Theme 2: Distributed Evolvable Memory and Storage

The computing infrastructure will include highly-heterogeneous distributed memory and storage resources. As workloads relentlessly increase their data needs, the memory reachable by processors as local memory will expand across an entire rack–creating a formidable memory wall that we will meet with novel processor structures and gracefully-degrading coherence mechanisms. To utilize heterogeneous memory and storage assets efficiently, we will develop new abstractions that allow applications to select the type of asset needed. Moreover, we will develop theory-grounded scalable algorithms to apportion these assets efficiently among thousands of competing applications in the datacenter and billions of allocation requests. Ubiquitous intelligent memory and storage blocks distributed across the memory hierarchy will be harnessed to operate in a coordinated manner.

Heterogeneous Intelligent Memory and Storage (IMS) blocks present in multiple locations of the memory hierarchy of a distributed machine.
Heterogeneous Intelligent Memory and Storage (IMS) blocks present in multiple locations of the memory hierarchy of a distributed machine (Courtesy of Steven Swanson).

Papers and Presentations:

Johnson Umeike, Neel Patel, Alex Manley, Amin Mamandipoor, Heechul Yun, Mohammad Alian
IPASS 2023