Congestion occurs in a network device when packets (data) arrive faster than they can be forwarded.
Since switches have large number of inputs (which may regulary have packets destined for a single output), they likely sites of network congestion
To reach the an equivalent level of performance, output buffered switches require more buffers and a faster switch fabric than input buffered switches.
Congestion will fill all buffers. Dead time or spare capacity is required to empty full buffers.
Buffering is provided to deal with bursts of traffic that exceed the line speed of the egress port.
Head-of-Line Blocking is a statistical phenomenon associated with the input-buffered, FCFS, 1xfabric switches, that can result in lower than line rate output capacity on certain ports.
With some traffic patterns HOL blocking may contribute to congestion by reducing the capacity of heavily utilized outputs when there are 2 or more such outputs. Where there is only 1 heavily utilized output, HOL blocking is not a factor.
How does this relate to the GIGAswitch/FDDI?
It is an input & output buffered switch. (Output if buffered to make it possible to not lose packets when injecting them into the outgoing FDDI. Half duplex FDDI may not be fast enough to keep up with the switch's output stream.) The switch has a 1x crossbar fabric.
Output capacity may be reduced by running in half-duplex. Or, per-packet overhead can be a factor. There can also be a secondary effect due to HOL blocking, when several output ports are simultaneously congested
What can be done about congestion?
What could DIGITAL do?
TCP will push the network towards congestion.
For any congested switch, he best way to reduce congestion is to add ports to busy destinations.
Dealing with congestion by addressing HOL blocking can be expensive, and will have limited effect, since most effects of congestion are not caused by HOL blocking. If utilized, such measures should be applied to high cost links.
Curtis Villamizar from ANS says that Stephen does not understand the severity of the problem. Basically, having port blocking is unacceptable. Buffering on the input is a mistake. The product should have been fixed by putting a buffering on the output side. Stephen agreed to relay Curtis's concerns back to DIGITAL.
Curtis also notes that Stephen's suggestion about adding another switch (to reduce the expense of the hunt group solution) is a form of solution since packet do get off the input queue of the first switch.