FlashShare
FlashShare is a research storage-stack design that carries workload criticality from the host kernel into SSD firmware to prioritize latency-sensitive I/O.
Cross-stack policy: Operators assign workload attributes, which travel through kernel I/O structures into a reserved NVMe command field. Modified queue handling prioritizes latency-critical requests, while SSD firmware partitions its data cache and adapts prefetching. This requires coordinated host and device changes, not an ordinary NVMe configuration switch. [1]
Completion path: Select-ISR polls latency-critical completions and uses interrupts for throughput-oriented work. Selective polling still consumes host resources; the separate XLER accelerator design offloads polling and queue-management work. These are distinct design variants, not interchangeable names for interrupt coalescing. [2]
Evaluation: Section 5 uses gem5 with SimpleSSD, eight simulated ARM cores, Linux 4.9.30 and a scaled 2 GB host memory. Its end-to-end response metric includes client, network, server and storage time. Real-device characterization supports the model but does not turn the complete cross-stack evaluation into a deployed-device benchmark. [1]
Relations
- related
- I/O completion polling, FLIN
Limitations
- Prioritizing latency-critical work can delay throughput-oriented work; a faster storage medium alone does not remove shared-stack interference. [1]
Citations
- Zhang et al. FlashShare: Punching Through Server Storage Stack from Kernel to Firmware for Ultra-Low Latency SSDs. OSDI 2018, pp. 477-492, sections 3-5.
- Zhang et al. FlashShare, OSDI 2018 author slides, completion paths and accelerator design
Further reading
Sources & caveats
- Primary-paper and author-slide review; no firmware execution, hardware benchmark or performance reproduction was performed.
- The paper's prototype characterization and full-system simulation are separate evidence sources. [1]