• Physical links and form factors
• SATA/AHCI overview
• NVMe over PCIe
• Submission/completion queues and interrupts
• Namespaces, multipath, features
• Controller internals and firmware
• Filesystems and block layers
• Performance: latency, IOPS, throughput
• Exercises

• Form factors: 2.5" SATA, M.2 (B/M key), U.2/U.3, EDSFF (E1.S/E3), Add-in Cards (AIC).
• Links: SATA (6 Gb/s), PCIe (Gen3/4/5, x1..x16); PCIe lanes dictate peak bandwidth.
• HBAs/RAID cards aggregate drives; power-loss protection matters for data integrity.

• AHCI defines registers and command list processing; optimized for HDD semantics.
• NCQ (Native Command Queuing) improves concurrency but limited queue depth (32).
• SATA bottlenecks SSDs due to link and protocol overhead.

• Designed for SSDs: parallel queues, doorbells, minimal register access.
• Admin and I/O queue pairs per CPU core; MSI-X interrupts enable scaling.
• NVMe over Fabrics (NVMe-oF): RDMA/TCP transports for disaggregated storage.

Doorbell write → device processes SQ entries → CQ entry + interrupt

• Each I/O queue is a circular buffer in memory; host writes SQ entries, device writes CQs.
• Queue depth and interrupt coalescing balance latency vs throughput.
• Polling (io_uring, SPDK) reduces latency by avoiding interrupt overheads.

// io_uring submission sketch
io_uring_sqe* sqe = io_uring_get_sqe(&ring);
io_uring_prep_read(sqe, fd, buf, len, off);
io_uring_submit(&ring);

• Namespaces partition capacity; ANA/MPIO support multipath and failover.
• Features: encryption, sanitize/secure erase, firmware updates, telemetry.
• SMART/health monitoring predicts endurance and failure.

• FTL maps LBAs to flash pages; wear leveling and garbage collection maintain performance.
• Write amplification impacts endurance; OP (over-provisioning) mitigates.
• DRAM-less controllers use HMB; PLP (caps) protect metadata on power loss.

• IO stack: VFS → filesystem → block layer → driver → device; merge and scheduler policies matter.
• Modern filesystems (ext4, XFS, APFS, Btrfs, ZFS) exploit SSD parallelism differently.
• Userspace stacks (SPDK) bypass kernel for ultra-low latency.

• IOPS bound by queue depth and controller; throughput by lanes × rate × efficiency.
• Workload mix (read/write, seq/random) changes optimal tuning (QD, coalescing, polling).
• Measure with fio; watch tail latency (p99/p999) and thermal throttling.

fio --name=randread --filename=/dev/nvme0n1 --ioengine=io_uring --iodepth=64 --bs=4k --rw=randread --numjobs=4

1. Compare SATA SSD vs NVMe SSD latency and throughput with fio; chart QD scaling.
2. Enable polling (io_uring) and measure latency improvements vs interrupt mode.
3. Simulate endurance impact by varying write amplification in an SSD model.