Abstract

We present Resourceful, a system that enables applications to obtain a fine-grained description of how certain activities they perform (e.g processing a user request, serving a database query or performing a background task) consume system resources (as managed by the OS kernel).

Understanding how the behaviour of the kernel affects application performance is hard. Existing tools only give coarse grained aggregate performance measurements, which are difficult to relate back to application behaviour. Resourceful allows engineers to understand the kernel in relation to their individual application semantics, meaning they can answer questions such as, ‘what resources does the kernel consume and where, when I process this user request?’.

In order to be meaningful for application optimization, scheduling systems or for devops teams, the resource consumption measurements are broken down per kernel subsystem: one can determine (amongst others) how much time was spent in each subsystem (how much time has this user request spent in the network or vfs subsystem? how much CPU , memory, how many cache misses happened for this particular request?).

We show how such measurements can be then compared across activities (in our example, comparing resources consumed by various requests) to explain the causes of performance variations or to determine workloads that interract badly: Why was this request slower than others? What other applications contended for the same resources during a particular request?