Load Balancing in L4Re

BBB-Link: https://bbb.tu-dresden.de/b/mat-xin-oyh-xzn

Presentation Language: English

Modern hardware increasingly combines processor cores of varying types and speeds on a single chip.
However, the L4Re operating system framework does not provide CPU load balancing.
This thesis proposes a user-level load balancing service for L4Re on x86-64 that distributes threads across physical cores during runtime.
It respects application-supplied constraints, such as thread pinning, core-sharing restrictions, and the placement requirements of virtual machine vCPUs.
AMP awareness is achieved by scaling the load of each CPU according to its measured compute performance.
The resulting load balancer is evaluated against the Linux EEVDF scheduler using ported CPU-bound benchmarks.