SHRiNKing Web Server Farms: A Method for Scaleable Performance Prediction and Measurement

The increasing size of web-server farms and the sheer volume of HTTP requests, makes hard to collect performance measurements and monitor the state of a farm in real time. Further, it increases the cost of a bad algorithmic or architectural decision, while predicting the performance of new algorithms and architectures is also hard. We propose a way to side-step these problems, by intelligently combining small-scale experiments and analysis. Our hypothesis is this: if we take a sample of the incoming requests, and feed it into a suitably scaled version of the web farm, we can extrapolate from the performance of the scaled system to that of the original. Our main nding is that when we suitably scale a web-server farm, then performance measures such as mean response time and throughput are left virtually unchanged. We show this using experiments and simple analysis.