100G QSFP28 DAC cables - experience

Thu Sep 14 18:12:29 UTC 2017

We're using a mix as well, some QSFP28 AOC, others DAC. One thing that you
need to keep in mind about the DACs is going to be the bend radius. These
things are girthy af, so make sure to either overestimate your runs
slightly, or buy one to test first.

On Thu, Sep 14, 2017 at 11:54 AM, Hugo Slabbert <hugo at slabnet.com> wrote:

>
> On Wed 2017-Sep-06 09:17:39 +0200, Jiri Prochazka <jiri at cdn77.com> wrote:
>
> Hi folks,
>>
>> I'm wondering if anyone have (either positive or negative) experience
>> with 100G QSFP28 DAC cables?
>>
>> Is there anyone who is using 100G DAC in large scale and would recommend
>> it (which means there are no issues compared to SR4 links)?
>>
>> I'm thinking about cables with lenght up to 1m, not more.
>>
>> We have had quite bad experience with 10G DAC in the past - but I do not
>> want to be slave of the past.
>>
>> Thank you for your thoughts!
>>
>> Jiri
>>
>
> We're deploying a decent chunk of 100G QSFP28 at the moment, but it's a
> mix of:
>
> - a handful of 100G QSFP28 copper DACs for some switch peerlinks
> - a bit >100x 100G QSFP28 AOC for interswitch links
> - a lot more 100G QSFP28 -> 4x25G SFP28 copper breakouts
>
> We're only a few weeks in at this point, so mileage may vary in the long
> run etc.
>
> The copper peerlinks are mostly 1M with some 3M.  We've had no issues with
> them so far.
>
> The AOC interswitch links vary more in length, but some of those are >10M
> (hence AOC rather than copper).  We've faced no issues with those.
> Granted, there is BGP with BFD running across those, so those should help
> in terms of liveness checks and such.
>
> I mention that because where we _have_ had issues are on the 100G -> 4x25G
> copper breakouts.  Those are for 25G edge connectivity.  It's a decent
> sample size with a bit north of 600x 25G ports.  The trouble we've had
> there have been with some links showing link up on the switch and server
> side but actually failing to pass any traffic, so we need to stuff some >L1
> liveness checks on there to ensure those links are good while we sort out
> the root issue.  It is not yet clear if this is a cable fault, driver
> issue, or something firmware-ish on the NICs.
>
> Also, fun fact: 25G only made its way into the 802.3ad bonding mode driver
> in the Linux kernel in March this year[1].
>
> --
> Hugo Slabbert       | email, xmpp/jabber: hugo at slabnet.com
> pgp key: B178313E   | also on Signal
>
> [1]https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/
> linux.git/commit/?id=19ddde1eeca1ee81f4add5e04da66055e09281ac
>