Shenango: Achieving High CPU Efficiency for Latency-sensitive Datacenter Workloads
summary
Shenango builds a single busy-spinning IOKernel plus per-application runtimes (C++/Rust) and uses DPDK to steer packets and reallocate cores every 5 µs by checking ring-buffer head/tail pointers for queued packets and threads. It sustains memcached-style 1 µs service times up to about 5.5 million requests per second before the IOKernel saturates, matches ZygOS tail latencies while giving batch jobs reclaimed cycles, and is implemented in roughly 13k lines of code.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:25Trend #1: Faster Networks
- 0:57The Rise of Kernel Bypass
- 1:36Trend #2: Slowing of Moore's Law
- 2:28Load Variation Makes Efficiency Challenging
- 3:07The Need for Multiplexing
- 4:05Multiplexing with Existing Approaches
- 7:10Challenges of Fast Reallocations
- 8:10Shenango's Contributions
- 9:19Shenango's Design
- 11:06Compute Congestion
- 11:53Congestion Detection Algorithm
- 13:58Implementation
- 14:16Evaluation Questions
- 14:40Experimental Setup
- 15:28CPU Efficiency and Network Performance with Memcached • Memcached batch processing application
- 17:51Shenango is Resilient to Bursts in Load
- 19:15Conclusion
related talks
13mLive in the Express Lane
21mPreview: Operating Systems
25mDPDK performance optimization for edge computing scenario
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
50mHigh Performance Network Programming in Rust
15mThe nanoPU: A Nanosecond Network Stack for Datacenters
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking