Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

RE: Ethernet Characteristics re: packet loss

Roy and all,

I've been trying to ignore this discussion because CSMA/CD efficiency isn't
that relevant given the high usage of full duplex today but I just can't
stand to see misconceptions propagated.

The "theoretical effective utilization" of 40% was a number derived from
theory on CSMA/CD algorithms that are easier to analyze but less efficient
than Ethernet's binary exponential backoff mechanism. Utilization is
strongly affected by packet size (because it takes just as long to arbitrate
the network to send a 64 byte packet as a 1518 byte packet. It is also
somewhat affected by network size though that effect is weaker. 

For network loads with a reasonable mix of large packets, Ethernet
utilization is much better than the 40% computed by research papers using
less effective algorithms even when not full duplex. The capture effect of
CSMA/CD enhances its efficiency (at the cost of producing somewhat poorer
sharing behavior).

Of course, real networks have lower utilization when the average is computed
over long times because traffic is bursty. Instantaneous utilization of an
Ethernet is always either 100% or 0%. Any other number depends partly on how
long a period is averaged.

If any of you want to understand efficiency and performance of CSMA/CD
throughly, read Mart Molle's excellent work on the subject.


-----Original Message-----
From: Roy Bynum [mailto:rabynum@xxxxxxxxxxxxxx]
Sent: Monday, March 19, 2001 4:56 PM
To: Stout, Tony; 'stds-802-3-hssg@xxxxxxxx'
Subject: Re: Ethernet Characteristics re: packet loss


What you were testing was obviously a full duplex setup.

In a the older CS/MA-CD half duplex environment the normal effective 
transfer bandwidth was about 30% on segments that have a large number of 
nodes.  On segments that had a low number of nodes, the theoretical 
effective utilization was 40%.  This limitation was because of collisions 
on the segment.

Full duplex setups do not support collisions.  With non-blocking Ethernet 
switches, it should not be hard to get 100% utilization between any two 
nodes.  For full duplex segments with large numbers of nodes, close to 100% 
effective utilization can be achieved if all of the nodes are sending to 
each other equally.  When any one node is the destination of an aggregation 
of multiple other nodes, such as on a workstation/server LAN the bandwidth 
utilization is limited by the bandwidth of the destination node.  With 
properly implemented flow control, the effective aggregate bandwidth on 
source nodes will match the destination node at close to 100%.

Thank you,
Roy Bynum

At 11:50 AM 3/19/01 -0500, Stout, Tony wrote:
>I have a quick question.... someone made a statement that it is a
>characteristic that once you go above 40% LAN capability, the you will
>to lose packets.  This doesn't seem to hold water with me.  I recently
>performed a test whereby we injected packets at 100%, 75%, 50%, 25% 10% &
>network capability.  We had 2 laptops, 1 desktop (PIII dual processor), a
>LAN Analyzer, and a SmartBits 200 generator -- at 100% both the laptops 40%
>of the packets we were sending -- the desktop captured 70% --- in both
>the application running on the platforms was using 100% of the processing
>time. -- my suspicion is that this is where the packets are being loss --
>the processing arena.
>Can you shed any light on this?
>Thank you,
>Tony Stout
>Information Systems Test Director
>Phone:  301-744-2690
>Mobile: (301) 399-9078
>Fax:  301-744-2688
>JITC Webpage:
>  <<Tony Stout (E-mail).vcf>>