|Thread Links||Date Links|
|Thread Prev||Thread Next||Thread Index||Date Prev||Date Next||Date Index|
Speaking as a customer, we would likely make use of almost as many 10M links as those > 10M. We currently use GigE links for both servers (1500 MTU and Jumbo) and as connections for distribution switches. We find the 0.5M Copper GBIC's less than useful, but about 35% of our links are <=10M. A low-cost 10-25M Coax would be quite useful as most of our connections in the data centers can utilize this length. The three predominant uses for our 1G connections today are spread fairly evenly about 1/3rd each:
MUX'ing 100M links into larger servers/clusters - This allows us to consolidate several 100Mb links into 1 or 2 Gb links with out regard to the load-sharing issues created by channel-aggregation protocols. (1500 MTU) (~1/2 of these are <=25M in length)
High-bandwidth pipes between clusters and their storage management systems. - Tape libraries, records archiving and retrieval and other data import/export connections. These are almost exclusively Jumbo Frame connections. (9K MTU) (~1/2 of these are <=25M in length)
Distribution links to remote buildings and floors centered away from the data center. (1500 MTU) This is mostly because, as Radia Perlman once pointed out, a link with enough capacity needs very little in the way of management.
So I suspect that 10G will be used by us in a manner similar to the way we utilize 1G, just moving more stuff... We did basically the same migration moving from shared 10Mb -> switched 10Mb, 10Mb -> 100Mb and 100Mb -> 1Gb.
All our 1Gb connections are SX/Multimode. We are seeing others utilize these GBIC's for >1Km on plain-old-in-the-ground-who-knows-what-the-modal-bandwidth-is 62.5/125 FDDI-spec glass. While we do have high GBIC failure rates (~10% per year), they are not distance related since they have been hard-failures and are on links < 100M in length.
We used the specs and installed 50um and have since been slowed in our use of it because the newer connectors (like MT-RJ and others) have made it problematic to field install the new connectors on 50um fiber. Since we pulled new 62.5um next to it we have used in exclusively for both our 1Gb Ethernet and our Fibre Channel for all connections, regardless of length.
< 1% of our links are single mode and < 2% are > 1Km...
Just some observations...
From: Rick Walker [mailto:walker@xxxxxxxxxxxxxxxxx]
Sent: Wednesday, April 19, 2000 6:58 PM
Subject: Re: 850 nm solutions
> Jim Tatum writes:
> But why does it matter? Why limit the users? Why not put in the table. It
> costs nothing. Just put in what the model and data tell us to. It is
> my opinion that a large percentage of 10GB style links are going to be
> very short, less than 10m. If you look at the way many fiber ports
> are being used today, many are in the 10m range. Also, since copper
> cables are going to be EXTREMELY challanged to go that distance at
> 10GB, why not let the market choose the lowest cost solution using
> 850nm VCSELs and 62.5um fiber?
FWIW, I agree that 10G across CAT-6 or other twisted pair would be very
difficult. However 10G across coaxial cable is fairly easy. It can be
done with 0.1" diameter coaxial cable using simple NRZ data encoding. A
simple FIR pre-equalizer can double this distance. Without a doubt
copper would be the cheapest solution for links under 10M. I would
estimate a mature chipset price of about $50 per end and $15 for the
This performance was demonstrated in 1998 using a 25GHz bipolar chipset.
See: Walker, R. C., K. Hsieh, T. A. Knotts and C. Yen, "A 10Gb/s
Si-Bipolar TX/RX Chipset for Computer Data Transmission" , ISSCC Digest
of Technical Papers 41(February 1998), 302,303,450.
A Copper PHY was voted down by the committee because it was thought that
there was no market for this type of low-cost short distance link.