Ruichuan Chen

DMTS @BellLabs, previously Postdoc @mpi_sws_ & PhD @PKU1898.

[Ruichuan Chen @ruichuan]

Coupled with efficient memory management across GPU and host memory, FaaScale handles bursty LLM inference workloads effectively, achieving up to 5× lower tail time-to-first-token latency and 31.3% cost reduction on real-world LLM traces.

[Ruichuan Chen @ruichuan]

Our FaaScale accepted to #MLSys'26.

FaaScale is a #serverless #LLM system that enables fast and resource-efficient model scaling. Its key idea is "pipelined multicast inference" that synergizes multicast with dynamic, cross-node pipeline-parallel execution during model transfer.

[Ruichuan Chen @ruichuan]

FaaScale implements this idea through a model scaling scheme that adaptively multicasts model blocks and dynamically forms inference pipelines on the fly.

[Ruichuan Chen @ruichuan]

Our paper "Torpor: GPU-Enabled Serverless Computing for Low-Latency, Resource-Efficient Inference" has been accepted to @usenix ATC '25. The better part is that it's been deployed and tested in real commercial #serverless cloud. Stay tuned!

[Ruichuan Chen @ruichuan]

Berlin Marathon, my first World Marathon Major!
42.195km, finish time 3:48:34
#BerlinMarathon #BerlinMarathon2024

[MLCommons @MLCommons]

We are excited to announce that applications for the 2024 MLCommons ML and Systems Rising Stars Program are now open! Early-to-late-stage PhD students and recent Ph.D. grads have until March 22nd to register. Apply today:

https://t.co/PYOeeSZIZK

[Ruichuan Chen @ruichuan]

A paper accepted to #UsenixSecurity '24.

This is a systems work with rigorous proofs. It shows, in a #FederatedLearning setting, how one can still select clients for participation securely, fairly and efficiently even if the server is adversarial.

[Ruichuan Chen @ruichuan]

Skiing down black pistes with a 70% gradient from three peaks at #serfausfissladis

[Ruichuan Chen @ruichuan]

Meeting a turtle while scuba diving at 25m deep 🤿 #scubadiving #hurghada #RedSea

[Ruichuan Chen @ruichuan]

Now, I'm officially a full Marathon runner.
Frankfurt Marathon, 42.195km, finish time: 3h53:33.
#runtheskyline #frankfurtmarathon

[Ruichuan Chen @ruichuan]

The second work designs a fully automated verification framework for understanding fine-grained #consistency semantics in realistic large-scale web applications with no user input, efficiently and correctly.

[Ruichuan Chen @ruichuan]

Two papers accepted to @EuroSys_conf #eurosys24. More details follow.

[Ruichuan Chen @ruichuan]

@EuroSys_conf The first work designs a distributed #differential_privacy #federated_learning framework that is highly efficient and resilient to client dropout. It handles client dropout in realistic FL scenarios, achieving optimal privacy-utility tradeoff and accelerating training by 2.4x.

[Ruichuan Chen @ruichuan]

@pramod_bhatotia Happy birthday!

[Ruichuan Chen @ruichuan]

@pramod_bhatotia Awesome, we should find a time to go skiing together;)

[Ruichuan Chen @ruichuan]

@marius https://t.co/LETzwoVWdk

[Ruichuan Chen @ruichuan]

Our Pheromone work accepted to @usenix NSDI'23. It proposes a data-centric approach to design #serverless platform that allows fine-grained data exchange between functions of a workflow, while providing a unified and efficient approach for function invocation and data exchange.

[Ruichuan Chen @ruichuan]

The nsdi'23 camera-ready of this "data-centric #serverless" work is now online: https://t.co/LETzwoVonM

The system is also open-sourced: https://t.co/K5Zbt4y4nO

Comments welcome!

[Ruichuan Chen @ruichuan]

Compared to well-established commercial and open-source platforms, Pheromone cuts the latencies of function interactions and data exchanges by orders of magnitude, scales to large workflows, and enables easy implementation of complex applications.

[Ruichuan Chen @ruichuan]

@balakc @usenix Thanks, will post a link here once the camera-ready is ready.

[Ruichuan Chen @ruichuan]

@pramod_bhatotia @usenix Thanks, Pramod!

[Ruichuan Chen @ruichuan]

@realhamed @usenix Thanks, Hamed!