Actually in a Server+Publisher environment, there are two causes of high memory usage:
- Reload processes (QVB.exe) may need massive amounts of RAM when loading/joining/storing large QVDs. If run in parallel, of course this high water mark may rise considerably. The good thing about these processes is that they will release memory as soon as they don't need it anymore. The bad thing is that high memory usage in this scenarion experiences spikes that are difficult to intercept and may still cause reloads to fail. Use Windows Performance monitor to check for reload problems.
- The QlikView server process (QVS.exe) will load QVW models in RAM and keep them there as long as they are being used. QVS actually serves as one gigantic cache. The good thing about this is that response times are very fast (for a well designed model on a suitable server/cluster). The bad thing about this is that QVS will NEVER release memory, unless you restart the service. You cannot easily control the QVS cache. See Maxim's post.
When you combine the two (intra-day reloads), and your memory usage reaches the upper limit, either one of them may fail to execute or to respond.
Check the size of your QVW applications. In desktop, load them with all data and store them on disk uncompressed. This will give you an indication of how much RAM they'll be using. Add 10% for every user that opens the same model. The community has lots of information and documents on how to calculate memory requirements for a QlikView server setup.
Oh and BTW which version of QVS did you install? Some older releases experienced lockups due to memory exhaustion when users started doing specific things in the AccessPoint.