Measurement Methodology

To understand how load affects response time, we measure latencies at various requests rates. Each server's maximum capacity is determined by having all clients issue requests in an infinite-demand (saturation) model, which is defined as load level of 1, and then relative rates are reported as load fractions relative to the infinite demand capacity of each server. This process simplifies comparison across servers, though it may bias toward servers with low capacity. Response time is measured by recording the wall-clock time between the client starting the HTTP request and receiving the last byte of the response. We normally report mean response time, but we note that it can hide the details of the latency profiles, especially under workloads with widely-varying request sizes. So, in addition to mean response time, we also present the $5^{th}$, $50^{th}$ (median) and $95^{th}$ percentiles of the latency distribution. Where appropriate, we also provide the cumulative distribution function (CDF) of the client-perceived latencies.

Figure 1: CDF of number of ready events (the return values from select()) in Flash