[Cerowrt-devel] packet capture hardware

Mon Feb 4 12:48:21 EST 2013

Changing the subject line to reflect this line of discourse.

On Mon, Feb 4, 2013 at 8:41 AM, <dpreed at reed.com> wrote:

> I hadn't researched the HPC FMC requirement for 10 GigE one yet.
>
>
>
> The 1 GigE one is expensive, but not because of parts cost.  This is the
> usual huge markup that goes with stuff sold to "Design Engineers" in
> companies - because they can charge, they do.
>

Well, it is also a function of volume. as a counter example, we can
probably leverage an upcoming manufacturing run of one of atheros's newer
chipsets, designed close to a cerowrt-able, debloatable spec, for about 30
bucks in 10k qtys. This still sort of implies a change in cerowrt's focus
from "fixing hardware you can get off the shelf" to *making something*
arduino-raspberri pi like, but has a great deal of appeal for me.
(inspiration: meraki) I am sufficiently annoyed at the entire industry at
this point. I am insufficiently wealthy.

Anyway, that chipset probably isn't fast enough to do packet captures at
line rate, so to continue on the thread of "designing a good box for packet
captures" but sort of half retaining the cerowrt concept and wandering
around others, in this email....

I think there is a real market need for something in the SFP form factor
that can do high rate packet captures and other sorts of analysis. I
imagine a SFP in, and Esata out going into a router would be a useful
diagnostic tool (and also something the NSA would love, which I have
ambiguous feelings about)

It could also be priced appropriately and maybe make some money.

I think there is also a market need for something that can be an analysis
box/home router that can also do captures at typical rates in the home
(20-30Mbit), but that's still just above what a wndr3800 can do when last I
tried. (it's mostly bound by the usb interface actually)

The dreamplug hw can do that, as best as I recall (getting one shortly)

>
> The zedboard PMOD interface seems to be more marketing appropriate for
> "cheap" stuff.  There is a PMOD for 100baseT, so you could throw a few of
> those on your system very cheaply.   Since the interface to PMODs is 8-bit
> parallel, all you might need is the magnetics and PHY for GigE, and you
> could make a soft GigE controller in the programmable logic part of the
> Zynq-7020.
>

I'd certainly like to make an eth controller capable of handling TSO/UFO
and breaking them up with fq/codel at the lowest possible level. On the
other hand I'm pretty sure a dual core a9 box is fast enough to drive gigE
with minimal buffering (but haven't played with the zedboard enough to
know. I do know the driver isn't bql'd. It's on my todo list)

One of the things I'm vague about is the path to making silicon, starting
with a FPGA design like this. Say we solve the universe:

* Build a better wifi interface (and other forms of wireless interface)
* Do gigE switching/routing/rate limiting with fq/codel in hw
* Has adsl and/or cable modem functionality
* Earthquake detector (just throwing that in there! :) )

What's the path to cost reducing that to, say, 15 bucks a chip in 3 years?

> I'd have to check that the signalling rates would be sustainable across
> the PMOD connector.
>

100Mbit is enough for the "home gateway" scenario.

>
>
> To make an FMC board, populate it with whatever GigE chip you like, etc.
> is trivial.  It should cost no more to fabricate than one of these little
> single chip GigE PCIe cards you can buy.   What chip would you like to
> use?   I (or others) could design the board and BOM, kit it up for
> manufacturing (by, say, Sunstone or other places that do PC boards and
> kitted assembly in small runs).
>

I like the idea of a soft chip on the fpga myself, actually. I'd like to
get smarter logic inside the tx ring. I don't care for any of the current
generation of ethernet chips very much. The ar71xx in cero has the
advantage of being rather simple, the e1000e is a very common chip, too.
The realtek is terrible with tons of errata.

So to just use a phy... well, broadcom's common phys need a nda to look at,
so do marvel's. It would be interesting to pursue making a switch/router
actually out of a sufficient number of phys, if there is sufficient I/Os
available on the fpga. Something like the vyatta...

and with a soft eth design it could scale up to 10GigE or higher.

>
> Trivial stuff - maybe one could even convince Digilent and/or Avnet to do
> the design/mfring.
>

I would like to think that the latency advantage of making a debloated box
would convince some people, like wall street, and large scale buyers to get
involved. That said, I look at the hits on things like the water videos at
modena and the uphill battle with multiple manufacturers thus far and get
discouraged...

>
>
> Wouldn't it be a lot better to have a pluggable and completely flexible
> highly scalable monitoring unit that could go down the wire level as
> needed, with the base cost being the $300 that a Zedboard goes from?
>

It looks like the fpga chip itself is 220 presently. I am not sure how
rapidly that will drop with time or volume.

ooh, I see they have a milspec version (my hobby is space stuff)

>
>
> And it would be completely "open hardware" and :"open source".
>

I would so totally dig that. The number of VCs in my rolodex is rather
small.

I agree with you that the zedboard is "the raspberri pi of high speed
digital logic" and that a zillion things can/will be done with it. However
it's at a painful price point presently for most "normal" people. This is
an advantage, actually, given some of the target markets...

(I kind of hate it when I wear my business hat rather than my engineering
one)

I think the scope of designing a full fledged standalone zedboard-like
board,
one that fits into the home router role, or a packet capture role, or a SFP
slot,
is rather large, and would need a payoff at the end...

Even something on the scale of the netfpga project over at stanford (which
only saw about 2000 manufactured and huge uni support), will take time and
money. It would be very fun, and potentially profitable at the end, but as
a hobby project... the learning curve is steep, the skills required very
diverse. (yes, fun, yes needs a community to form around it)

(And cero as it stands eats way too much of my time and I really would like
to get someone else(s) building it so I can focus on more nagging issues up
the stack)

As for designing an add-on 100Mbit board to the zedboard, much easier. I'm
not huge on the PMOD connectors (fragile. Worse, the SD card sticks out the
side, and I already broke one zedboard's SD connector off), and a big
unknown is how fast they can be driven....

>
> -----Original Message-----
> From: "Dave Taht" <dave.taht at gmail.com>
> Sent: Sunday, February 3, 2013 8:47pm
> To: dpreed at reed.com
> Cc: "Mark Constable" <markc at renta.net>,
> cerowrt-devel at lists.bufferbloat.net
> Subject: Re: [Cerowrt-devel] stanford talk/deluged in hardware/yurtlab
>
>  Darn I wish I'd made it to that show today.
>
> On Sun, Feb 3, 2013 at 5:11 PM, <dpreed at reed.com> wrote:
>
>> http://www.prweb.com/releases/2012/2/prweb9154394.htm (10 GigE FMC card)
>>
>>
>>
>
> impressive. Seems to require a hpc (high pin count) board, which zed isn't.
>
>
>>
>> http://www.xilinx.com/products/boards-and-kits/1-2AJPAV.htm (1 GiGE FMC
>> card)
>>
>
> 625 eu. While I am painfully aware of how much it costs to step ahead of
> the bleeding edge, I think the odds are pointing harder and harder at doing
> a non-fpga design that does what I want...
>
> I may go back to looking at octeons or ti's new octeon killer.
>
> And/or leveraging a newer atheros reference board.
>
>
>>
>>
>> -----Original Message-----
>> From: "Dave Taht" <dave.taht at gmail.com>
>>  Sent: Sunday, February 3, 2013 1:39pm
>> To: dpreed at reed.com
>> Cc: "Mark Constable" <markc at renta.net>,
>> cerowrt-devel at lists.bufferbloat.net
>> Subject: Re: [Cerowrt-devel] stanford talk/deluged in hardware/yurtlab
>>
>>
>>
>> On Sun, Feb 3, 2013 at 10:26 AM, <dpreed at reed.com> wrote:
>>
>>> It would be trivial to do this with a Zedboard.
>>>
>>
>> Well, need two network ports. Haven't figured out much on interfacing the
>> thing to offboard gear (I'd have liked it if it had a pci interface). So is
>> interfacing up a second network card "trivial" on the I/Os provided?
>>
>> And wanted esata, or some high speed disk I/O interface for captures.
>>
>> I'd rather like to continue forward on the zedboard front. The prospect
>> of designing an ethernet chip that actually could incorporate fq_codel etc
>> is very exciting. The RGII interface is available to access directly, in
>> particular.
>>
>>
>>
>>
>>
>>>
>>> -----Original Message-----
>>> From: "Dave Taht" <dave.taht at gmail.com>
>>> Sent: Sunday, February 3, 2013 1:17pm
>>> To: "Mark Constable" <markc at renta.net>
>>> Cc: cerowrt-devel at lists.bufferbloat.net
>>> Subject: Re: [Cerowrt-devel] stanford talk/deluged in hardware/yurtlab
>>>
>>>  Well, I see it for 320. Then you need to add a SSD, and a decent
>>> network card, and I suppose it could be made to work. Awful big, tho, in an
>>> era where I can get 1/2TB on an 2.5 inch SSD.
>>>
>>> What I'd wanted was closer to a dreamplug - 160 bucks, two network
>>> ports, but with an internal SSD. bonus points if it fit into a 1U rack and
>>> ate as little power as possible.
>>>
>>> Principal use case here is to be a "network monitor" with enough oomph
>>> to run stuff like cacti/mrtg/snmp tools, as well as do captures off of a
>>> mirrored switch port.
>>>
>>>
>>>
>>> On Sun, Feb 3, 2013 at 10:10 AM, Dave Taht <dave.taht at gmail.com> wrote:
>>>
>>>>
>>>>
>>>>  On Sun, Feb 3, 2013 at 10:03 AM, Mark Constable <markc at renta.net>wrote:
>>>>
>>>>> On 2013-02-03 09:18am, Dave Taht wrote:
>>>>> > I'm grumpy, as it doesn't have an esata interface internally,
>>>>> apparently.
>>>>>
>>>>> https://www.google.com?q=HP+N40L+MicroServer
>>>>>
>>>>> I know this is no where near an embedded device but I just got one of
>>>>> these
>>>>> on sale (new model out) for $220 and I think it's the most useful
>>>>> all-round
>>>>> cheap server box I've ever seen. Some people have it running 16 GB ram
>>>>> and
>>>>> I've got mine booting off an SSD via external eSATA. Very well built
>>>>> with 2
>>>>> x half height PCI slots (4 x eth port card?). Only missing USB3 ports
>>>>> and
>>>>> hot-swap drive space. And, very quiet with just an SSD.
>>>>>
>>>>
>>>> I'd be very interested to know how fast it could do packet header
>>>> captures.
>>>>
>>>> Line rate (gigE) would be good.
>>>>
>>>> Does it do BQL? (what is the onboard ethernet chips)
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>> _______________________________________________
>>>>> Cerowrt-devel mailing list
>>>>> Cerowrt-devel at lists.bufferbloat.net
>>>>> https://lists.bufferbloat.net/listinfo/cerowrt-devel
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Dave Täht
>>>>
>>>> Fixing bufferbloat with cerowrt:
>>>> http://www.teklibre.com/cerowrt/subscribe.html
>>>>
>>>
>>>
>>>
>>> --
>>> Dave Täht
>>>
>>> Fixing bufferbloat with cerowrt:
>>> http://www.teklibre.com/cerowrt/subscribe.html
>>>
>>
>>
>>
>> --
>> Dave Täht
>>
>> Fixing bufferbloat with cerowrt:
>> http://www.teklibre.com/cerowrt/subscribe.html
>>
>
>
>
> --
> Dave Täht
>
> Fixing bufferbloat with cerowrt:
> http://www.teklibre.com/cerowrt/subscribe.html
>

-- 
Dave Täht

Fixing bufferbloat with cerowrt:
http://www.teklibre.com/cerowrt/subscribe.html
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.bufferbloat.net/pipermail/cerowrt-devel/attachments/20130204/c9024c11/attachment-0002.html>