vPIM: Efficient Virtual Address Translation for Scalable Processing in-Memory Architectures
DescriptionProcessing in-memory (PIM) using 3D stacked memory requires scalability and programmability. Scalability can be achieved by connecting PIM stacks over a network. For programmability, PIM needs virtual memory. Radix page table has high translation cost due to serialized cross-stack memory accesses; Cuckoo-hash-based page tables introduce excessive accesses across memory stacks. We propose an address translation scheme for multi-stack PIM systems which integrates network-contention-aware hashing design into Cuckoo hash to reduce high-cost cross-stack accesses. On top, we propose pre-translation that repurposes some cores to pre-translate virtual addresses. We achieve speedup of 4.4x and 1.7x compared to radix and cuckoo hash table.
TimeThursday, July 13th1:40pm - 1:55pm PDT
Location3004, 3rd Floor
In-memory and Near-memory Computing Architectures, Applications and Systems