Salvatore Salamone
As client/server computing becomes more common, network servers need to handle an ever-increasing number of I/O requests from client PCs. As networks grow, the bottleneck that limits network performance is most often associated with passing data between the server and the network over an Ethernet connection.
Network administrators trying to minimize this bottleneck have had several solutions to choose from. One is to upgrade the client and server machines to higher-speed networks, such as FDDI (Fiber Distributed Data Interface) or 100-Mbps Ethernet. Another option is to move the servers to high-speed networks and connect them to LAN segments with either routers or Ethernet-switching hubs. A third option, called load bal
ancing, offers a cost-effective way of handling the traffic that passes between a server and the rest of the network.
As the term implies, load balancing ``balances'' the traffic between a server and a network over multiple network adapter cards. Load balancing requires a bridge or Ethernet switch between the server and the clients (see the figure at left) and software running on the server.
For servers running NetWare, load-balancing software comes in the form of an NLM (NetWare loadable module), such as Balance.NLM from Network Specialists (Lyndhurst, NJ) or Switch.NLM from Kalpana (Sunnyvale, CA). As adapter cards are added to the server, the throughput increases proportionally. A typical load-balancing solution can increase the aggregate bandwidth between the server and clients from 10 Mbps to 20 or 30 Mbps via the addition of one or two Ethernet adapter cards in the server.
In addition to adding bandwidth, load balancing provides a level of fault tolerance. If one of the network adap
ter cards or network cables fails, traffic between the server and the network is simply passed to the other connections. That's in sharp contrast to the total disruption of traffic flow that takes place when a single high-speed connection linking a server to a network fails.
When does the performance enhancement from load balancing save money compared to moving to higher-speed LAN technologies such as FDDI and fast Ethernet? Surprisingly, the answer is simple if the server uses an Intel 486-class CPU. A 486-class server running a mix of common word processing, database, and file transfer operations in NetWare can deliver only about 20 Mbps, according to the results of an Infonetics Research (San Jose, CA) server-bandwidth congestion test that compared Ethernet switching/load balancing, fast Ethernet, and FDDI.
For the most common NetWare environments with 486 servers, that means load balancing with Ethernet switching is a practical, cost-effective choice, according to Michael Howard, president o
f Infonetics. The Infonetics study also determined that when a network server is upgraded to a Pentium-class CPU or better, then the processor is no longer the limiting factor. Howard says that for more powerful servers (i.e., Pentium or better) running a mixture of client/server applications, an upgrade to 100-Mbps technology will likely be needed.
Such sentiment is echoed by switching-hub vendors, who will most likely play a role in all three scenarios. ``In the short term, multiple [network interface cards] are the best solution today, because [this approach] lets you use inexpensive 10Base-T Ethernet cards and gives you a degree of fault tolerance,'' says Jim Goede, product marketing manager at Lannet (Irvine, CA). ``Down the road, you will see people moving to 100Base-T or FDDI to connect servers to the network.''
Figure: A Typical Network with Load Balancing
Load-balancing software running on a server balances traffic to and from numerous Ethernet LAN clients, while an Ethernet switch
ing hub is used to direct traffic between the server and the appropriate LAN segment.
Illustration: Graph: PCMCIA to Take Off on Desktop
Although PCMCIA slots are not prevalent on desktop PCs now, many PC vendors and analysts expect them to become much more common on the desktop for a variety of reasons: easy exchange of I/O cards and data among portable and desktop PCs; PCMCIA's small form factor, which makes for smaller-footprint desktop PCs; lower power consumption; and easier swapping of peripherals. ``PCMCIA is also happening on the desktop because people want to be able to use the cards that they bought for their notebook,'' says Jack Peterson, director of marketing at SCM Microsystems (Los Gatos, CA, (408) 395-9292), a company that plans on announcing products later this year that combine traditional I/O devices with PCMCIA. Martin Shan, product manager for DEC's PC MTE family of desktops that feature an optional PCMCIA slot, notes that PCMCIA peripherals are currently more expensi
ve than their ISA counterparts but says those prices are already dropping.
Load Balancing's Inexpensive Performance Boost
