LinuxCon + CloudOpen + ELC-E Europe 2014

System administrators and developers are challenged by understanding the elapsed time of heterogeneous and distributed programs in order to improve response time. Profilers are useful to identify hotspot code in an application, but they do not take the wait time into account. A task can block for a device, a local task using any inter-process communication or a task running on another computer. We present a technique using solely kernel tracing to recover the active path of a computation, crossing machines barrier. The technique uses scheduling, interrupt and network events, but does not require system call tracing nor user space instrumentation. Therefore, this technique works for any programming language, and includes interactions with kernel threads. We show the analysis result of actual use-cases, including shell scripts, RPC request, file server and web app.