Effective Load Balancing Using Processor Performance States

Title: Processor performance states aware dynamic load balancing Suresh B Siddha

Traditionally, Operating system scheduler assumes that each processor in a multiprocessor system environment has similar functionality and performance. Load balancing part of the scheduler is highly dependent on this assumption. With the advancement of new processor and platform technology as well as the need for better power managed processors, this constant performance assumption in a multiprocessor environment may not be true going forward. Today, there are processors that provide ability to switch the processor frequency at run time. Operating system techniques like ondemand governor, use this fast frequency change feature, and do power management by aggressively changing the frequency depending on CPU utilization. This may result in processors in a multiprocessor environment run with different performance characteristics.

This paper describes the scheduler challenges with special focus on load balancing due to the performance asymmetries. This paper will also describe how these challenges can be addressed by incorporating the processor performance state information into the Linux Kernel Domain Scheduler. Finally, it will provide the power performance data with this new prototype scheduler.