We ran two tests to get an idea of the server's CPU and file system performance. For the CPU test, we used Bluecurve Dynameasure SQL Professional 2.0. For the file test, we used Bluecurve Dynameasure File Professional 2.0. The file test was run twice against the server - once against NT 4.0 and once against NetWare 5.0. The CPU test was run only against NT 4.0 because Bluecurve's SQL tests can't run under NetWare.
The file test was designed to highlight the performance of the file subsystem. We transferred 12K-byte files between all the clients and the server bidirectionally. We ran the test six times, increasing the number of simulated users with each iteration. We started with one user, then increased to 10, 20, 30, 40 and 50 users.
Note that these simulated users are worst-case, not typical, users, which we used to more thoroughly test the server's file system. They wee configured to send out a new file request as soon as the server services a previous request. In reality, the number of actual users that a server can support should be higher than our numbers. The CPU test was designed to stress the CPU subsystem. We used the Dynameasure SQL test to perform a mixed-read on-line transaction processing request against a Microsoft SQL Server 6.5 database on the server. The request included table joins to run up the CPU utilization.
As in the file test, we increased the number of simulated users with each iteration, from one to 5, 10, 15 and finally 20 simulated users.
Our network test used Ganymede Software's Chariot 2.2 to stress the network interface card (NIC) driver. We set up four bidirectional streams of traffic that all terminated in the server tested, balanced over two Fast Ethernet NICs. Each stream terminated at a separate client with a Fast Ethernet NIC. Half of the streams simulated HTTP GIF requests, and the other streams simulated HTTP text requests.
The Chariot test yielded a single number for network throughput with the processors at nearly 100% utilization. Because the processors reached a bottleneck before the NICs, a higher number means better network performance; an efficiently designed NIC driver will not utilize the processors as much per transaction and, therefore, can handle more traffic before the system's CPUs are maxed out. We measured the average aggregate bandwidth over a 10-minute run. We repeated the test five times and averaged the numbers to get a single value.
The core of the network was a Fast Ethernet switch. Each client and the server had its own switch port.
RELATED LINKS
Back to the main review
