 --rx_subdev "A:0 B:0" --rx_channels 0,1 --tx_rate 125e6 --tx_subdev "A:0 B:0" --tx_channels 0,1 --args "addr=192.168.10.2,second_addr=192.168.20.2,mgmt_addr=10.2.1.19,master_clock_rate=125e6,use_dpdk=1" [INFO] [UHD] linux; GNU C++ version 7.3.0; Boost_106501; UHD_3.14.0.HEAD-0-gabf0db4e EAL: Detected 8 lcore(s) EAL: Some devices want iova as va but pa will be used because.. EAL: IOMMU does not support IOVA as VA EAL: No free hugepages reported in hugepages-1048576kB EAL: Probing VFIO support... EAL: VFIO support initialized EAL: PCI device 0000:02:00.0 on NUMA socket -1 EAL: Invalid NUMA socket, default to 0 EAL: probe driver: 8086:10fb net_ixgbe EAL: using IOMMU type 1 (Type 1) EAL: Ignore mapping IO port bar(2) EAL: PCI device 0000:02:00.1 on NUMA socket -1 EAL: Invalid NUMA socket, default to 0 EAL: probe driver: 8086:10fb net_ixgbe EAL: Ignore mapping IO port bar(2) PMD: ixgbe_dev_link_status_print(): Port 0: Link Down EAL: Port 0 MAC: aa bb cc dd ee f1 EAL: Port 0 UP: 1 PMD: ixgbe_dev_link_status_print(): Port 1: Link Down EAL: Port 1 MAC: aa bb cc dd ee f2 EAL: Port 1 UP: 1 EAL: Init DONE! EAL: Starting I/O threads! USER2: Thread 1 started [00:00:00.000003] Creating the usrp device with: addr=192.168.10.2,second_addr=192.168.20.2,mgmt_addr=10.2.1.19,master_clock_rate=125e6,use_dpdk=1... [INFO] [MPMD] Initializing 1 device(s) in parallel with args: mgmt_addr=10.2.1.19,type=n3xx,product=n310,serial=313ABDA,claimed=False,addr=192.168.10.2,second_addr=192.168.20.2,master_clock_rate=125e6,use_dpdk=1 [INFO] [MPM.PeriphManager] init() called with device args 'product=n310,time_source=internal,master_clock_rate=125e6,clock_source=internal,use_dpdk=1,second_addr=192.168.20.2,mgmt_addr=10.2.1.19'. [INFO] [0/DmaFIFO_0] Initializing block control (NOC ID: 0xF1F0D00000000004) [INFO] [0/DmaFIFO_0] BIST passed (Throughput: 1344 MB/s) [INFO] [0/DmaFIFO_0] BIST passed (Throughput: 1341 MB/s) [INFO] [0/DmaFIFO_0] BIST passed (Throughput: 1348 MB/s) [INFO] [0/DmaFIFO_0] BIST passed (Throughput: 1347 MB/s) [INFO] [0/Radio_0] Initializing block control (NOC ID: 0x12AD100000011312) [INFO] [0/Radio_1] Initializing block control (NOC ID: 0x12AD100000011312) [INFO] [0/DDC_0] Initializing block control (NOC ID: 0xDDC0000000000000) [INFO] [0/DDC_1] Initializing block control (NOC ID: 0xDDC0000000000000) [INFO] [0/DUC_0] Initializing block control (NOC ID: 0xD0C0000000000002) [INFO] [0/DUC_1] Initializing block control (NOC ID: 0xD0C0000000000002) Using Device: Single USRP: Device: N300-Series Device Mboard 0: ni-n3xx-313ABDA RX Channel: 0 RX DSP: 0 RX Dboard: A RX Subdev: Magnesium RX Channel: 1 RX DSP: 0 RX Dboard: B RX Subdev: Magnesium TX Channel: 0 TX DSP: 0 TX Dboard: A TX Subdev: Magnesium TX Channel: 1 TX DSP: 0 TX Dboard: B TX Subdev: Magnesium [00:00:03.728707] Setting device timestamp to 0... [INFO] [MULTI_USRP] 1) catch time transition at pps edge [INFO] [MULTI_USRP] 2) set times next pps (synchronously) [00:00:05.331920] Testing receive rate 125.000000 Msps on 2 channels [00:00:05.610789] Testing transmit rate 125.000000 Msps on 2 channels [00:00:15.878071] Benchmark complete. Benchmark rate summary: Num received samples: 2557247854 Num dropped samples: 0 Num overruns detected: 0 Num transmitted samples: 2504266704 Num sequence errors (Tx): 0 Num sequence errors (Rx): 0 Num underruns detected: 0 Num late commands: 0 Num timeouts (Tx): 0 Num timeouts (Rx): 0 Done!
Tuning Notes
General Host Performance Tuning App Note
Perform the general host performance tuning tips and tricks.
Increasing num_recv_frames
If you experience Overflows
at higher data rates, adding the device argument num_recv_frames=512
can help.
Full Rate Streaming (UHD 3.x only)
If you're streaming data at the full master clock rate, and there is no interpolation or decimation being performed on the FPGA, you can skip the DUC and DDC blocks within the FPGA with the following parameters:
skip_ddc=1 skip_duc=1
Full Rate on X3xx
If you're streaming two transmit channels at full rate (200e6) on the X3xx platform, you should additionally set the following device arg:
enable_tx_dual_eth=1
Isolate Cores/CPUs
Isolating the cores that are used for DPDK can improve performance. This can be done by adding the isolcpus
parameter to your /etc/default/grub
file in the GRUB_CMDLINE_LINUX_DEFAULT=""
(the "GRUB_CONFIG
"). For example, to isolate cores 2 through and including 4, add this entry:
isolcpus=2-4
NOTE: After saving the GRUB_CONFIG
file, execute sudo update-grub
and then reboot the computer for the changes to take effect. It is OK to isolate core used by DPDK via this method.
Disable System Interrupts on Cores/CPUs
Disabling system interrupts can improve the jitter and performance generally by 1-3%. This can be done by adding the parameters nohz_full
and rcu_nocbs
to your GRUB_CONFIG
. For example, to disable system interrupts on cores 2 through and including 4, add this entry:
nohz_full=2-4 rcu_nocbs=2-4
NOTE: After saving the GRUB_CONFIG
file, execute sudo update-grub
and then reboot the computer for the changes to take effect. It is OK to isolate core used by DPDK via this method.
Streaming on Multiple Channels using 1 Thread Per Stream
If you're streaming on multiple channels simultaneously, you can create multiple streamer objects on separate threads. This can be accomplished with the benchmark_rate
example by using the parameter --multi_streamer
.
Elevated Streaming Thread Priority
In UHD 4, streaming thread priorities can be elevated with the uhd::set_thread_priority_safe()
function call. This can be accomplished with the benchmark_rate
example by using parameter --priority high
. Note that if the Thread Schedule Priority has not been enabled for the current user then this parameter requires sudo
to work.
Extra nice
Priority
Beyond elevating streaming thread priority, one can also increase the nice
priority level to maximum to increase the amount of CPU time for the process and its thread by prepending the following to the command being issued:
sudo nice -n -20
Limit Execution CPUs/Cores
With Linux one can limit the cores being used by the threads in a process via a taskset
prepended to a command being executed. When combining with the other techniques listed here, one can fairly well constrain a process and its thread to specific cores and give them maximum CPU time. Note that when using DPDK the MAC cores specified in uhd.conf
will already be included as part of the taskset
(but those cores will not be isolated nor have system interrupts disabled on them unless using that setting as noted above); it generally won't hurt to include the DPDK cores as part of the overall taskset
but it's not required. For example, to limit process/thread execution to cores 2 through and including 4, one would prepend the following (note that sudo
is not required):
taskset -c "2-4"
NOTE: For best performance do not include the cores used by DPDK in the dpdk_corelist
argument in the taskset
.
Stopping Extraneous Processes
The Linux kernel spawns a number of processes and threads that spend most of their time sleeping; one cannot stop these processes/threads. That said, in a typical Linux install (for example Ubuntu) there will be a graphical desktop (e.g., gdm
) and various daemons started up by user request (e.g., containerd
, docker
, snapd
). Most of these daemons can be controlled by systemctl
; some must be manually quit. Any process that runs regularly stands a chance of interrupting the UHD and DPDK threads, and thus stopping, quitting, or disabling those processes can increase performance. For example, the following commands stop various daemons that execute regularly on Ubuntu; these need to be executed with sudo
and some might not exist on your specific system and for these the command will do nothing so it's OK to run it:
systemctl stop containerd containerd.socket systemctl stop docker docker.socket systemctl stop dbus dbus.socket systemctl stop snapd snapd.socket systemctl stop udev systemd-udevd-control.socket systemd-udevd-kernel.socket
The following command stops the main Ubuntu desktop GUI, so running it will render the system accessible only via networking (e.g., ssh
), so make sure network access is enabled before executing this command:
systemctl stop gdm
Disable Hyper-threading
In some applications which require the highest possible CPU performance per core, disabling hyper-threading can provide roughly a 10% increase in core performance, at the cost of having fewer core threads. Hyper-threading is disabled within the BIOS and how to do this varies by motherboard manufacturer. With other techniques listed here, disabling hyper-threading should only be done as a last resort to eek absolute maximum performance from the CPU.
Additional Tuning Notes from Intel
- Performance report from Intel on DPDK 17.11: https://fast.dpdk.org/doc/perf/DPDK_17_11_Intel_NIC_performance_report.pdf
- How to get best performance with NICs on Intel platforms: https://doc.dpdk.org/guides/linux_gsg/nic_perf_intel_platform.html
Known Issues / Troubleshooting
Regular Overflows on multi-CPU Systems
On Multi-CPU systems each CPU is a NUMA node. The NUMA node contains all of the cores on the CPU. For best performance when using CPU/core isolation ("isocpus" per AAA above) make sure that all of the cores are in the same NUMA node.
Underruns Every Second with DPDK + Ubuntu
With Linux kernels 5.10 and beyond, we have observed periodic underruns on systems that otherwise have no issues. These Linux kernel versions are the default for Ubuntu 20.04.3 LTS and later. The underrun issue is due to the RT_RUNTIME_SHARE
feature being disabled by default in these versions of the Linux kernel (shown as NO_RT_RUNTIME_SHARE
). The following procedure can be used to enable this feature. This process was tested on Linux kernel version 5.13; the procedure may be slightly different on other kernel versions. To determine the Linux kernel version of your system, in a terminal issue the command uname -r
.
$ sudo -s $ cd /sys/kernel/debug/sched $ cat features | tr ' ' '\n' | grep RUNTIME_SHARE NO_RT_RUNTIME_SHARE $ echo RT_RUNTIME_SHARE > features $ cat features | tr ' ' '\n' | grep RUNTIME_SHARE RT_RUNTIME_SHARE