How would you go about establishing a number of Player Agents to run a load test. An example would be a load test performed with 2000 VU on a web based application. How many machines would I need to simulate that or what requirements should I look at to establish this.
At the minimum two, one for the majority of the load and the other running one virtual user of each type as a control factor (start the virtual users on this second control generator first). If the virtual users performance degrades at the same rate, then the bottleneck in performance is in the application (or possibly the network if you insist on testing across a slow wide-area network), not the generators.
if, on the other hand, the lightly loaded generators users begin to show better response times than the remainder of the set, then you are hitting a bottleneck in your load generator that is coloring your test results.
How many generators per box? Those figures are a lot like EPA estimates. It will vary from manufacturer to manufacturer. It will vary depending upon the complexity and size of your virtual user, how many parameters and how much prcessing occurs in your client. The more math, the higher the file I/O, the greater number of parameter substitutions, and other factrs all contribute to the overall system weight of a virtual user.
2000 might be fine for a HP V class server to host, but maybe not for the Windows 2000/XP box sitting on your desk.
My rule of thumb, always include the control factor, limit the network size and complexity, be very conservative on the number of users per box and monitor very heavily all of the host stats.
As per me, a VU takes about 5 MB of space in memory. You reserve say 128 MB for OS e.g. Win2K an rest of the memory you can have for VU's per agent. For example, if you RAM is 1024 then
1024 - 128 = 896 MB
896/5 = 175 VUs Approx.
You can reduce it to 150 for safety purpose.
A highly advanced bug is indistinguishable from a feature.