PD3: Prefetching Data with DPUs for Disaggregated Memory
summary
PD3 implements a DPU-side prefetch path that parses client requests, looks up a bucketized host view, issues RDMA pulls to remote memory, and DMA-loads items into compute-local RAM before the execution engine accesses them. The DPU lookup reaches 70 million ops throughput (13× baseline) and the end-to-end system matches local-memory Garnet performance using only 128 MB of local cache (under 1% of the working set), eliminating remote request/transfer latency.
AI-written summary · about
chapters (11)
- 0:00PD3 and team intro
- 1:43Memory disaggregation challenges
- 3:09Cache misses bottleneck
- 4:53DPUs for prefetching data
- 5:36Prefetching with host view
- 7:19Host view optimization
- 8:44False positives in prefetching
- 9:48Data transfer engine
- 11:12Loading zone for prefetch
- 12:16Evaluation and performance
- 13:40Summary
related talks
19mDShuffle: DPU-Optimized Shuffle Framework for Large-scale Data Processing
19mCluster Optimization: Harnessing DPUs for Ambient Mesh Workloads
39mFunction Dedicated Network Deployment Model in DPU/IPU-Like Device Acro...
17mDuhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks
11mWhen Cloud Storage Meets RDMA
19mEnabling Big Memory workloads to run
16mSolidigm CSAL Solution Brings Advanced IO Shaping, Caching and Data Placement Into NVIDIA DPU DOCA S
18mArchitecting Multi-tenant Data-center Networks for GPU Customers