Microsoft Wolfpack, NCR LifeKeeper, Vinca StandbyServer for NT, and Data General NT Cluster-in-a-Box use a failover object model that can include IP addresses, applications, data files, and so forth. For example, you can use any of these products to create a high-availability cluster of a SQL Server or set of file services. We picked SQL Server and file services to test the functionality of each of these four clustering solutions.
We set up the four solutions in the Windows NT Magazine Lab's enterprise test environment, using a mix of servers and workstations running Bluecurve's Dynameasure 1.5 to generate and measure client load. (For more information about this testing software, see John Enck, "Dynameasure by Bluecurve: Born to Measure," November 1996.) We also used NT's Performance Monitor to track CPU utilization as services failed over from one cluster node to the other.
The test bed consisted of the host platform for each solution (servers and shared disk array--see the individual product reviews for details) connected to a Cisco Catalyst 5000 100Base-TX fast Ethernet switch, with a collection of Digital Equipment, Compaq Computer, Hewlett-Packard, and clone Pentium desktops (all P100 or higher, with 32MB of RAM), on 100Mbps Ethernet with Adaptec 10/100Mbps NICs and hubs. Although we didn't use an optimal network topology (we weren't testing performance in this review), we used several virtual LANs and NT Server 4.0 with Multi-Protocol Routing (MPR) software running on an NCR Globalyst S40 as a router.
We wanted to generate enough load so that the servers would have to do some work to fail over the attached clients (you can easily fail over a resource when no one is using it). We used Dynameasure to run 100 to 150 simulated SQL Server users. While the simulated users accessed the servers, we failed the primary node (i.e., we turned off the server). At this point, all SQL Server Named Pipe connections (and TCP/IP sockets) were dropped, but the databases failed over to the secondary nodes. We had to restart the Dynameasure test when the SQL Server service came back online and the database migrated to the secondary node, but we restarted the SQL users with the new connections via the alias.
Typical failover times were 15 seconds to 30 seconds, with or without user load, for NT Cluster-in-a-Box, LifeKeeper, and Wolfpack (we only recorded Wolfpack times for file share failover; for more information, see the Wolfpack review, page 72). StandbyServer's failover times varied more than the other products (see its review, page 78). We will run future tests on these solutions to determine the performance effects the clustering software introduces, how much load each system can handle, and so forth.
