[Cerowrt-devel] [Bloat] still trying to find hardware for the next generation worth hacking on

[Dave Taht]

On Mon, Aug 18, 2014 at 3:51 PM, Frank Carmickle <frank at carmickle.com> wrote:
>
> On Aug 17, 2014, at 1:13 PM, dpreed at reed.com wrote:
>
> http://www.habeyusa.com/products/fwmb-7950-rangeley-network-communication-board/
> looks intriguing.
>
>
> This, and others like it from Supermicro and Asrock, are what I am
> interested in.  I'm not so sure this is what Dave is looking for.  These are
> not sub $100 boards by any stretch.  I happen to like IPMI serial over LAN
> and don't really want to live without it.  I built a Habey fanless chassis
> and a Asrock board system with 16gb of RAm a few months ago.  The Habey
> fanless chassis is pretty nice except that it smelled awful.  They have an
> external inline power brick with a barrel connector that is converted to atx
> by a small board in the chassis.  This way I also get to run Debian without
> worrying about filling up some small flash.
>
> Dave, are you looking for something for the masses?

No, I'm looking for something to do further, open, effective, bleeding
edge and game changing research on... a project that will run for
another 3+ years...

...that eventually will be something for the masses.

Cost is not a huge object right now. Cheap in 3 years would be a good goal.

On the high end, for example, are things like netfpga, and lower cost
options like the parallella architecture.

http://netfpga.org/
http://www.parallella.org/

It does seem like better hardware offloads need to be developed to ensure that
rate scheduling, AQM, and fair queuing, can move into new hardware.

I agree with other folk on this thread that we are in a transition
period where the cost-effective chips have offloads that interfere
with further bufferbloat related research, even on the technologies
that are relatively open (ethernet, wifi).

This situation is far worse on access technologies such as 3g, and
lte, cable, and gpon, which are black boxes out of reach of open and
academic researchers.

About the only FTTH technology that is relatively accessible is single
mode fiber ethernet, and it appears to be uncommon except in
municipalities that have mandated ISP competition (for example:
http://arstechnica.com/tech-policy/2010/03/how-amsterdam-was-wired-for-open-access-fiber/
)

In the cases where we can make a difference, the only cpu technologies
that appear
capable of doing GigE and software rate limiting and codel seem to be
x86 based. (And only intel! even a amd box, recently benchmarked,
couldn't forward at > 700Mbit)

I even have doubts about the atom box mentioned earlier, the available
cache sizes are quite small, and ivy bridge based platforms seem
better, but only testing could tell.

Although I like the turrus project, I too am dubious about the future
for power pc in these marketplace.

I've had a chat with atheros about some of their new stuff, I'm not
sure what of that I can share here (yet). Will let you all know when I
can.

I think all options are open. I'd like to have a project that could
indeed be funded (for a change!) - AND make a difference - and be fun
and useful for everyone here and a variety of other research-y
projects like project bismark, the EFF work, commotion wireless, etc,
etc. An ideal project would be to produce an entirely open high speed
router that anyone could make.

Shorter term... it would be nice to find a board that "just worked"
well with the SQM system, up to 300mbit, however. That covers the
largest band of ISP deployments for the next several years.

And I, at least, have competing goals - there are the gateway problem,
and the DSLAM/CMTS/GPON head end problem... and then there's fixing
wifi, which is nearest and dearest to my heart.

At the moment I'm planning on switching to some x86 box to prototype
solutions for make-wifi-fast, (due to ease of debugging mostly,
availability of mini-pcie, also). 802.11ac is presently a mess...

So I appreciate the suggestions so far, keep them coming! All options
are on the table and what we'll congeal on or not congeal on is wildly
variable, but I would like to find a way focus enough folk on solving
any one problem to get decent results inside of a year.

Note: I am at sigcomm and linuxcon this week in chicago and my replies
will be sparse.

My morning's keynote at the congestion control workshop slides are up at:

http://snapon.lab.bufferbloat.net/~d/sigcomm2014.pdf

(Sadly, not recorded)




>
> --FC
>
>
>
> _______________________________________________
> Bloat mailing list
> Bloat at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/bloat
>



-- 
Dave Täht

NSFW: https://w2.eff.org/Censorship/Internet_censorship_bills/russell_0296_indecent.article