Efficient Scheduling Policies for Microsecond-Scale Tasks
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06
summary
Simulations model 1μs tasks with measured overheads (100 ns to move a task, 5 μs to add a core) and show work stealing consistently yields the best latency/efficiency tradeoff; static allocations sometimes beat reactive reallocation for very short tasks. Two new core-allocation policies, delay range and utilization range, proactively revoke cores and, when implemented on Caladan, produce up to 13–22% higher CPU efficiency without hurting median or tail latency.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:29Datacenter Goals
- 1:07Multiplex to Achieve Both Goals
- 1:56Multiplexing Systems
- 2:10Lingering Inefficiency
- 3:39Policy and Mechanism
- 4:37Challenge
- 6:04Our Approach
- 6:35Key Findings
- 7:04System Model
- 7:37Overheads
- 7:59Load Balancing Policies
- 9:06Which Load Balancing Policy is Best?
- 9:58Core Allocation Policies
- 11:35Load Balancing for Non-static Allocatio
- 12:04Which Allocation Policy is Best?
- 13:01Experimental Results
- 13:51Takeaways
- 14:20Summary