subrate SFP?

Saku Ytti saku at ytti.fi
Fri Aug 30 05:59:24 UTC 2013


I got quite a bit of replies from sellers selling me cuSFP, insisting they
work.

So I'd like to clear up on this. For 10/100 to work on SFP slot, the PHY in
the host needs to be multirate. Exception is SGMII which supposedly
supports magic mode where SFP can ask it to send same bit 10 times, then
SFP can discard 9/10 bits, to remain very dumb yet deliver 100M client on
1GE host.

RGMII does not support this trick and this trick does not bring you down to
10M. One box that we have right now, which can't do any of this is ME-4924.

There is absolutely no reason that you couldn't deliver 'media converter'
or '2 port switch' in a SFP casing, to get that 1 10/100 port in every
4500-X or EX4550 port you need to cater some legacy. If my desire is odd (2
people have expressed off list they want same) this won't be built. But if
this is somewhat common demand and missing product, we can certainly get
such SFP built.

Obviously this SFP would cost bit more than normal cuSFP, as it needs to do
rudimentary buffering, packet dropping and it needs to have frame parser.


On 29 August 2013 23:38, joel jaeggli <joelja at bogus.com> wrote:

> On 8/29/13 6:08 AM, Saku Ytti wrote:
> > How do people deal with situation where you need <=48 SFP/SFP+ ports, but
> > you occasionally need one or two cu 10/100 ports?
> arista 7050s support 100 Mb/s on their copper sfp I have leveraged that,
> if you can break out the 40Gb/s ports you have as many as 64 ports of
> 10Gb/s. there are other switches that I've seen do this but they're not
> common.
>
> My problem is mostly around PDU/CDU management, in racks that otherwise
> would be 10Gb/s only and in general I've addressed it with dedicated
> switches  that support many of these devices rather than just two.
> > For some reason it's becoming quite rare for SFP port to natively support
> > 10M and 100M rates.
> >
> > Technically obviously solution to me would be subrate SFP, which presents
> > itself as 1GE to host, offering 100M or 10M to client. This would
> obviously
> > break QoS at the host as host would still think it's 1GE and SFP itself
> > would need to drop+buffer. But for my applications it would be fine, the
> > 10M or 100M ports are typical some MGMT access interfaces.
> > I can't imagine such SFP being complex or expensive, considering we have
> E1
> > over IP in a SFP, which includes control-plane and forwarding-plane
> inside
> > SFP form-factor
> >
> > Is this demand peculiar? Could I source such SFP somewhere by showing
> there
> > is demand?
> >
> > Putting 2 port switches or fibre converters with external PSU just to
> > support few ports seem dirty.
>
>


-- 
  ++ytti



More information about the NANOG mailing list