Operating System Support for Large Memory Server
With the advent of in-memory computing, such as data analytics, graph processing, and machine learning, the demand for high-density memory has steadily increased in recent years. Due to the challenge of scaling DRAM density in both CPU and GPU worlds, however, new hardware features are introduced. I advocate that system researchers and engineers should take advantage of such new facilities to build large memory servers. In this talk, I will first present why modern Linux operating systems are insufficient for new memory hierarchies, such as tiered or heterogeneous memory systems, and then introduce our memory management techniques to unlock the potential of the multi-tiered memory hierarchy. Second, to overcome the memory capacity of GPUs, I will introduce a new approach providing an illusion of virtual memory space for GPUs comprised of local GPU, spare memory of neighbor GPUs, and host memory. I will also present how our proposed memory manager is implemented in the GPU driver to support diverse GPU workloads without application modification. Finally, I will briefly introduce ongoing research efforts to build fast and deep learning serving systems for multi-GPU servers.
Jeongseob Ahn is currently an Associate Professor in the Department of Software and Computer Engineering at Ajou University and is also affiliated with the Department of Artificial Intelligence. His research interests lie in building high-performance computer systems with the interactions between architecture and system software. His works have been published in top-tier systems as well as architecture conferences such as ISCA, MICRO, EuroSys, and ATC. Before joining Ajou, he worked at Oracle Labs in Silicon Valley and spent a year at the University of Michigan. He received his Ph.D. in Computer Science from KAIST in 2015.