From Louan Handke on Sat, 01 May 1999
What is the difference between switch hub and unswitched hubs
The traditional ethernet hub (concentrator, repeater, etc) is a relatively simple device which just amplifies the signals on any of its ports out to all of the other ports. A "switch" or "intelligent" hub is more of a multiport bridge. It "learns" which MAC addresses (ethernet hardware assignments) are on each of it's ports and only "repeats" (rebroadcasts) data frames to the appropriate port.
In a traditional hub only one system on a given network segment can be "talking" at any given time. The whole network segment is virtually a single wire. Any time two or more systems attempt to send packets at close to the same time there is a "collision." This is called CSMA/CD --- carrier sense (listen for quiet), multiple access (any card and "speak up"), with collision detection.
Whenever a collision occurs the cards involved send a short jamming signal, and then they perform a psuedo random "backoff" delay before attempting to re-broadcast. Since it is incredibly unlikely that two cards will choose the same amount of backoff delay one of them will usually "win" and get to send first. That's fine with only a couple of cards in contention. However, as utilization approaches 20% or more, the number of collisions skyrockets and the overall average throughput drags to a crawl.
The traditional answer was to segment the systems --- putting servers in close proximity to their clients (work groups), put routers between segments, and put lots of interfaces in your workgroup servers (four to eight ethernet interfaces was not unusual for big netware servers).
Etherswitches are used to alleviate some of these problems. On a 24 port etherswitch its theoretically possible for 12 pairs of systems to be concurrently exchanging data frames. This allows for much larger segments (called VLANs --- virtual local area networks).
On the downside, etherswitches are typically much more expensive than their more passive cousins. They have to contain processors, memory, and firmware. In addition their processors have to be pretty quick (usually quick RISC chips with a mess of ASICs I guess). Also there are degenerate cases. If all of your servers are located on one or two legs of an etherswitch then it won't help much. All of the clients will be waiting for that one (or those couple of) port(s) to be clear --- a classic bottleneck.
Again the solution is to have lots of smaller servers --- segment the network, and replicate the data and services so that they clients tend to use local copies of everythings. Hierarchies scale!
(Not to say that etherswitches don't have their place --- its just to say that their deployment should be based on an understanding of the situation and the benefits vs. the costs of the technology. Most vendors have little interest in your needs --- they want to sell you the shiny expensive toy).