NFT @ 10312025: Transparent Multi-GPU Scheduling in the Polyglot World

#researchatnecstlab - #NECSTFridayTalk

On Friday, October 31, 2025, we had a new talk for the series #NECSTFridayTalk.

During this talk, we had, as speaker, Ian Di Dio Lavore, PhD at Dipartimento di Elettronica, Informazione e Bioingegneria.

In the following, you can find the details about the talk:
📌 Title: Transparent Multi-GPU Scheduling in the Polyglot World

📌 Abstract: GPUs are increasingly available in both cloud and personal computing environments, yet their integration with common programming languages such as JavaScript, Python, and Java still limits their widespread use for data processing acceleration. Exploiting multi-GPU systems remains complex due to intricate interconnect topologies and diverse inter-GPU communication patterns, often requiring advanced asynchronous programming skills.

We present a novel scheduler that enables transparent asynchronous execution across multiple GPUs without prior knowledge of program dependencies or system architecture. Integrated within the polyglot GraalVM ecosystem, it provides seamless access to multi-GPU acceleration from multiple high-level languages. Our benchmarks show that the scheduler achieves 80–90% of the performance of hand-optimized CUDA code on Volta and Ampere multi-GPU systems.

Finally, I will discuss our ongoing efforts to extend this work toward distributed multi-GPU environments, further broadening accessibility and scalability of GPU-accelerated computing.


#NECSTLab #Computerscience