The Windows NT Magazine Lab can model a variety of network
arrangements to test the performance of enterprise-class server hardware and
software. This capability presents several requirements that our hardware and
setup must satisfy.
For example, Internet testing requires simulating transactions from the
outside world, which necessitates IP addresses other than our own. Messaging
with Microsoft Exchange requires a domain model, so we set up a domain
controller. This system now also functions as a WINS server. The simulated
clients must participate in this domain, and the high loads we generate require
that we spread out client loads across multiple logical--and physical--networks
to prevent network I/O bottlenecks.
You can test Microsoft SQL Server (or any database) either in a domain or a
workgroup--in a 1000+ user environment, you're more likely to find a domain than
a workgroup. The Lab uses Bluecurve's Dynameasure for end-to-end capacity and
performance testing. Optimal arrangement of the testing hardware means separate
systems for the control server (which manages test operations and stores results
data), the management console (for administering tests), and the test server
(the system under load)--all of which need high-speed I/O pipes to the rest of
the client/server environment.
When we moved from 15 to 35 client workstations simulating user load, we
decided to spread these thousands of virtual users among several networks to
more closely simulate the real world and to remove network I/O as a potential
performance trap that would affect test results.
This move required additional networking hardware and another high-power
server. As you can see in Figure 1, we set up four VLANs: one for the resource
systems and three for the clients. One server acts as the DHCP and MPR system;
the Cisco switch keeps the VLANs physically separated, while still offering the
best possible performance; and the added hubs give us extensibility. This
functional network is reasonably easy to administer.