[Make-wifi-fast] Testing the LEDE nbd/staging branch

[Make-wifi-fast] Testing the LEDE nbd/staging branch

[Richard Smith]

I noticed a need-more-testing post on the CeroWrt list and finally got 
around to building a nbd/staging+airtime fariness wndr3700v2 image for 
one of my spare units.

I just subscribed to the list and looks like there are some followup 
patches not in my build so I'll have to re-test with the followups next.

One question:  Are these patches "Just Work" or do I need to do some 
tweaking of qdiscs, etc like the sqm or cake stuff needs?  If so what 
are those?

I did some testing this morning and I seem to have an additional problem 
than just the slowdown.

When I run flent my against the ndb/stageing+patches the wifi on my 
laptop pukes after a few seconds and I have to disable/re-enable to get 
it back. Doesn't happen running the same test against a wdnr3800 running 
OpenWrt CC (my main AP).

Here's a snipit of the kernel error.

[301744.796063] iwlwifi 0000:3a:00.0: Microcode SW error detected. 
Restarting 0x82000000.
[301744.796074] iwlwifi 0000:3a:00.0: CSR values:
[301744.796079] iwlwifi 0000:3a:00.0: (2nd byte of CSR_INT_COALESCING is 
CSR_INT_PERIODIC_REG)
[301744.796087] iwlwifi 0000:3a:00.0:        CSR_HW_IF_CONFIG_REG: 
0X00489008
[301744.796094] iwlwifi 0000:3a:00.0:          CSR_INT_COALESCING: 
0X0000ff40
[301744.796101] iwlwifi 0000:3a:00.0:                     CSR_INT: 
0X80000000
[301744.796108] iwlwifi 0000:3a:00.0:                CSR_INT_MASK: 
0X00000000
[301744.796115] iwlwifi 0000:3a:00.0:           CSR_FH_INT_STATUS: 
0X00010000
[301744.796122] iwlwifi 0000:3a:00.0:                 CSR_GPIO_IN: 
0X00000019

-- 
Richard A. Smith

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Toke Høiland-Jørgensen]

Richard Smith <smithbone@gmail.com> writes:

> I noticed a need-more-testing post on the CeroWrt list and finally got
> around to building a nbd/staging+airtime fariness wndr3700v2 image for
> one of my spare units.
>
> I just subscribed to the list and looks like there are some followup
> patches not in my build so I'll have to re-test with the followups
> next.
>
> One question: Are these patches "Just Work" or do I need to do some
> tweaking of qdiscs, etc like the sqm or cake stuff needs? If so what
> are those?

They are "just work", but there are some issues that I am currently
trying to work out. If you just run off Felix' staging tree you are
fine, but won't get per-flow FQ to each station.

You can remove package/kernel/mac80211/patches/220-fq_disable_hack.patch
which will get you the per-station FQ, but will break crypto. Should be
fine for testing unencrypted stuff though.

-Toke

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Dave Taht]

I guess I'm confused - does felix's repo also have your airtime
fairness scheduler?

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Richard Smith]

On 08/29/2016 04:01 PM, Dave Taht wrote:
> I guess I'm confused - does felix's repo also have your airtime
> fairness scheduler?

I may have confused you.  I added the patches from your "wifi airtime 
fairness patches could use eyeballs and testing" post to the nbd/staging 
branch.

The email (indirectly) suggested that was the branch to use.

-- 
Richard A. Smith

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Toke Høiland-Jørgensen]

Richard Smith <smithbone@gmail.com> writes:

> On 08/29/2016 04:01 PM, Dave Taht wrote:
>> I guess I'm confused - does felix's repo also have your airtime
>> fairness scheduler?
>
> I may have confused you.  I added the patches from your "wifi airtime fairness
> patches could use eyeballs and testing" post to the nbd/staging branch.
>
> The email (indirectly) suggested that was the branch to use.

Yup, that's the right thing to do.

I think I managed to fix the remaining crypto bug. Will send an updated
patch tomorrow (need to extract it from my pile of test versions). In
the meantime, here are some images with the patch applied if someone
wants to go ahead and test before then:

https://kau.toke.dk/experiments/softq-debug/

-Toke

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Dave Taht]

I have tested toke's wndr3800 build below, and with psk2+ccmp crypto
enabled, using an osx box
to drive the tests, I can get 210Mbit wifi downloads and uploads for
nearly any number of flows. Fairness is nearly perfect. w00t!

Regrettably my laptop is still running the broken code (and toke's got
patches to send), so I can't test the linux wifi on that laptop.

I(or someone with a non-borken laptop running) need to test tkip and
other forms of aes... see table here for how to configure

https://wiki.openwrt.org/doc/uci/wireless#wpa_modes

One really impressive thing, during a 24 flow download test, was how
good packet aggregation got.

Total packet count::    ideal 7702277      lookaround 4715
Average # of aggregated frames per A-MPDU: 31.8

I have never, driving tests with tcp, seen it get that good (nor have
I tested via an osx box much)

The script that I'm basically using, if you'd like to replicate my
results with your hardware/conditions with various forms of crypto on
(and please compare against off!).

#!/bin/sh

flent -t "psk2-cc-ht40-osx-ether" -H 192.168.1.201 -l 30 rrul
flent -t "psk2-cc-ht40-osx-ether" -H 192.168.1.201 -l 30 rrul_be

for i in 1 2 4 8 12 16 24;
do
flent -t "psk2-cc-ht40-$i-flows-osx-ether" -H 192.168.1.201 -l 30
--test-parameter=upload_streams=$i tcp_nup
flent -t "psk2-cc-ht40-$i-flows-osx-ether" -H 192.168.1.201 -l 30
--test-parameter=download_streams=$i tcp_ndown
done




On Mon, Aug 29, 2016 at 2:09 PM, Toke Høiland-Jørgensen <toke@toke.dk> wrote:
> Richard Smith <smithbone@gmail.com> writes:
>
>> On 08/29/2016 04:01 PM, Dave Taht wrote:
>>> I guess I'm confused - does felix's repo also have your airtime
>>> fairness scheduler?
>>
>> I may have confused you.  I added the patches from your "wifi airtime fairness
>> patches could use eyeballs and testing" post to the nbd/staging branch.
>>
>> The email (indirectly) suggested that was the branch to use.
>
> Yup, that's the right thing to do.
>
> I think I managed to fix the remaining crypto bug. Will send an updated
> patch tomorrow (need to extract it from my pile of test versions). In
> the meantime, here are some images with the patch applied if someone
> wants to go ahead and test before then:
>
> https://kau.toke.dk/experiments/softq-debug/
>
> -Toke



-- 
Dave Täht
Let's go make home routers and wifi faster! With better software!
http://blog.cerowrt.org

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Richard Smith]

On 08/29/2016 06:22 PM, Dave Taht wrote:
> I have tested toke's wndr3800 build below, and with psk2+ccmp crypto
> enabled, using an osx box
> to drive the tests, I can get 210Mbit wifi downloads and uploads for
> nearly any number of flows. Fairness is nearly perfect. w00t!

Unfortunately,  I still get the same behavior where it borks my Intel AC 
8260 card.

-- 
Richard A. Smith

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Toke Høiland-Jørgensen]

Richard Smith <smithbone@gmail.com> writes:

> On 08/29/2016 06:22 PM, Dave Taht wrote:
>> I have tested toke's wndr3800 build below, and with psk2+ccmp crypto
>> enabled, using an osx box
>> to drive the tests, I can get 210Mbit wifi downloads and uploads for
>> nearly any number of flows. Fairness is nearly perfect. w00t!
>
> Unfortunately, I still get the same behavior where it borks my Intel
> AC 8260 card.

Hmm, that's certainly "interesting"...

You say your baseline is a CC access point; there are so many moving
parts in this that it would be good to verify whether the crash is
caused by the softq patches specifically. Could you do an identical
build where you remove
package/kernel/mac80211/patches/337-ath9k-Switch-to-using-mac80211-intermediate-software.patch
(and any subsequent patches that touch the same code; i.e., the airtime
fairness patches), and see if that also causes crashes? :)

-Toke

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Richard Smith]

On 08/30/2016 06:48 AM, Toke Høiland-Jørgensen wrote:

> Richard Smith <smithbone@gmail.com> writes:
>
>> On 08/29/2016 06:22 PM, Dave Taht wrote:
>>> I have tested toke's wndr3800 build below, and with psk2+ccmp crypto
>>> enabled, using an osx box
>>> to drive the tests, I can get 210Mbit wifi downloads and uploads for
>>> nearly any number of flows. Fairness is nearly perfect. w00t!
>>
>> Unfortunately, I still get the same behavior where it borks my Intel
>> AC 8260 card.
>
> Hmm, that's certainly "interesting"...
>
> You say your baseline is a CC access point; there are so many moving
> parts in this that it would be good to verify whether the crash is
> caused by the softq patches specifically. Could you do an identical
> build where you remove
> package/kernel/mac80211/patches/337-ath9k-Switch-to-using-mac80211-intermediate-software.patch
> (and any subsequent patches that touch the same code; i.e., the airtime
> fairness patches), and see if that also causes crashes? :)

Sure and if that works I'll walk though the patches seeing which one 
causes the trouble.

I only have before/after work to test so it may take a day or 2 to go 
though all the combinations.

-- 
Richard A. Smith

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Toke Høiland-Jørgensen]

Richard Smith <smithbone@gmail.com> writes:

> On 08/30/2016 06:48 AM, Toke Høiland-Jørgensen wrote:
>
>> Richard Smith <smithbone@gmail.com> writes:
>>
>>> On 08/29/2016 06:22 PM, Dave Taht wrote:
>>>> I have tested toke's wndr3800 build below, and with psk2+ccmp crypto
>>>> enabled, using an osx box
>>>> to drive the tests, I can get 210Mbit wifi downloads and uploads for
>>>> nearly any number of flows. Fairness is nearly perfect. w00t!
>>>
>>> Unfortunately, I still get the same behavior where it borks my Intel
>>> AC 8260 card.
>>
>> Hmm, that's certainly "interesting"...
>>
>> You say your baseline is a CC access point; there are so many moving
>> parts in this that it would be good to verify whether the crash is
>> caused by the softq patches specifically. Could you do an identical
>> build where you remove
>> package/kernel/mac80211/patches/337-ath9k-Switch-to-using-mac80211-intermediate-software.patch
>> (and any subsequent patches that touch the same code; i.e., the airtime
>> fairness patches), and see if that also causes crashes? :)
>
> Sure and if that works I'll walk though the patches seeing which one
> causes the trouble.

Awesome, thanks!

> I only have before/after work to test so it may take a day or 2 to go
> though all the combinations.

No worries. Just holler when you know more :)

-Toke

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Richard Smith]

On 08/30/2016 01:56 PM, Toke Høiland-Jørgensen wrote:

>
>> I only have before/after work to test so it may take a day or 2 to go
>> though all the combinations.
>
> No worries. Just holler when you know more :)

Seems it's actually not specific to the staging branch and apparently 
I'm not good at remembering what my baseline AP is. :)

Baseline is running dd r40727 and not cc like I thought.

Anyway it happens on both staging and baseline but just seems to be much 
less frequent on r40727 than on the staging build.  Small sample size 
bias I think.

I'll try out the AP's at work tomorrow and see if it happens there as well.

-- 
Richard A. Smith

Re: [Make-wifi-fast] Testing the LEDE nbd/staging branch

[Dave Taht]

On Tue, Aug 30, 2016 at 6:07 PM, Richard Smith <smithbone@gmail.com> wrote:
> On 08/30/2016 01:56 PM, Toke Høiland-Jørgensen wrote:
>
>>
>>> I only have before/after work to test so it may take a day or 2 to go
>>> though all the combinations.
>>
>>
>> No worries. Just holler when you know more :)
>
>
> Seems it's actually not specific to the staging branch and apparently I'm
> not good at remembering what my baseline AP is. :)
>
> Baseline is running dd r40727 and not cc like I thought.
>
> Anyway it happens on both staging and baseline but just seems to be much
> less frequent on r40727 than on the staging build.  Small sample size bias I
> think.
>
> I'll try out the AP's at work tomorrow and see if it happens there as well.
>
> --
> Richard A. Smith

Yay! If you have some other client chipset to test with? I'd like to
see iwl in particular put through a wringer....

-- 
Dave Täht
Let's go make home routers and wifi faster! With better software!
http://blog.cerowrt.org