MTS, Microsoft AI Superintelligence

Rengan
Xu

LLM infrastructure — model and systems optimization.

About

Research shaped by
systems thinking.

I am a Member of Technical Staff at Microsoft AI Superintelligence (MSI). Previously, I was a Research Scientist at Meta SuperIntelligence Labs. I focus on scaling reinforcement learning post-training for high throughput—spanning attention, MoE, grouped GEMM, parallelism strategies, and E2E system efficiency.

Education

Ph.D., Computer Science
University of Houston, 2016
Advisor: Dr. Barbara Chapman

B.S., Computer Science
Hefei University of Technology, 2009

Research focus

LLM

Reinforcement Learning (RL)

Inference, Post Training

GPU Kernels

Publications

Selected work

Research across ads delivery, learned representations, accelerator programming, and high-performance systems.

2026
In-Kernel Broadcast Optimization: Co-Designing Kernels for RecSys Inference
PyTorch Blog · May 5, 2026
Jian Jiao, Boda Li, Hongtao Yu, Yuanwei (Kevin) Fang, Zhengkai Zhang, Zhuoran Zhao, Yuxin Chen, Sijia Chen, Yang Chen, Zijian Shen, Shuyao Bi, Ao Cai, Junhan Hu, Shuqi Yang, Wei Wei, Lu Fang, Rengan Xu, et al.
Article ↗
2024
Enhancing Performance and Scalability of Large-Scale Recommendation Systems with Jagged Flash Attention
arXiv:2409.15373 · September 2024
Rengan Xu et al.
PDF ↗
2024
Async Learned User Embeddings for Ads Delivery Optimization
arXiv:2406.05898 · SIGIR 2024 Workshop on Multimodal Representation and Retrieval
Mingwei Tang, Meng Liu, Hong Li, Junjie Yang, Chenglin Wei, Boyang Li, Dai Li, Rengan Xu, et al.
PDF ↗
2017
Implementing the OpenACC Data Model
IPDPSW 2017 · IEEE, pp. 662–672
Michael Wolfe, Seyong Lee, Jungwon Kim, Xiaonan Tian, Rengan Xu, Sunita Chandrasekaran, Barbara Chapman
PDF ↗
2015
Compiler Transformation of Nested Loops for General Purpose GPUs
Concurrency and Computation: Practice and Experience · Wiley
Xiaonan Tian, Rengan Xu, Yonghong Yan, Sunita Chandrasekaran, Deepak Eachempati, Barbara Chapman
PDF ↗
2014
SPEC ACCEL: A Standard Application Suite for Measuring Hardware Accelerator Performance
2014
Guido Juckeland et al., including Rengan Xu
PDF ↗

2019
Densifying Assumed-Sparse Tensors
ISC High Performance · Springer, pp. 23–39
Derya Cavdar, Valeriu Codreanu, Can Karakus, John A. Lockman, Damian Podareanu, et al.
PDF ↗
2016
An Analytical Model-Based Auto-Tuning Framework for Locality-Aware Loop Scheduling
International Conference on High Performance Computing · Springer, pp. 3–20
Rengan Xu, Sunita Chandrasekaran, Xiaonan Tian, Barbara Chapman
PDF ↗
2015
Multi-GPU Support on Single Node Using Directive-Based Programming Model
Scientific Programming · Volume 2015
Rengan Xu, Xiaonan Tian, Sunita Chandrasekaran, Barbara Chapman
PDF ↗
2014
NAS Parallel Benchmarks for GPGPUs Using a Directive-Based Programming Model
LCPC 2014 · Springer, pp. 67–81
Rengan Xu, Xiaonan Tian, Sunita Chandrasekaran, Yonghong Yan, Barbara M. Chapman
PDF ↗
2014
Reduction Operations in Parallel Loops for GPGPUs
PMAM ’14 · ACM
Rengan Xu, Xiaonan Tian, Yonghong Yan, Sunita Chandrasekaran, Barbara Chapman
PDF ↗
2018
The OpenACC Data Model: Preliminary Study on Its Major Challenges and Implementations
Parallel Computing · Volume 78, pp. 15–27
Michael Wolfe, Seyong Lee, Jungwon Kim, Xiaonan Tian, Rengan Xu, Sunita Chandrasekaran, Barbara Chapman
PDF ↗
2016
ACC-SVM: Accelerating SVM on GPUs Using OpenACC
First Workshop of Mission-Critical Big Data Analytics
Rengan Xu, Dounia Khaldi, Abid M. Malik, Barbara Chapman
PDF ↗
2016
Optimizing GPU Register Usage: Extensions to OpenACC and Compiler Optimizations
ICPP 2016 · IEEE, pp. 572–581
Xiaonan Tian, Dounia Khaldi, Deepak Eachempati, Rengan Xu, Barbara Chapman
PDF ↗
2014
Accelerating Kirchhoff Migration on GPU Using Directives
WACCPD ’14 · pp. 37–46
Rengan Xu, Maxime R. Hugues, Henri Calandra, Sunita Chandrasekaran, Barbara M. Chapman
PDF ↗
2014
A Validation Testsuite for OpenACC 1.0
IPDPSW 2014 · IEEE, pp. 1407–1416
Cheng Wang, Rengan Xu, Sunita Chandrasekaran, Barbara Chapman, Oscar Hernandez
PDF ↗
2014
Compiling a High-Level Directive-Based Programming Model for GPGPUs
LCPC 2013 · Springer, pp. 105–120
Xiaonan Tian, Rengan Xu, Yonghong Yan, Zhifeng Yun, Sunita Chandrasekaran, Barbara Chapman
PDF ↗
2014
OpenUH — An Open Source OpenACC Compiler
Poster · GTC 2014
Rengan Xu, Xiaonan Tian, Yonghong Yan, Zhifeng Yun, Sunita Chandrasekaran, Barbara Chapman
2013
Exploring Programming Multi-GPUs Using OpenMP and OpenACC-Based Hybrid Model
IPDPSW 2013 · IEEE, pp. 1169–1176
Rengan Xu, Sunita Chandrasekaran, Barbara Chapman
2013
Filesystem Aware Scalable I/O Framework for Data-Intensive Parallel Applications
IPDPSW 2013 · IEEE, pp. 2007–2014
Rengan Xu, Mauricio Araya-Polo, Barbara M. Chapman
2013
OpenACC Programming Experiences Using Scientific Applications
Poster · GTC 2013
Rengan Xu, Cheng Wang, Sunita Chandrasekaran, Barbara Chapman
PDF ↗
2012
Directive-Based Programming Models for Scientific Applications — A Comparison
SC Companion 2012 · IEEE
Rengan Xu, Sunita Chandrasekaran, Barbara Chapman, Christoph F. Eick
PDF ↗
2012
Parallel I/O Framework for Data-Intensive Parallel Applications
Poster · PDSW ’12
Rengan Xu, Mauricio Araya-Polo, Barbara Chapman
PDF ↗

Contact

Let’s connect.

For research conversations, collaborations, or professional inquiries, reach me by email or find my work and profile online.

renganxu@gmail.com GitHub ↗ LinkedIn ↗ Google Scholar ↗

Research shaped bysystems thinking.

Education

Research focus

Selected work

In-Kernel Broadcast Optimization: Co-Designing Kernels for RecSys Inference

Enhancing Performance and Scalability of Large-Scale Recommendation Systems with Jagged Flash Attention

Async Learned User Embeddings for Ads Delivery Optimization

Implementing the OpenACC Data Model

Compiler Transformation of Nested Loops for General Purpose GPUs

SPEC ACCEL: A Standard Application Suite for Measuring Hardware Accelerator Performance

Densifying Assumed-Sparse Tensors

An Analytical Model-Based Auto-Tuning Framework for Locality-Aware Loop Scheduling

Multi-GPU Support on Single Node Using Directive-Based Programming Model

NAS Parallel Benchmarks for GPGPUs Using a Directive-Based Programming Model

Reduction Operations in Parallel Loops for GPGPUs

The OpenACC Data Model: Preliminary Study on Its Major Challenges and Implementations

ACC-SVM: Accelerating SVM on GPUs Using OpenACC

Optimizing GPU Register Usage: Extensions to OpenACC and Compiler Optimizations

Accelerating Kirchhoff Migration on GPU Using Directives

A Validation Testsuite for OpenACC 1.0

Compiling a High-Level Directive-Based Programming Model for GPGPUs

OpenUH — An Open Source OpenACC Compiler

Exploring Programming Multi-GPUs Using OpenMP and OpenACC-Based Hybrid Model

Filesystem Aware Scalable I/O Framework for Data-Intensive Parallel Applications

OpenACC Programming Experiences Using Scientific Applications

Directive-Based Programming Models for Scientific Applications — A Comparison

Parallel I/O Framework for Data-Intensive Parallel Applications

Let’s connect.

Research shaped by
systems thinking.