[Cerowrt-devel] debloating the Pi, maybe not.

[Cerowrt-devel] debloating the Pi, maybe not.

[Dave Taht]

I was donated a raspberry pi in cambridge and got it booted today.
It's a very cool
box, an arm6, running debian, with hard floating point, and getting a complete
development environment running, getting it to do mesh networking,
graphics, apache, etc, on a 16GB SD card, was a breeze.

But as for network performance:

It only comes up at 100Mbit, and as noted elsewhere the ethernet chip
is hooked up over USB.

dmesg

[3.078644] smsc95xx 1-1.1:1.0: eth0: register 'smsc95xx' at
usb-bcm2708_usb-1.1, smsc95xx USB 2.0 Ethernet, b8:27:eb:18:fc:ac

The best the 3.2 kernel based networking can do is about 75Mbit/sec.
This eats 68% of cpu in the "system" (sy in top) and 30% in si (system
interrupt)

I had had some hope of this being a good box to put some test tools on,
but with performance like this, I am back to leveraging the mips stuff.

That said, I do intend to poke at the driver and update to 3.6 when
I get a chance, since we have patches lying around for both this
chip and the usbnet code.

It does indeed bloat up under load, but not badly.

64 bytes from 172.21.42.14: seq=49 ttl=64 time=0.625 ms

64 bytes from 172.21.42.14: seq=0 ttl=64 time=4.184 ms
64 bytes from 172.21.42.14: seq=1 ttl=64 time=3.421 ms
64 bytes from 172.21.42.14: seq=2 ttl=64 time=5.707 ms
64 bytes from 172.21.42.14: seq=3 ttl=64 time=2.843 ms
64 bytes from 172.21.42.14: seq=4 ttl=64 time=6.142 ms
64 bytes from 172.21.42.14: seq=5 ttl=64 time=6.026 ms
64 bytes from 172.21.42.14: seq=6 ttl=64 time=2.373 ms
64 bytes from 172.21.42.14: seq=7 ttl=64 time=13.746 ms
64 bytes from 172.21.42.14: seq=8 ttl=64 time=3.296 ms
64 bytes from 172.21.42.14: seq=9 ttl=64 time=0.699 ms
64 bytes from 172.21.42.14: seq=10 ttl=64 time=2.948 ms
64 bytes from 172.21.42.14: seq=11 ttl=64 time=12.975 ms
64 bytes from 172.21.42.14: seq=12 ttl=64 time=4.238 ms


-- 
Dave Täht

Fixing bufferbloat with cerowrt: http://www.teklibre.com/cerowrt/subscribe.html

Re: [Cerowrt-devel] debloating the Pi, maybe not.

[Oliver Neukum]

On Friday 23 November 2012 17:43:24 Dave Taht wrote:
> I was donated a raspberry pi in cambridge and got it booted today.
> It's a very cool
> box, an arm6, running debian, with hard floating point, and getting a complete
> development environment running, getting it to do mesh networking,
> graphics, apache, etc, on a 16GB SD card, was a breeze.
> 
> But as for network performance:
> 
> It only comes up at 100Mbit, and as noted elsewhere the ethernet chip
> is hooked up over USB.
> 
> dmesg
> 
> [3.078644] smsc95xx 1-1.1:1.0: eth0: register 'smsc95xx' at
> usb-bcm2708_usb-1.1, smsc95xx USB 2.0 Ethernet, b8:27:eb:18:fc:ac
> 
> The best the 3.2 kernel based networking can do is about 75Mbit/sec.
> This eats 68% of cpu in the "system" (sy in top) and 30% in si (system
> interrupt)

Wow. What did you do to saturate the system like this? Blow empty
UDP packages into the net?

> I had had some hope of this being a good box to put some test tools on,
> but with performance like this, I am back to leveraging the mips stuff.
> 
> That said, I do intend to poke at the driver and update to 3.6 when
> I get a chance, since we have patches lying around for both this
> chip and the usbnet code.

Nothing for performance in the mean time. Note that we don't try to
do interrupt mitigation in usbnet.

	Regards
		Oliver

Re: [Cerowrt-devel] debloating the Pi, maybe not.

[Dave Taht]

On Mon, Nov 26, 2012 at 4:10 PM, Oliver Neukum <oneukum@suse.de> wrote:
> On Friday 23 November 2012 17:43:24 Dave Taht wrote:
>> I was donated a raspberry pi in cambridge and got it booted today.
>> It's a very cool
>> box, an arm6, running debian, with hard floating point, and getting a complete
>> development environment running, getting it to do mesh networking,
>> graphics, apache, etc, on a 16GB SD card, was a breeze.
>>
>> But as for network performance:
>>
>> It only comes up at 100Mbit, and as noted elsewhere the ethernet chip
>> is hooked up over USB.
>>
>> dmesg
>>
>> [3.078644] smsc95xx 1-1.1:1.0: eth0: register 'smsc95xx' at
>> usb-bcm2708_usb-1.1, smsc95xx USB 2.0 Ethernet, b8:27:eb:18:fc:ac
>>
>> The best the 3.2 kernel based networking can do is about 75Mbit/sec.
>> This eats 68% of cpu in the "system" (sy in top) and 30% in si (system
>> interrupt)
>
> Wow. What did you do to saturate the system like this? Blow empty
> UDP packages into the net?

A single TCP stream.

I built netperf 2.6 with --enable-demo, ran netserver on it, and hit it
with single

netperf -l60 -H pi

I don't remember if the performance I got from it was with -t
TCP_STREAM or -t TCP_MAERTS (probably the latter).

I will re-run the test with iperf, also, the next time I boot it.

Other than this it's a pretty darn cool board, and if I had time I'd
dream up a zillion projects for it... It does look rather easy to
compile or cross compile kernels for it, too...

>
>> I had had some hope of this being a good box to put some test tools on,
>> but with performance like this, I am back to leveraging the mips stuff.
>>
>> That said, I do intend to poke at the driver and update to 3.6 when
>> I get a chance, since we have patches lying around for both this
>> chip and the usbnet code.
>
> Nothing for performance in the mean time. Note that we don't try to
> do interrupt mitigation in usbnet.

ok.

>         Regards
>                 Oliver
>



-- 
Dave Täht

Fixing bufferbloat with cerowrt: http://www.teklibre.com/cerowrt/subscribe.html