To test each rack-mountable server's raw computing power, we used Neal Nelson & Associates' Business Benchmark. This measures relative performance of several core operations, including office and database workloads, transaction processing, and combined integer and floating-point calculations. It also includes a suite of tests designed to gauge server load and file I/O. Tests ran while the servers were disconnected from the network.

To evaluate network performance, we used Network Computing's FileMetric 1.0 written by Coffee Computing Corp. (www.coffeecomputing.com). It's a multithreaded client/server, master/worker software package that benchmarks file-server performance using a group of client machines to produce traffic. The file server can be any combination of hardware/software that can serve files to Windows95 or NT clients. FileMetric was set up to test file reads across SMB (Server Message Block) file systems, each of which can produce multiple test engines.

We used 14 Fast Ethernet engines running on Pentium clients on an idle segment to hit each server with the same test data. Each client machine ran a client engine, which created as many threads (workers) as needed to conduct the tests. During a test run, the number of engine threads can vary widely, depending on the I/O load of the server. Using a script designed to test network I/O, we ran each test a number of times and averaged the results for a read throughput mean. Each server was configured with 256 MB of RAM, 512 KB of cache, a single processor and Windows NT 4.0 with Service Pack 3, and ran identical network protocols and services.