packing, spreading and scheduling latency

This proposal has been accepted as a session.

*

One Line Summary

When packing tasks can improve power efficiency and scheduling latency

Abstract

Spreading tasks on CPUs is the default policy to improve performance by minimizing resources sharing and scheduling latency but the final behavior doesn’t always follow the rule. This talk will describe how some information like the power domain topology or the wakeup latency can be used in the scheduler to improve power efficiency and scheduling latency. Based on the results of an implementation, the talk will explain when it’s worth to pack tasks but also why it can sometimes deserve the power efficiency of the system.

Presentation Materials

slides

Speakers

Leave a private comment to organizers about this proposal