Computer Science Seminar by Maxime Gonthier: Scheduling Under Memory Constraint for Runtime Systems

Speaker: Maxime Gonthier, post doctoral researcher, University of Chicago

Title: Scheduling Under Memory Constraint for Runtime Systems

Abstract: Hardware accelerators like GPUs now provide a large part of the  computational power used for scientific simulations.

Despite their efficacy, GPUs possess limited memory and are connected to the main memory of   the machine via a bandwidth limited bus.

Scientific  simulations often operate on very large data, that surpasses the GPU's memory capacity.

Therefore, one has to turn to out-of-core computing: data are kept in a remote, slower memory (CPU memory), and  moved back and forth from/to the device memory (GPU memory), a process also  present for multicore CPUs with limited memory.

In both cases, data movement quickly becomes a performance  bottleneck.

Task-based runtime schedulers have emerged as a convenient and  efficient way to manage large applications on such heterogeneous  platforms. 

In this presentation, we propose a scheduler for task-based runtimes that improves data locality in an out-of-core  setting, to reduce data movements.  We compare this scheduler to existing schedulers in runtime systems and using StarPU, we show that our new scheduling strategy achieves comparable performance when memory is not a constraint, and significantly better performance when application input data exceeds memory, on both GPUs and CPU cores.

Speaker Bio: Maxime Gonthier received his PhD from the École Normale Supérieure de Lyon. He is now working as a PostDoc at the University of Chicago. His research revolves around data locality and task scheduling algorithms under memory constraints on the StarPU runtime. He is also working on resilient algorithmic solutions for storage systems and energy-aware solutions for HPC.