Server Power & Performance:
A Buyers Guide To Upgrading Your System

Date: January 14, 2004
Author: Jason Rabel

Conclusions:

I would have liked to do some more extensive testing, however like I said it was a production machine, and I had to minimize the down time. A few other useful tools for those interested are: Apache Bench, flood, and httperf (also here's a link to siege again).

From the results above, one can summarize that as the concurrent users increases, response time increases, transaction rate decreases, and concurrency increases. Ideally a web server would want to establish a connection, send the data, and close the connection to the client as quickly as possible. Determining just how much a web server can handle is a function of not only the system hardware and software, but also the data that is being sent. Since every requested page in our tests was generated on-the-fly from data stored in the database, more resources are needed from the server versus requesting just a static html file.

From the tests above, the extra RAM really becomes apparent when concurrent users increases, showing gains roughly from 20-25% in response times, and also similar gains in transactions per second. That's a lot of bang for the buck, especially when compared to other upgrades like faster processors, or even a whole new server.

Adding more physical memory to our server has eliminated the swapping to virtual memory (hard drive), and has helped keep response time quick, even under peak traffic conditions. Users expect and demand near-instant responses to their requests, they don't care about the complexities involved, they just want whatever it is they are requesting. Keeping performance levels acceptable can be difficult, especially with a tight budget. Doing a little investigative work beforehand can uncover that something as simple as adding more memory to a system is all that is required, instead of spending a lot more money on a whole new server.