From DD-WRT Wiki

(Difference between revisions)
Jump to: navigation, search
Revision as of 22:35, 30 May 2021 (edit)
Tmittelstaedt (Talk | contribs)
(Flow Acceleration, SFE and Cut-Through Forwarding)
← Previous diff
Revision as of 22:36, 30 May 2021 (edit) (undo)
Tmittelstaedt (Talk | contribs)
(Flow Acceleration, SFE and Cut-Through Forwarding)
Next diff →
Line 162: Line 162:
From a discussion on [https://forum.dd-wrt.com/phpBB2/viewtopic.php?t=329126 Build r46690 ] From a discussion on [https://forum.dd-wrt.com/phpBB2/viewtopic.php?t=329126 Build r46690 ]
-"ctf is much faster. with sfe i reach about 700 mbit with lan->wan tests. with ctf i get full gbit speed. so 950 - 980 mbit and lower cpu usage"+"ctf is much faster. with sfe i reach about 700 mbit with lan->wan tests. with ctf i get full gbit speed. so 950 - 980 mbit and lower cpu usage.... i'm running it here for 2 weeks now on several routers (r7000, ac66, n66)"
--BrainSlayer --BrainSlayer
'''What kind of speed boost does this give me?''' '''What kind of speed boost does this give me?'''

Revision as of 22:36, 30 May 2021



NVRAM (non-volatile RAM) also called flash space is the place where the permanent settings are stored. This includes

You can run the following commands using Command Line:

  • Show the whole NVRAM content:
nvram show
  • Show NVRAM content containing the pattern <search_pattern> (useful for quickly finding things in NVRAM):
nvram show | grep <search_pattern>
Replace <replace_pattern> with what you are actually looking for. This can simply be a word or a regular expression
  • Show value of a certain variable:
nvram get <variable_name>
  • Change variable value in RAM only:
nvram set <variable_name>="<value>"
(Quotes not needed for numeric/boolean values, needed for texts like or 443 8080 or sometext).
Now you can play with the new settings or test user scripts but this new value will be lost after reboot unless you do nvram commit.
  • Delete a variable (and its value). Useful when you made a typo in <variable_name> above:
nvram unset <variable_name>
  • Save all changed variables to NVRAM:
nvram commit

Reboot the router for new settings in NVRAM to go into effect.

  • To erase the nvram block when running DD-WRT:
erase nvram
  • To erase nvram in the Broadcom CFE (bootloader):
nvram erase
  • These are often confused!


/* boardflags */

  1. define BFL_BTCOEXIST 0x0001 /* This board implements Bluetooth coexistance */
  2. define BFL_PACTRL 0x0002 /* This board has gpio 9 controlling the PA */
  3. define BFL_AIRLINEMODE 0x0004 /* This board implements gpio13 radio disable indication */
  4. define BFL_ENETROBO 0x0010 /* This board has robo switch or core */
  5. define BFL_CCKHIPWR 0x0040 /* Can do high-power CCK transmission */
  6. define BFL_ENETADM 0x0080 /* This board has ADMtek switch */
  7. define BFL_ENETVLAN 0x0100 /* This board has vlan capability */
  8. define BFL_AFTERBURNER 0x0200 /* This board supports Afterburner mode */
  9. define BFL_NOPCI 0x0400 /* This board leaves PCI floating */
  10. define BFL_FEM 0x0800 /* This board supports the Front End Module */
  11. define BFL_EXTLNA 0x1000 /* This board has an external LNA */
  12. define BFL_HGPA 0x2000 /* This board has a high gain PA */
  13. define BFL_BTCMOD 0x4000 /* This board' BTCOEXIST is in the alternate gpios */


BCM3303 v0.7 clock frequencies: (todo...)

MIPS/CPU Backplane Comment

BCM3303 v0.8 clock frequencies:

MIPS/CPU Backplane Comment
183 91
188 94
197 98
200 100
206 103
212 106
216 108
217 109
225 113
238 119
240 120
250 125

Flow Acceleration, SFE and Cut-Through Forwarding

As Internet Bandwidth speeds available to the end users have continued to increase, manufacturers have had to increase speed and power of routers to keep up with the higher network speeds. This has 2 negative consequences, first is that powerful router CPU's at high clock speeds are much more expensive, and second that more powerful hardware increases heat production which shortens the life of the equipment. Adding a cooling fan can help this but it decreases router reliability because the fan can fail causing the router to overheat, and new SOHO router prices need to be maintained at approximately 25% of prices of a new Windows desktop otherwise the price/performance of using something like pfsense on an Intel-based PC, or the commercial Untangle on a custom PC is better than any hardware router. (although the reliability is much less and power consumption of the PC is much higher)

As a result research began on seeing if it would be possible to increase packet-handing efficiency from the standard netfilter-based NAT forwarding in Linux and thus maintain the same embedded CPU architecture while accelerating packet throughput.

The earliest acceleration effort appeared around 2010 as an experimental "binary blob" module in Broadcom's SDK named "fast NAT" and marked experimental. That was incorporated into TomatoUSB firmware Build 47 in June 2010. It was then disabled in TomatoUSB firmware in Build 48 the following month due to instability. This was the predecessor to the CTF binary blob module from Broadcom. It is purely software NAT acceleration and worked with MIPS processors.

The next acceleration effort was SFE. This was originally named "Fast Classifier" (Fast Crashifier by one wag) It was first posted to GitHub (2 years after the Qualcomm acquisition Atheros in 2011), by Qualcomm in 2013 Qualcomm Shortcut Forwarding Engine and per it's README:

"It was written as a demonstration of what can be done to provide high performance forwarding inside the kernel. There were two initial motivations:

1) To provide a platform to enable research into how QoS analysis systems can offload work and avoid huge Linux overheads.

2) To provide a tool to investigate the behavior of various processors, SoCs and software sets so that we can characterize and design new network processor SoCs."

For the next 4 years Broadcom continued to work on Fast NAT (CTF) with it making occasional appearances (disabled by default) in different forks of Tomato as bcm_nat, and steadily spreading through manufacturer's K26-based firmware for Broadcom based devices by Netgear, Linksys, and others. Most manufacturer's firmware DID NOT contain Linux kernels with many enhanced features, (such as packet inspection) thus Fast Nat was able to work with them, while the various open projects (Tomato, Openwrt, and all their forks, as well as dd-wrt) used more complete (and modern) Linux kernels so the Broadcom CTF module was unstable or impacted various services (like port forwarding)

There was also parallel work on Fast Classifier (SFE) being done, it was incorporated by Google into the ChromeOS 3.x source code ChromeOS 3.18 SFE source.

During this period a series of unfavorable comparisons in the media to manufacturer's firmware were made and posted, and users also pushed for CFE to be included (with often disappointing results when enabled) into their preferred open source router firmware. For example the ctf.ko module Broadcom released for the K2.6 kernels was tested back by kong around 2014 in dd-wrt and discovered to be very buggy so dd-wrt never implemented it.

Finally, SFE was ported into OpenWRT LEDE in June 2017 by gwlim Qualcomm Fast Path For LEDEand into dd-wrt by BrainSlayer a month later. Kong's SFE announcement This really helped the cause of high-speed link usage on open source firmware on routers as once the code was part of the source code of these flagship distributions, every other fork and project started incorporating SFE (and very likely a number of hardware manufacturers as well)

Broadcom had not been idle with CFE, either. By 2012 they had concentrated on the Northstar line of CPUs and in addition to the software NAT acceleration, introduced hardware acceleration in Northstar. CTF was handled by a proprietary software module ctf.ko to create Layer 1 acceleration, while Flow Acceleration (FA) was handled by adding special hardware into it's devices with gigabit Ethernet ports that helps distribute the processing of packets in the router to create Layer 2 acceleration. With Broadcom:

Level 1=CTF Only

Level 2=FA + CTF

With Layer 2, instead of the primary CPU handling all packet delivery within the router, many packets can be moved directly from interface to interface without the primary CPU being involved in all of the packet handling. FA was officially incorporated into dd-wrt in May of 2021. (after years of waiting for Broadcom to port it to the K4 kernels and release a binary) CTF works with many Broadcom chips while FA works with the Northstar line of Broadcom chips, and SFE works with Atheros, Broadcom, and all other chips (including Intel).

In summary, CTF and SFE are software optimization techniques to accelerate NAT, while FA is a Hardware NAT acceleration mechanism.

Once dd-wrt is loaded on your router, click Status and look at the router CPU. If it is a newer Broadcom CPU you should have CTF. If it is Broadcom and an ARM chip then you can check this list to see if your CPU is a Northstar CPU If it is, you will get both FA and CTF.

Note the following screenshot of a Linksys EA6200 with both options. CTF is accessible from the dropdown for Shortcut Forwarding Engine (SFE) while the Flow Acceleration is listed below it. If your CPU is not a Northstar CPU but does have CTF then you will have the dropdown for CTF in the Shortcut Forwarding Engine but not the Flow Acceleration selection. If it's not Broadcom you will only have SFE

Note that for Flow Acceleration to be visible and selectable, the WAN interface of the router MUST be configured EITHER with DHCP or with a Static IP address, it must also have QoS disabled, and AFTER the configuration is saved and applied, the router MUST be rebooted for the option to appear. CTF Level 1 does not require QoS to be disabled, however, so if you need QoS you can still use it.

Which scheme is faster

From a discussion on Build r46690

"ctf is much faster. with sfe i reach about 700 mbit with lan->wan tests. with ctf i get full gbit speed. so 950 - 980 mbit and lower cpu usage.... i'm running it here for 2 weeks now on several routers (r7000, ac66, n66)" --BrainSlayer

What kind of speed boost does this give me?

If you have a fairly fast router CPU such as 300Mhz or above, and a slower cable or DSL line, then you probably won't see any difference. If you have a 500Mbt or above line, you can see a big difference. For example without FA you might max out at 700Mbt on a gigabit Internet line.

For testing throughput, a baseline should be established with Shortcut Forwarding Engine (SFE) disabled and Flow Acceleration disabled. Then enable SFE and test, then switch to CTF and test, then switch back to SFE and enable FA, then finally switch to FA and CTF. Test results are going to be specific to the dd-wrt version and the router model.


  • While SFE is open source, CTF and FA are binary blob drivers from Broadcom. NOBODY other than Broadcom knows exactly how they work or what exactly they are doing, the links below are all speculations on what is actually going on. The fact that CTF requires Broadcom devices strongly hints that some part of the packet handling code in the module were coded in assembly language and that FA requires Northstar and ARM devices indicates specialty hardware is being used.
  • Packets that are modified by the router (such as packets being stuck into an OpenVPN tunnel) cannot go through SFE, CTF or FA and the router will continue to process them as before.
  • It is possible that some types of packets (such as VoIP packets) can be disrupted by having the CTF and FA options turned on. (SFE generally is not a problem)
  • Traffic Monitoring, (such as YAMon) PPPoE, and NAS fileserving don't work with these options.
  • Port forwards of some protocols may be interfered with.
  • Because the FA and CTF code is not published it is impossible for a 3rd party to audit it for security.
  • All of these apply to wired ethernet to ethernet connections. Of course, most wifi protocols do not operate on Gigabit speeds anyway. There are SOME of the higher end Broadcom wireless radios that have their own CPU's and will offload some packet processing from the main CPU but that is separate from FA

Broadcom docs:

Youtube video of Northstar Flow Accelerator Computex 2012 Demo This one is fascinating almost as much for the lack of identification of who is speaking, who he works for, what router model is sitting there taken apart, what router software is in use and what testing methodology is being used. The group filming the video implies it is a Broadcom representative.

It appears that Broadcom uses the same technology in their higher end hardware, under the names TruFlow and vSwitch acceleration.

Datasheet on the BCM957414A4142CC The BCM957414 is a Dual-Port 25 Gb/s SFP28 Ethernet PCI Express 3 x8 Network Interface Card that mentions the "TruFlow™ flow processing engine" and "vSwitch acceleration" features

Datasheet on the Stingray PS250 This sheet expands on TruFlow, stating:

"The powerful TruFlow™ configurable flow accelerator adds a powerful packet inspection and processing capability in the hardware, moving common flow-processing workloads into hardware and freeing CPUs for application workloads"

Datasheet for the bcm58712 CPU that states:

"Broadcom’s TruFlow™ packet flow processing technology enables Open vSwitch acceleration through CPU offload"

OpenWRT investigations on the older K26 CTF module

OpenWRT Technical Explanation of ctf.ko module and claims

Initial OpenWRT discovery and speculation on CTF

OpenWRT discussion of profiling the acceleration module

Hardware Modifications

Hardware Modifications (mods) are not necessarily DD-WRT specific or compatible.