Here, There and Everywhere: The Past, the Present and the Future of Local Storage in...
summary
Cloud local storage evolution is traced from Espresso (SPDK user-space poll), which cuts kernel context-switch overhead but wastes CPU (99% of nodes show <60% effective utilization and 5–12 µs syscall overhead), to Doppio (ASIC offload), which removes host CPU but lags SSD advances, to Ristretto (ASIC+SOC) that balances offload and cloud features. A hybrid Latte (ML dispatcher + admission controller) prototype combines local cache with EBS backend, hits ~18% cache rate, delivers near-physical latency and lower cost than EBSX 30µs (EBSX is ~20× costlier) while reducing tail latency and surviving GC.
AI-written summary · about
chapters (11)
- 0:00Local storage evolution
- 1:22Kernel limits with NVMe SSDs
- 2:49SPDK Espresso cloud limits
- 3:54ASIC Doppio challenges
- 5:39Ristretto ASIC/SOC co-design
- 7:22Local storage limits vs EBS
- 8:48EBSX with persistent memory
- 9:11Latte hybrid storage
- 10:57Evaluation setup
- 13:49Cost comparison
- 14:13Local storage summary
