Overview
Explore efficient scheduling policies for microsecond-scale tasks in this 15-minute conference talk from NSDI '22. Delve into the challenges datacenter operators face in supporting microsecond-latency applications while maximizing CPU resource utilization. Learn about various systems that allow multiple applications to run on the same server, and discover why many struggle to balance latency and efficiency, especially with tasks as short as 1μs. Examine different policy options for allocating cores across applications and load-balancing tasks across cores within an application through simulation comparisons. Gain insights into the surprising effectiveness of static core allocations for small tasks and the superior performance of work stealing among load-balancing policies. Understand how implementing the best-performing policy choices can yield significant efficiency improvements without compromising latency. Follow the researchers' journey as they build upon the Caladan core-allocating system to demonstrate these improvements in practice.
Syllabus
Intro
Datacenter Goals
Multiplex to Achieve Both Goals
Multiplexing Systems
Lingering Inefficiency
Policy and Mechanism
Challenge
Our Approach
Key Findings
System Model
Overheads
Load Balancing Policies
Which Load Balancing Policy is Best?
Core Allocation Policies
Load Balancing for Non-static Allocatio
Which Allocation Policy is Best?
Experimental Results
Takeaways
Summary
Taught by
USENIX