100BaseT EMI ?
becker at scyld.com
Wed Mar 27 10:21:28 PST 2002
On Wed, 27 Mar 2002, Huntress Gary B NPRI wrote:
> I am drafting a proposal for a small cluster that will be used in a
> tactical military environment (Naval), thus I am bound by some
> environmental issues, particularly in the area of electromagnetic
> Right now I don't want to address the whole cluster, but is anyone
> aware of EMI issues specifically related to 100 Base T ethernet?
Most Ethernet clusters don't have a problem with EMI (Electro-Magnetic
Interference), but it's something to keep in mind when designing a large
system or your own enclosure. EMI and electrical/fire safety are often
forgotten elements of putting together clusters.
In general Ethernet is relatively immune to EMI with short runs.
However within a cluster the situation is worse than you might guess
from non-cluster experience. Most of the network traffic is
synchronized, and coincides with power-use spikes.
[[ OK, ok, that was primarily to point out why this topic is relevant to
the Beowulf list. ]]
Here are a few points to keep in mind:
Bad connections are responsible for most of the generated EMI:
don't un-twist pairs or strain the cable at the connector.
Kinked, crimped or stretched cables are very bad:
use 10X diameter (typical: 2" radius) around corners with UTP
Generating EMI and noise susceptibility are often symmetrical: a good
antenna works both for Tx and Rx.
10baseT generates a strong electrical noise spike at 20Mhz while active.
100baseTx operates continuously and spreads the noise, but some
data patterns can generate strong noise spikes (look up 4B/5B).
Gigabit copper puts more energy into the cable, but has less chance of
a single frequency noise spike (8B/10B). Gigabit transceivers are
designed to handle "echoing" mismatched cables, but are relatively more
susceptible to external noise.
You should pick cluster software or hardware control that doesn't use a
KVM switch. Using a KVM switch results in far more cables, and thus more
opportunities for emitted EMI. It's easy to filter noise from the power
cable, which operates at 60Hz. It's somewhat harder for keyboard and
mouse (which have sharp 5V swings from very busy chips), difficult for
Ethernet (1V carefully rounded waveforms from network-only
transceivers), and very difficult for video (1V sharp-edged pattern
> My gut feeling is that good quality shielded cables will be the first
> line of defense, however I'm concerned about radiated emission levels
> at the connectors.
It's counterintuitive, but UTP (unshielded twisted pair) is often better
than STP (screened/shielded twisted pair). The metallic shield can
couple a point source of noise or imbalance to the make the whole shield
On a general note: Most clusters are built using steel cases on steel
shelves or a rack. This provides both shielding along with electrical
and fire safety. Don't use an open frame design unless you have an
overall metal enclosure. Think about fires before using wood or
plastic, even for shelves.
Donald Becker becker at scyld.com
Scyld Computing Corporation http://www.scyld.com
410 Severn Ave. Suite 210 Second Generation Beowulf Clusters
Annapolis MD 21403 410-990-9993
More information about the Beowulf