NSDI 2023

Meta Info

Homepage: https://www.usenix.org/conference/nsdi23

Paper list: https://www.usenix.org/conference/nsdi23/technical-sessions

Accepted Papers

• Spring: https://www.usenix.org/conference/nsdi23/spring-accepted-papers

• Fall: https://www.usenix.org/conference/nsdi23/fall-accepted-papers

Papers

Large Language Model (LLM)

• Bamboo: Making Preemptible Instances Resilient for Affordable Training of Large DNNs [Paper] [Code]

  • UCLA & CMU & MSR & Princeton

  • Resilient distributed training

Model Serving

• Shepherd: Serving DNNs in the wild [Paper] [Personal Notes]

  • UWaterloo & Yale & UC Berkeley

  • Handle the short-term workload unpredictability.

  • Aggregate request streams into moderately-sized groups; leverage preemption and model-specific batching.

RDMA

• Understanding RDMA Microarchitecture Resources for Performance Isolation [Personal Notes] [Paper] [Benchmark Suite]

  • Duke & Microsoft & SJTU

  • Develop a test suite to evaluate RDMA performance isolation solutions.