Futexes are complex enough that make it a good candidate for applying fuzzy testing techniques to the various inputs and operations that can be performed. Instead of re-writing any testing infrastructure, make use of the trinity system call fuzzing tool. While trinity does apply fuzzing to futexes, it does so with very limited knowledge, using only generic callbacks to the different steps involved. As such the main goal is extending testing coverage and making it more ad-hoc.
While there are plenty of benchmarks that compare different IO multiplexing techniques such as epoll vs poll/select, there's really nothing out there that particularly measures epoll system call latencies under different scenarios. Design and implement a series of performance benchmarks for this call under the 'perf bench' framework.
The state of the art of sysv/posix ipc benchmarking is a combination of ad-hoc programs scattered over the internet. While some mechanisms, such as sysv semaphores, have a lot of coverage, others really lack (message queues), and some are simply non-existent; albeit some of the legacy flavors we aren't too concerned with, other than them being functionally correct.
Consolidate all this by taking the good benchmarks and implement new ones where needed, such as message queues, everything around the 'perf bench' framework. This will both improve the lives of us who hack on kenrel ipc as well as improve coverage overall. Eventually mmtests should be updated as well.