Re: Does Ten-Gigabit Ethernet need fault tolerance? (nonredundant NICs)
Some clarifications below.
From: Joe Gwinn <gwinn@xxxxxxxxxx>
To: stds-802-3-hssg@xxxxxxxx <stds-802-3-hssg@xxxxxxxx>
Cc: rabynum@xxxxxxxxxxx <rabynum@xxxxxxxxxxx>
Date: Wednesday, July 28, 1999 12:44 PM
Subject: Re: Does Ten-Gigabit Ethernet need fault tolerance? (nonredundant
>At 8:28 PM 99/7/27, Roy Bynum wrote:
>>You wrote "There is no problem with parts of the segment having
>>NICs,". With a full duplex, point to point link, there are only two
>>at each end. It was my understanding that 10GbE was to be a full duplex,
>>point implementation only. Within a data switch, at layer 2 or layer 3,
>>links to multiple systems can be implemented, creating virtual segments;
>>10GbE link is itself, full duplex, point to point. It is the single 10GbE
>>I am writing about when I refer to fault tolerance. I am not writing
>>allowance at the virtual segment level, as in the case of allowing for
>>switch failures within the virtual segment. Are we talking about two
>I suspect you are correct. In my nomenclature, a "link" is bidirectional,
>having two independent fibers, and connects NICs to hubs. It seems to me
>that when one speaks of a link in GbE, it's the path from NIC to NIC (via
>an unnamed hub); the hub is assumed. In any case, we will need to arrive
>at a common and self-consistent nomenclature.
802.3 generally defines "link segments" which are symmetrical and are at the
lowest level. In other words, there is (mostly) no differentiation between
types of link segment. A link from a NIC to a switch is the same as a link
between two switches and is the same as a link between two NICs by which I
mean a link from NIC A to NIC B without any other device in the middle.
Generally, there is no definition of link as between two NICs with a switch
in the middle. The definition used in 802.3u was
"1.4.110 link segment: The point-to-point full-duplex medium connection
between two and only two MDIs.
>Is 10GbE expected to have hubs, or will all links be physically point to
>point between pairs of NICs? One would expect that there will be hubs in
>10GbE as well, as inability to handle anything but pairs of NICs would be
Up to 1 Gbps speed, there has been two modes of operation written into the
standard: half-duplex and full-duplex. At the time we started the 1 Gbps
work, there was no standard full-duplex mode of operation (that came in
March 1997 with 802.3x) so we felt compelled to include half-duplex
operation. Practical experience is that no one (in my personal knowledge)
is building any half-duplex equipment at 1 Gbps. Since we now have a
full-duplex component to the standard, we could craft a 10 Gbps standard
that is only full-duplex. I suspect this is what most participants want so
we might save ourselves the bother of the half-duplex work.
Hubs are generally shared devices using half-duplex operation so following
my previous opinion, I expect 10 Gbps to be an all-switched environment with
>So, my mental picture of 10GbE has been a star topology, with a hub in the
>center and the NICs at the points of the star, with duplex fiber-optic
>links connecting each NIC to the hub, one link per NIC.
I agree, except that the hub changes to a switch.
>In RTFC, one has the same number of NICs as before, but there are two (or
>more) hubs, each hub having its own star of links to those NICs. Each NIC
>now has multiple duplex ports, one per hub. The whole affair, containing
>NICs and hubs, is called a "segment". Bridges between segments are
>two-headed NICs, one head per segment, even if the bridge NIC happens to be
>physically colocated with a hub.
>If one of the NICs is connected to only one hub, and nothing is broken,
>rostering will cause that NIC to be included in the segment, and thus to be
>accessible to other NICs. This behaviour is automatic. If a NIC having
>only one hub connected breaks or loses contact with that hub, rostering
>will isolate that unfortunate NIC. This behaviour is also automatic.
>>Joe Gwinn wrote:
>>> At 9:12 PM 99/7/24, Roy Bynum wrote:
>>> >Does RTFC allow a minimally trained individual to simply plug two fiber
>>> >pairs into the 10GbE interface to implement fault tolerance and if a
>>> >pair, parallel to the first, is not plugged in the fault tolerance is
>>> >implemented? This will be the simplest and most common implementation
>>> Yes, this will work, by design. The rostering algorithm will just treat
>>> the missing path as broken, and press on. There is no problem with
>>> of the segment having non-redundant NICs, although those NICs will be
>>> out of the segment if those NICs or their links fail.