ICSA Colloquium Talk - 28/09/2023

Title: Persistent Memory meets GPU computing

Abstract: Extracting knowledge from massive amounts of data needs massively parallel computing and access to large amounts of memory. A wide range of application domains, including deep learning, graph analytics, and scientific computing, rely on massive parallelism of Graphic Processing Units (GPUs) for their primary computing needs. On the other hand, the new Persistent Memory (PM) technologies enabled by the likes of Intel Optane DC memory, Samsung’s memory-semantic SSD, can bridge both the capacity and performance gap between traditional DRAM memory and the storage, besides being non-volatile. However, persistent memory has remained off-limits for GPU. In this talk, we will discuss a new system called GPM or GPU with Persistent Memory which enables GPU programs to perform loads and stores to data stored on PM. We demonstrate several use cases, including GPU-accelerated databases and checkpointing of intermediate results during DNN training that benefit from both GPU’s massively parallel computing and PM’s fast, byte-grain persistence.  We will detail the key engineering challenges in designing GPM using commercially available hardware. Finally, we will argue that new hardware support is necessary for GPUs to fully harness the potential of persistence memory.

Bio: Arka (prava) is an associate professor at the Indian Institute of Science (IISc), Bengaluru. He is also a technical consultant to AMD's Server Performance Group. His research often straddles the boundary between computer architecture and system software. Arka’s recent research often focuses on GPU computing. He also works on improving virtual memory management in the hardware and the operating system. Arka received Intel Rising Star Faculty Award in 2021 and the Google India Research Award in 2022. He holds a Ph.D. in Computer Science from the University of Wisconsin-Madison.