The LAN desktop: shared or switched? - clients should share 100 Mbps backbone for best LAN architecture - Network Management: Special Focus - Technology Information - Cover Story

In client/server environments, the real question is not whether to share or to switch, but whether to share the bandwidth directly or through a switch.

Switching is a very efficient way to increase aggregate bandwidth of a network. Improvements are made by isolating segments and allowing simultaneous transmissions to occur.

There is little doubt that switching in the backbone is the most cost-effective way of increasing performance and that there is little or no place for shared segments in the backbone.

What about desktop connectivity? Is switching all the way down to each desktop the ultimate solution? Is it true, as many vendors claim, that sharing can be eliminated and full bandwidth delivered at a low cost to each user?

The answer depends mostly on the network architecture. If you have a fully distributed network where there are no centralized servers, switching to each user is probably the right approach.

But most networks use client/server architecture with centralized servers. And there are strong indications that the pendulum is swinging back in favor of more centralized network services.

In this environment, the choice is not whether client connections should be shared or switched, because sharing is inherent in a client/ server LAN. The question is, what is the best way of sharing?

Right now, the "ultimate" desktop solution is a switch that provides 10 Mbps switched connections to the desktop and one or more 100 Mbps connections to the backbone.

Such a device usually provides an order of magnitude of improvement in performance compared to a shared hub with 10 Mbps connections all around, including the backbone. Switching to the desktop works.

But wait -- we have not really eliminated sharing. In a client/server environment, there is virtually no direct communication between clients. All communication originating at the clients is switched to the backbone and vice versa. So the 100 Mbps backbone connection is shared!

Are there other ways to share the backbone connection? Yes: you can connect the clients directly to the backbone using a 100 Mbps hub. Assuming your wiring infrastructure can support client connections at 100 Mbps, your choices are sharing the backbone connection indirectly through a switch or directly through a hub.

It turns out the cost of a 100 Mbps hub port is equal to or even less than the cost of a 10 Mbps switch port. So choosing how you share your backbone connection is not a price issue.

Which approach provides better performance? There are many ways of measuring performance, and by any reasonable measure the 100 Mbps hub will beat the switch. For example, delays are an order of magnitude lower, the peak transfer rate is an order of magnitude higher, and a large file transfer is 10 times faster at I 00 Mbps.

The switch will beat the hub only on one measure: throughput. The throughput of the switch is about 20% better. But throughput (maximum total load on the network at a point where delays are unacceptable) is a poor measure of performance.

The best measure for comparing desktop architectures is average available bandwidth or bandwidth on demand. The graphic shows bandwidth available to a client as a function of total load presented by other clients (sharing the same backbone connection).

We are using an idealized Ethernet with a "typical" throughput of 40% for a segment with many nodes and 50% for a segment with only two nodes (dedicated switch connection).

When no one else is on the network, the client connected directly to the backbone gets the full bandwidth of 100 Mbps, while the client connected via a 10 Mbps dedicated segment gets only 10 Mbps or 10% of the available bandwidth (thus the claim that switching to the desktop provides the full bandwidth is false).

If other clients are presenting a load, the amount of available bandwidth decreases for the direct connection but stays constant for the dedicated connection. Since most networks are lightly loaded most of the time, however, the available bandwidth is significantly better for the direct connections for most of the useful range.

Average bandwidth on demand depends on the probability distribution of the load on the network. If we assume a uniform distribution (a very conservative assumption since most networks are lightly loaded most of the time), the direct connection provides 50 Mbps of average bandwidth on demand while the switched connection provides 10 Mbps. By this performance measure, the direct connection is at least five times better.

If the major reason for considering a 10 Mbps switched desktop connection is performance and if your infrastructure (wiring and NICs) is 100 Mbps-ready, you should consider using a directly shared backbone architecture instead. You will get at least a five-fold increase in bandwidth on demand and your cost may even be lower.

COPYRIGHT 1997 Nelson Publishing
COPYRIGHT 2004 Gale Group