Developer Conference 2015

perf has become the defacto standard for linux performance analysis. Often the most basic commands are just enough to pinpoint bottlenecks and get a feeling what is going on. But perf is much more powerful...

In this talk we will explore what other features perf provides: Besides simply checking where the CPU spends most of its cycles, one can easily add custom tracepoints and capture arguments or return values from functions. We will do so on examples in the networking stack. This will be useful to isolate bugs or even validate patches.

In an outlook the talk will introduce the audience what performance counters are available on today's CPUs but not easily accessible in perf. A short introduction will be given on how the register and masks are to be interpreted from the CPU manuals and how they can be used from perf. As examples we will analyse the MOESI protocol and memory bus traffic between different NUMA nodes.