[Beowulf] Jury rigged ethernet?

Robert G. Brown rgb at phy.duke.edu
Mon Jun 27 13:52:19 PDT 2005


Ed Karns writes:

> ... below ...
> 
> 
> On Monday, June 27, 2005, at 10:22  AM, Reuti wrote:
> 
>> Hi,
>>
>> point to mention is the most likely unknown wave resistance of the  
>> existing wires. You have
>>
>> 150 Ohm for IBM Type 1 .... possible to use.
>> 100 (sometimes 105?) Ohm for CAT3/5 twisted pair ... The best  
>> candidates ... if you have enough pairs between this cabling and above  
>> ...
> 
> ..........
> 
> These cables below are not recommended. Using coax will only make your  
> connections unreliable unless you can make or use a breakout box with  
> the proper connection converters (RG type coax conversion to RJ type  
> connectors = bad news)  ...

As I also pointed out to several people in offline remarks, impedance
matching is a known problem with known solutions for at least some of
the more common STP to UTP situations.  The biggest problem occurs where
one goes from STP (say, the lines into said "tight space") to UTP (say,
a UTP/Cat 5 patch cable).  You tend to get reflections at these
junctures that can increase error rates if the UTP and STP runs are of
comparable length and the run lengths are "bad" lengths relative to
carrier wavelengths (where standing waves can be established via the
reflections).

One solution is to avoid such junctures and use STP end to end, primary
and patch cables alike. However according to this cisco white paper
(that examines the issue from the point of view of organizations needing
to reuse old e.g. token ring or phone wiring):

  http://www.cisco.com/en/US/netsol/ns340/ns394/ns74/ns149/networking_solutions_white_paper09186a00800b4249.shtml

it isn't a big issue for short runs, either of the primary STP cable or
short patch cables on the ends.  To quote:

  ...short lengths such as 90 meters of Token Ring A-suffix STP cable
  and 5 meters of UTP patch cords on each end should not cause
  significant problems.

at least through 100BT.  

They do discuss using impedance matching adapters for various cable
types, finding that in a lot of cases they don't really help much.

Honestly, I don't think that a 30 foot run (ten meters, barely longer
than their five meter limit in toto) will be too much of a problem,
although one might have to experiment a bit and get very clean
connections to the terminators on both ends.  Termination per se is
likely to be the biggest issue, with or without adapters or baluns or
impedance matching as empirically determined to be necessary.

> 
>> 50 Ohm for coax 10Base2 (RG58)
>> 93 Ohm for coax of the IBM 3278 Terminal connection (RG59)
>> ...
>>
>> There are adapters (called baluns) around, to convert it at both ends  
>> of the wire.
> 
> DO NOT incorporate any impedance matching devices in your connections  
> if possible. (Do not use these "baluns" as they will only add  
> complications, increasing line losses and decrease reliability.)

Ya.  Don't use coax at all is good advice.

> You will find that the impedance matching capabilities of most modern  
> ethernet adapters' chip sets will overcome any irregularities in the  
> actual cabling over short lengths. As a point of reference, two wire  
> coat hangers in parallel have approximately 150 Ohms impedance ... so  
> using the available (unshielded) twisted pairs of the cables available  
> will produce the more reliable results when attached to modern ethernet  
> adapters. (Using the center conductors of the coax would probably work,  
> but then you would also have to deal with termination of the coax  
> shields.)

Ah, I see.  "Coat-hanger net":-)

I'll have to try this one at home.  If I can find a few meters worth of
coat hangers...;-)

> FYI: You may be better off disregarding the cables available and simply  
> using one of the "plug in the wall" power line network adapters  
> instead. This would be done by using one of the three phases (" A  
> single existing cable provides 3 phase power, ...")

Yeah, but the Cisco paper suggests that worrying about it is probably
pointless as it will very probably just work at these dimensions, at
least through 100BT, as I originally suggested.  GigE I'm not so certain
about, but if he wires up all four pairs it is easy enough to test.

However, it certain CAN depend in some detail on just what KIND of STP
he's got in there, though, and until this is known and researched who
knows how things will go?

   rgb

> 
> Ed
> FireWireStuff.com
> 
>>
>> Cheers - Reuti
>>
>>
>> Ed Karns wrote:
>>> Not a good idea, although it may actually work. ... not necessarily   
>>> doomed to failure, but not industry standard practice.
>>> Test before you commit.
>>> Ed
>>> FireWireStuff.com
>>> On Sunday, June 26, 2005, at 12:04  PM, beowulf-request at beowulf.org   
>>> wrote:
>>>> Today's Topics:
>>>>
>>>>    1. Jury rigged ethernet? (Huntress Gary B NPRI)
>>>>  .....
>>>>
>>>> --------------------------------------------------------------------- 
>>>> -
>>>>
>>>> Message: 1
>>>> Date: Wed, 22 Jun 2005 22:37:29 -0400
>>>> From: Huntress Gary B NPRI <HuntressGB at Npt.NUWC.Navy.Mil>
>>>> Subject: [Beowulf] Jury rigged ethernet?
>>>> To: "'beowulf at beowulf.org'" <beowulf at beowulf.org>
>>>> Message-ID:
>>>>       
>>>> <A506359028CAD21191C10008C75D895C0C6D5248 at NPRI54EXC14.NPT.NUWC.NAVY.M 
>>>> IL >
>>>>     Content-Type: text/plain
>>>>
>>>> This is more than a little off topic but if the answer is  
>>>> encouraging  it will strengthen my case to purchase a small-ish  
>>>> cluster  significantly.
>>>>
>>>> Basically I am exploring the idea of placing a cluster in a  
>>>> dedicated  "non-traditional" physical location where space is at a  
>>>> dramatic  premium.  This particular space has existing cabling that  
>>>> I cannot  change and it is my only electrical access.  A single  
>>>> existing cable  provides 3 phase power, a few DC discrete signal  
>>>> pairs, and a few  twisted shielded pairs that were used for serial  
>>>> communication.   The  cable is approximately 30 feet long.  I would  
>>>> like to adapt two  twisted pairs for ethernet.
>>>>
>>>> What are the chances that I can coax 10Mbps out of this connection   
>>>> with modern (hopefully tolerant) ethernet cards at either end?  Is   
>>>> this doomed to failure?
>>>>
>>>> Thanks,
>>>>
>>>> Gary H.
>>>>
>>> _______________________________________________
>>> Beowulf mailing list, Beowulf at beowulf.org
>>> To change your subscription (digest mode or unsubscribe) visit  
>>> http://www.beowulf.org/mailman/listinfo/beowulf
>>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://www.scyld.com/pipermail/beowulf/attachments/20050627/633851c9/attachment.bin


More information about the Beowulf mailing list