SPECmarks are leading us astray

It is suggested that success at designing systems to satisfy latency-sensitive requirements has not yet been achieved, and that it is necessary to develop benchmarks that explicitly measure system latency, rather than measuring it indirectly through its effect on throughput. If both throughput-sensitive and latency-sensitive benchmarks are required, the result might be the creation of systems that are easily tuned (or even self-tuned) to provide either good latency or good throughput, as the case demands.<
>