Very Computer

Can anyone tell me where I can find `rule-of-thumb'
numbers on acceptiable network errors and collision
rates? I have a network that is having performance
problems and I don't hove a baseline to compare it
to. Any help would be appreciated. Thanks.

Please respond with email.

Branden

Collision rates are a worthless metric. Unless they are exceeding 100%, they
tell you very little about the condition of the network.

--

Networks and Communications Consulting        (408) 395-5700
21885 Bear Creek Way                          (408) 395-1966 fax
Los Gatos, CA 95030
"...specialists in Local Area Networks and Computer Communications Systems"

Quote:>Network *errors* (i.e. CRC errors) should be virtually non-existent on any
>Ethernet. I have rarely seen more than one per DAY that was not the cause
>of a malfunctioning device or improper cable configuration.

Any transmitter which uses "early transmit" [ie xmission begins
before the entire packet is buffered by the network controller]
is vulnerable to host latency.

If the latency is excessive, the tx fifo runs dry and the poor
nic has no choice but to generate a packet with a bad crc.

Malfunctioning device?  A bit of a theological question.  The
hardware was designed this way, and is functioning as spec'd.

--
#include <std.disclaimer>

While this may be true for 10Mb/s Ethernet, I wouldn't set this
expectation for
100Mb/s systems. Even though the BER is lower for 100BASE-TX, you are
sending more
bits in less time and will therefore perceive a 'compression' of errors.
That is,
when observed over an hour instead of a day you may actually be seeing
errors that
are well within the BER specification of the physical layer. In
practice, this
perceived increase in errors is not a problem. Moving toward Gigabit
Ethernet
one would need to scale the BER requirement along with the data rate in
order
to maintain the same expectation. This would not be a good thing for the
technology
since BER is directly related to cost. One of the reasons for this is
that the
BER of the PHY would need to be so high to make the implied guarantees
that it
would take years to actually simulate and test each part to meet the
tighter
specifications without making extrapolations.

I would advise against creating these 'rules of thumb' since they lead
to
dangerous myths and preconceptions. Still, if you must, an acceptable
measure
would be to calculate the BER based on the CRC error count and compare
it to
the standard requirement. This comparison is somewhat unfair to the
physical layer
since this number will be different than the BER rate of the physical
layer
itself because CRC errors are a measure of total system BER. Don't be
alarmed by
a few orders of magnitude difference.

Since a single bit error is enough to cause a CRC error, I would count
each
CRC error as a single bit error. This may not strictly be the case for
errors
caused by bursty noise, but some approximation needs to be made in a
packetized
network.

Much like the touted 30% collision myth, CRC errors do not spell
disaster until the
BER is very close to 1...your definition of 'very' may vary :-)

BTW, I agree 100% on collision rates.

The opinions above are mine, all mine. (And I'm a happy miser.)

The problem is that an event which is NOT a network error APPEARS as a network
error, if one looks only at a CRC error count. If you correlate this statistics with
a count of Transmit Underruns of the devices in the network (and this is a common
SNMP MIB object), you can determine what the "real" network error rate is.

--

Networks and Communications Consulting        (408) 395-5700
21885 Bear Creek Way                          (408) 395-1966 fax
Los Gatos, CA 95030
"...specialists in Local Area Networks and Computer Communications Systems"