Despite the performance of today's and tomorrow's computers there will always be workloads which can take advantage of more and more of it. One way to push the performance with the same hardware is to reduce the amount of overhead while executing the program.
Linux specifically has ways to customize the programming environment to turn off unneeded features and guide the scheduler and resource allocator to optimize the program execution. We can do more so and for that we only have to look back to the early days of computing with machines dedicated to running single applications. Of course we do not want to incur the same cost but with unikernels we don't have to.
This talk will present some of the findings in the quest to reduce system noise (latency) along all these axes which comes from common practice around Linux, the most commonly used OS in HPC, and also results from ongoing research conducted with Boston University.
Interest and basic knowledge of compute is sufficient. Unless one is interested in really high performance computing one would not use any of the techniques, though.
First the talk will go into some of the details of inefficiency resulting from execution of programs in modern systems. This is contrasted with compute in the early days of computing. Solutions are presented which range from OS configuration, over replacing some of it, to making deep changes of the entire compute environment.
// Ulrich Drepper
works for Red Hat in the office of the CTO on technology on emerging technologies – mostly in the areas of high-performance, scalable, and efficient compute. The projects cover wide areas of technologies from infrastructure software such as compilers, runtimes, and OS kernels to hardware designs for processors and systems.