DPA-Store: An Ordered Network Data Path Key-Value Store
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
DPA-Store runs a lock-free learned index inside the BlueField-3 DPA memory, fetches values via host DMA to a host-side tree replica, and uses batched writes plus single-atomic subtree stitch to keep traversals consistent. It sustains 33 million ops/s for GETs and 13 MOPS for ranges with a per-access DPA memory latency ~465 ns, a hot-pair NIC cache boosting throughput, but raw inserts stall at about 1.7 million inserts/s due to host↔DPA memory limits.
AI-written summary · about
chapters (12)
- 0:00DPA-Store overview
- 1:46DPA-Store architecture
- 3:11DPA subsystem challenges
- 4:43Learned index design
- 5:48Architecture details
- 7:57Memory management
- 9:02Cache line optimization
- 10:35Performance evaluation
- 12:04Comparison with B-tree
- 13:29Insert performance limits
- 14:13Comparison with RoLEX
- 14:56Future work and conclusion
related talks
48mRebuilding a RT Datastore: the design, deployment & performance story
15mImproving Performance of Flash Based Key-Value Stores Using Storage Class Memory...
25mDisaggregating Persistent Memory and Controlling Them Remotely
22mPinK: High-speed In-storage Key-value Store with Bounded Tails
17mPreview: Persistent Memory
35mDeveloping Software for Persistent Memory
14mBurstable Cloud Block Storage with Data Processing Units
28mDPDK vs ODP: A comparison