ICSA Colloquium - 15/12/2022

Talk Title

Memory-Centric System Architecture

Talk Abstract

Disaggregated Memory in the data center will solve problems of memory cost and capacity, but will introduce new challenges due to its remoteness and latency. New system foundations will be required for addressing these new challenges in the context of data-intensive applications such as analytics and deep learning and in anticipation of continued advances in semiconductor memory device technology. As "memoryness" spreads beyond RDMA in space through disaggregation and in time through persistent memory,  the need to rescue translation contexts from process abstraction has become paramount. Elephance co-founder Daniel Bittman addressed that problem in his work on Twizzler by placing a foreign object table in every memory object, thereby setting in motion for memory what S3 did for storage, namely a service architecture built around object abstractions. Elephance Memory Inc.'s MemOS(TM), a memory operating system specifically designed to run on disaggregated memory nodes, is a fork of Twizzler targeting the optimization of memory-intensive computation and communication on CPUs and GPUs alike. We will share our software approach at a high level and round up the talk with our current assumptions and projections governing the impact of memory disaggregation on microarchitecture and the memory hierarchy.

Bio

Pankaj Mehra is the President and CEO of Elephance Memory. He volunteers as Adjunct Associate Professor at UC Santa Cruz. Previously, he was VP of Storage Pathfinding and Product Planning in Samsung Electronics' Memory business. He was SVP and WW CTO of Fusion-io and was VP and Senior Fellow at SanDisk and Western Digital following the acquisition of Fusion-io. He was Distinguished Technologist, Chief Scientist of HP Labs Russia, and Founder of Taxonom.com during his 14 years at Tandem-Compaq-HP. Pankaj has previously founded startups IntelliFabric, Whodini, and AwarenaaS. He has held academic and research positions at NASA Ames, IIT Delhi, and IBM TJ Watson Research Center. Pankaj's 3 books and one hundred published papers and patents span processors, memory, interconnects, and storage; and, machine learning and context aware computing.