Nu: Achieving Microsecond-Scale Resource Fungibility with Logical Processes
summary
A system called Nu implements logical processes that split a Unix process into many small stateful proclets and a centralized controller, and it migrates those proclets across machines to reclaim resources. A 32-server prototype over 100G Ethernet runs a social network, a MapReduce (Phoenix) job, and a KVS and shows proclet migration in ≈100 μs (40 μs for <256 KB, up to 11 GB/s for large heaps), keeps 99.9th‑percentile latency near or below 1 ms during memory pressure (disruption lasting 0.54–2.1 s) and limits throughput loss to <6% for 24–37 ms.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:33Operational reality of today's datacenter
- 4:25Key Idea: fine-grained decomposition
- 5:38Challenges and design overview
- 6:35The logical process abstraction
- 7:22Proclet communication
- 8:04Centralized controller
- 8:29Proclet migration
- 10:07Asynchronous Call
- 11:04Computation shipping
- 12:15Able to achieve high utilization w/o disruption?
- 15:08How fast can we migrate proclets?
related talks
17mQuicksand: Harnessing Stranded Datacenter Resources with Granular Computing
10mLightning Talk: Millisecond Scale-to-Zero and the Quest to Never Pay for Idle Again
16mPreview: Cloud Computing
19mE3: Energy-Efficient Microservices on SmartNIC-Accelerated Servers
25mGreenOps at Scale: The 50TB Memory Diet for a Petabyte Platform
25mSustainability at the Core with Cloud Native Processors
25mRunning Unikernels in Prod for a Few Years
28mEmpowering the Next Generation of High-Performance, Power-Efficient Servers