Processing in Storage Class Memory
Talkat 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage '20) · 2020-07-13..07-14
summary
Position: processing-in-memory becomes integral to storage-class memory to cut memory-bandwidth and CPU usage by running simple, parallel tasks inside SCM. Implementation uses UPMEM DDR4 DIMMs with one DPU per 64 MB, 64 KB private SRAM and DMA-driven loads into an interleaved bank layout. Measured raw move of 36 GB in 0.16 s (~250 GB/s) across 576 DPUs (9 ranks × 64 DPUs); snappy decompression shows up to 4.5× speedup versus host CPU for large files, with smaller files losing benefit. Proposals: data-triggered functions, concurrent host/DPU access, mixed memory.
AI-written summary · about