Espresso: Constructing Cost-Efficient CXL JBOF via Inter-SSD Computing Resource Sharing
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15
summary
Espresso builds a JBOF that disaggregates SSDs into disaggregated compute/data ends and uses the cache-coherent CXL fabric to harvest idle onboard ARM cores and DRAM from other drives for metadata processing and mapping-table caching. It requires a minor NVMe-driver tweak to redirect queues, implements decentralized resource monitors/managers and DMA coordination, and is evaluated end-to-end: +50.4% SSD resource utilization, ≈19% monetary cost reduction, 38–41% throughput wins versus alternatives while matching performance with roughly half the per-SSD compute.
AI-written summary · about
chapters (9)
- 0:00Espresso and storage challenges
- 3:13Cost-efficient SSD with resource sharing
- 6:07CXL for fine-grained sharing
- 8:39Espresso workflow and init
- 10:06Disaggregated SSD architecture
- 12:41Processor harvesting via NVMe redirection
- 13:46DRAM harvesting and consistency
- 14:06Validation results and performance
- 15:50Cost efficiency and workload benefits