# 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](https://www.usenix.org/conference/nsdi23/presentation/thorpe)] \[[Code](https://github.com/uclasystem/bamboo)]
  * UCLA & CMU & MSR & Princeton
  * Resilient distributed training

### Model Serving

* Shepherd: Serving DNNs in the wild \[[Paper](https://www.usenix.org/conference/nsdi23/presentation/zhang-hong)] \[[Personal Notes](https://paper.lingyunyang.com/reading-notes/conference/nsdi-2023/shepherd)]
  * 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](https://paper.lingyunyang.com/reading-notes/conference/nsdi-2023/husky)] \[[Paper](https://www.usenix.org/conference/nsdi23/presentation/kong)] \[[Benchmark Suite](https://github.com/host-bench/husky)]
  * Duke & Microsoft & SJTU
  * Develop a *test suite* to *evaluate* RDMA performance isolation solutions.