[Cerowrt-devel] router chipset selection for cero-2

Mon Aug 13 13:57:52 EDT 2012

When I said "not cheaper" I didn't say that "price was no object!!" :)

Both of the x86 systems proposed would cost well north of 400 dollars
when outfitted properly. This is a lot to request of the volunteer
cerowrt effort to buy on their own! (I note that donations of
sufficient hardware of any stripe would be gladly considered)

I still retain hope we'll be able to get a distributed network
analysis system deployed using the existing cerowrt hardware, too, and
at any scale,
that's going to cost a lot as it is.

The present retail cost on a 3700v2 is under 100 dollars, and from
that we can assume that the build cost is under 30. I was figuring on
(at worst) doubling the build cost for cero-2. By looking at various
new chipsets now, that may only be available in evaluation board form
(now), my hope would be to influence the development of future
products *before* they got deployed, this time around.

This is a more ambitious goal than the cero-1 effort. There we wanted
to find a popular 100% open source chipset to work with, that was
widely sold at retail, work with it, and deeply understand what was
right and wrong with the software load...

which we're kind of done with. Peering dimly 18 months into the future
as to what would become a popular chipset for home routers is hard,
but I don't think it will be x86 based.

I HAVE been tempted towards explicitly adding support for this x86
adsl capable box and exploring how far up it can scale:

http://www.traverse.com.au/geos21-dual-adsl2-x86-router-appliance

David Woodhouse has one and I've been meaning to hack on his for ages.
Getting a good grip on adsl behavior would be nice to have.

To add to the negatives of the x86 idea some more, however:

Working on a resource constrained platform with an embedded mindset is
very useful, in that memory intensive design mistakes (like, for
example, doing the UI in java, leveraging bloated libraries) aren't
made. Proving that fq_codel would 'just work' on commonly deployed
low-end hardware was very important to us to prove to the industry as
a whole.  While some problems remain, as codel can be more memory
intensive than we'd like, I like to think the myth that fq and aqm
can't run on all interfaces of a given router is now *thoroughly*
debunked. This stuff barely registers on a cpu trace on the wndrs.

Similarly, manufacturers care about pennies in their markets and
shipping the bare minimum amount of hardware required to meet a market
niche is what they are all about.

Next:

If we'd produce an x86 version of cero, everyone on the planet would
try tossing their own junked 486 through 6 core  box at the problem
with wildly varying results. The additional driver support required in
the x86 world boggles the mind, just on ethernet alone.

The support burden for the cerowrt effort is already beyond sustainable.

I'm perfectly happy for x86 folk to track ubuntu/redhat/suse/etc
directly, and to try and push bufferbloat related patches into kernels
relevant to those distros, and more importantly, the people that are
working on those distros (this has been done for fedora and ubuntu -
thx John and Kamal!). The openwrt mainline works on over 30 arches
including x86, too. (thx everybody at openwrt!). Still, in the x86
world, few seem to have picked up on how needed BQL support is needed
at the driver level to see any useful results from an aqm, and
furthermore everybody over there is busily adding all sorts of
cpu-saving network (tso,gso,ufo,gro,etc) offloads, which nearly
universally add latency.

now, I'm capable of arguing both sides of an issue (lawyer for a dad)
in the same email...

so, in defense of using an x86 platform for (at least some) future
cerowrt related development:

A) it's EASY to track kernel head and do experiments and new code development.
The compile, test, debug cycle is much shorter,
and getting stuff pushed upstream is much easier.

This translates out to much shorter time to market than any other cpu
arch (although arm is tracking mainline linux quite well now)

B) I already do do x86 based work a goodly percentage of the time (on
a pair of laptops, one of which just died). For most of the last
quarter of last year I worked on x86 exclusively (for bql and sfqred).
Judging from the upcoming (and sadly unfunded) plans for wifi and
improvements to fq_codel, a lot more work should be done initially on
the x86.

But in either of those cases using off the shelf used x86 hardware
seems more cost effective than picking a new x86 based routing box
from a given vendor, and going with it, unless said vendor was
supporting the effort directly.

It does bug me that although we get great support from various
individuals related to atheros, nobody at netgear calls us back. We
picked the netgear because it had a great chipset, and the build
quality is excellent (tested to -40C/+70C). Numerous other users of
the same chipset were in the running (notably buffalo), but the build
quality wasn't as good...

I wouldn't mind working with the next generation (mips 75k) atheros
chipsets with  a manufacturer that cared...

And certainly on my over-long todo list is to get someone to add BQL
and fq_codel support to the dreamplug and smileplug arm based boxen...

So anyway, this gets me back to my original questions about the
comcerto 2000, and I guess I'm going to have to make some phone calls.

On Mon, Aug 13, 2012 at 7:59 AM, Luke Hamburg <luke at solvent-llc.com> wrote:
> How about something like the Jetway JBC372F36 ?
>
> -Atom N2600 (new) CedarView CPU: fairly capable + very low wattage (fanless)
> -mini-PCIe card slot (comes populated with a B/G/N-card but that is easily
> replaced)
> -dual gigabit Ethernet ports, USB ports, serial/COM ports
> -rugged industrial design (all metal enclosure) - with (4) antenna holes!
>
> The Comcerto C2200 sure does have some wonderful-sounding specs - but for
> now I couldn't find any available (or even planned) hardware. I realize this
> Atom unit doesn't satisfy all of the requirements (no hardware RNG, not sure
> about hooking to advanced interfaces e.g. GPON) but surely this beats a
> wndr3700 ?
>
> Luke
>
>
> On Sat, Aug 11, 2012 at 9:44 PM, Dave Taht <dave.taht at gmail.com> wrote:
>>
>> I can't remember with whom I was talking to about alternatives to mips
>> for home router processors, but this is the first new one I've seen in
>> the arm world that comes close to being one...
>>
>>
>> http://electronicdesign.com/article/digital/dualcore-cortexa9-tackles-communication-gateway-chores-74092
>>
>> Most of the new arms are targetted at the burgeoning handheld markets.
>> A router has no use for video
>> and a big use for pci busses and multiple ethernet chips, which the
>> handheld targetted chips usually don't have...
>>
>> All that said, the arm ecosystem appears to be healthier than the mips
>> ecosystem, overall.
>>
>> The marvell kirkwood (dreamplug) is getting long in the tooth, the
>> octeon is too expensive (and the cool onboard hardware locked away
>> with binary blobs), the various atheros chipsets I'm aware of a little
>> too weak, the broadcoms a little too proprietary, and perhaps this new
>> chip from mindspeed would be "just right".
>>
>> So... Anyone know about the comcerto 2000, or of anything else out there?
>>
>> I figure cerowrt's ar71xx chipset currently has less than 18 months of
>> market life left to it.
>>
>> While I would hope to have finished fixing the home router market by
>> then, deploying fq_codel, getting ipv6 made default, solving the
>> naming problems, deploying dnssec, etc and genericall  fixing all the
>> head-ends and the rest of the known internet universe, etc, I'm not
>> planning on it. :)
>>
>> So thinking about what next piece of open, and more powerful hardware
>> to complete the research/work with is starting to weigh on my mind.
>>
>> Big Goals:
>>
>> Faster, better, not cheaper
>> Open source in all core components
>> 802.11ac
>> Hardware rng
>> Capable of gigE ipv6 *routing*
>> Capable of being hooked to various advanced interfaces (gpon, cable, adsl)
>>
>>
>> --
>> Dave Täht
>> http://www.bufferbloat.net/projects/cerowrt/wiki - "3.3.8-6 is out
>> with fq_codel!"
>> _______________________________________________
>> Cerowrt-devel mailing list
>> Cerowrt-devel at lists.bufferbloat.net
>> https://lists.bufferbloat.net/listinfo/cerowrt-devel
>
>
>
> _______________________________________________
> Cerowrt-devel mailing list
> Cerowrt-devel at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/cerowrt-devel
>

-- 
Dave Täht
http://www.bufferbloat.net/projects/cerowrt/wiki - "3.3.8-6 is out
with fq_codel!"