LinuxCon + CloudOpen + ELC-E Europe 2014

A lot of silicon in modern SoCs is doing nothing most of the time, being dedicated to so called "tracing". This used to be domain of well known, expensive and proprietary tools, but in the last year there has been more and more interest in providing kernel infrastructure driving it, however this is *not* focus of this presentation :-) Instead, I will show one of possible ways of consuming the vast amount of data obtained from the mentioned resources, mainly on processor trace. This data usually provides detailed data about instructions being executed with minimal (in some cases even no) performance overhead. After being decoded, it allows generation of very precise (not just sample based) profile of executed code. I have been working on using the user space perf tool to use this data and want to present results, discussing possibilities and limitations of the method.