Scheduling on Multicores

EDIC candidacy exam
Exam president: Prof. Martin Odersky
Thesis advisor: Prof. Rachid Guerraoui
Co-examiner: Prof. Willy Zwaenepoel

Abstract
In this proposal, we survey recent work which shows
that in non-uniform memory access machines the placement of
software threads to hardware cores can have a significant effect
on the performance of concurrent applications. The surveyed
work inspired us to devise a technique that utilizes machine
learning and performance counters in order to select between
different thread-placement policies. We propose Slate, an implementation
of this technique that is able to select and apply
thread-placement policies online.

Background papers
Abstracting Multi-Core Topologies with MCTOP (EuroSys '17)
Pandia: comprehensive contention-sensitivethread placement (EuroSys '17)
The Linux Scheduler: a Decade of Wasted Cores (EuroSys '16)