hard drives

Tags: hardware

Spinning Rust

• Operating systems implement all sorts of spinning rust scheduling, since they can make a good guess about how long operations will take (compared to jobs)
  • SSTF - shortest seek time first, but vulnerable to starvation due to a stream of nearby jobs
  • SCAN - loops back and forth, enqueueing for the next scan
  • SPTF - Shortest access first
• Disks can also merge IOs together

NVMe

• NVMe drives will lie about write speeds under low traffic
• NVMe drives are good for append-operations (aka structured logs)
• linux has nvme to deal with this
• in theory, the drives and the controllers are different, but in practice, these are sold as one whole thing
• https://nvmexpress.org/wp-content/uploads/NVM-Express-1_4c-2021.06.28-Ratified.pdf
• Leverages PCIe bus
• NVMe driver replaces the old block layer requester

Writes and Reads

• NVMe writes and reads are done through two circular buffers, a submission queue and a completion queue.
• When the host system has a command to send to the SSD, it places the command in the submission queue and then rings the doorbell. This updates the tailpointer of the queue
• Similar thing happens for completion queues, the drive updates the head doorbell when something is done

VS SATA and SCSI

• NVMe supports 64k queues with 64k commands per queue, so you have extra parallelism
• https://www.architecting.it/blog/performance-analysis-sas-sata-nvme/