[nSLUG] ethernet and the modern linux

George N. White III gnwiii at gmail.com
Fri Sep 26 16:23:38 ADT 2008

On Fri, Sep 26, 2008 at 11:45 AM, Daniel Morrison <draker at gmail.com> wrote:

> Wow.  My opinion differs with nearly everything that's been written in
> this thread so far, but I'll try to limit myself mostly to this point,
> which is flat out wrong.

"Your opinion is important to us".

I generally find that differences in opinion on technical issues come from
differences in assumptions about the environment.   That makes this a
linux issue because it is about a small pocket of linux in an ocean of
Windows PC's (and the Windows machine belong to the managers,
so as long as the managers think the network is healthy there aren't
going to be changes to make a few linux systems work better.

It is also a Nova Scotia issue because budgets are small, in the best
Scottish tradition.

> 2008/9/26 Jason Kenney <jdkenney at gmail.com>:
>> The reason being the number of pairs of wires used.  10 uses 1 pair up
>> and 1 pair for down, 100 uses 2 pairs for up and 4 pairs for down.
>> But 1000 uses all 4 pairs bidirectionally!
> 10BaseT and 100BastTX use two pairs: pins 1&2 and pins 3&6.
> Yes, there are some (much less common) variants such as 100BaseT4,
> which uses 4 pairs of CAT3 instead of two pairs of CAT5.  If you're
> using this, you probably now all about it already.
> 1000BaseT uses all four pairs.
> Another variant named 1000BaseTX uses only 2 pairs over CAT6, but is
> now uncommon.
> (How could 100BT use "2 pairs for up and 4 pairs for down"?  Which
> side is "up"?  It's symmetrical! <LOL>)
> BTW, WIkipedia says:
> Autonegotiation is a requirement for using 1000BASE-T[3] according to
> the standard. At least clock source has to be negotiated, as one has
> to be Master and the other Slave. Several physical layer device and
> drivers will allow you to force 1000 Mbit/s full duplex to eliminate
> autonegotiation issues. In this out of standard case, net admin had to
> care that one peer is clock Master, and the other is clock Slave.
> Other opinions:
> 1) turning auto-neg OFF is the leading cause of mismatches, because people:
>  - forget to set both sides of the link
>  - forget to 'save' the settings permanently
>  - forget to document the settings
>  - move patches without consulting the documentation
> Second leading cause of mismatches is buggy firmware/drivers.
> Auto-neg is an excellent choice for stable 24/7 networks.  The only
> reason for it to fail at boot (apart from buggy NIC implementations --
> get rid of those NICs!) is if the switch is not UP when the system
> boots.  In a 24/7 environment this should never happen.  Even if it
> does happen, auto-neg will occur when the switch comes up anyway!

In an environment with a whack of Desktops running XP, the switch
probably gets hammered after a power hit.   We should make everyone
use a laptop so they don't  get hit by power outages.

> 2)
>> One view is that a network where systems think the cable is disconnected is broken and should
> be fixed.
> Your view is that networks where systems think the cable is
> disconnected should just be left broken?

Not exactly.  My view is that the world has many such networks in big
and nobody is going to fix them.  It has been thus for years.  The
only workaround is
to find the problem switch lives and stand around fuming until someone
smells the
smoke an turns a firehose on it, and even then you have no guarantee that the
replacement will be better.

>>  My view is that shit happens, and we should configure systems to be robust.
> i.e. auto-negotiation -- much more robust than manually configuring
> the wrong settings.

so use the right settings.  We have been using fixed-100 for years with SGI.
The only downside is that when the machine room flooded and we moved systems
out I had to ask to have the ports set to fixed 100 in the new location.  With
PC's running WinXP and Intel interfaces, fixed-100 full duplex did not work
until last Jan. when Intel finally released a driver that supports fixed-100.

>>  Apple, who have lots of experience with a tightly controlled list of network interfaces,
> <ROFL>
>> recommend settings switches to force half-duplex, which is what you get when autoneg fails.
> On Apple equipment, maybe!
> Apple makes nice laptops and desktops.  Their servers are a joke.

You don't get high-end gear on that Scottish budget.   Low end Dell and
Apple both use Intel system boards -- often the same model with
minor tweaks.

> 3) udev has nothing to do with bringing interfaces up and down.
> Use auto-negotiation, and avoid all the junk dealing with differing
> module parameters, different ethtool options, etc.
> Full duplex is faster.  Half duplex results in collisions when both
> sides attempt to communicate with each other at the same time.
> Doesn't matter if you cannot sustain 100 Mbit data stream -- I'm
> talking about latency, not bandwidth.

FD is faster when it works, but if the machine loses the net connection
nothing moves.   If HD would work reliably, it may be the best approach.

If HD allows us to get more done than FD, but is too slow, then we have
an argument management can understand:  "we turned on the supercharger
(FD) and the network kept blowing up, so we have to run without boost,
but then we don't get enough power -- send money!"

George N. White III <aa056 at chebucto.ns.ca>
Head of St. Margarets Bay, Nova Scotia

More information about the nSLUG mailing list