A quick look at secondary storage and file systems.

Objectives

Notes

• I will do a whirlwind tour of Chapters 11 - 15.
• Hardware
  • Hard Disk Drives (HDD)
    • Old
    • Mechanical
    • Slow and big.
    • Transfer rate - tens to hundreds of MB/sec
    • positioning time, seek time, rotational latency , milliseconds
    • Head crash.
    • Read blue box bottom of 451.
  • Nonvolatile Memory (NVM)
    • new
    • MUCH faster
    • Becoming larger and cheaper.
    • Subject to limitations of rewrite.
      • Which appears to actually be erase then write.
      • Which makes write slower than read.
    • Lifetime is measured in Drive Writes per Day.
    • Take a look at This page.
      • 1 SDWPD (100% 128KB sequential writes) : Easiest test, overwrite the entire drive with sequential writes.
      • 1 RDWPD (100% 16KB random writes): More stressful, write large random blocks.
      • 0.25 RDWPD (100% 4KB random writes): Most stressful, write little random blocks.
      • These critters cost around $4k (30TB ssd).
  • RAM drives
    • Treat a large chunk of NVRAM as a device by adding a file system.
    • Much faster than SSD
    • But temporary
• HDD Scheduling
  • It is likely that there are multiple I/O requests occur at nearly the same time.
  • The drive is most likely only able to schedule one of these at a time.
  • These are placed in a queue
  • So the OS is responsible for scheduling which one occurs
  • Our old friend FCFS is easiest
    • But could have the device seeking all over the place (HDD)
  • SCAN
    • HDD algorithm
    • Start at one end of the disk,
    • Service the requests in order of how they line up in the scan.
    • Then scan on the way back.
    • Start over again.
    • This is called the elevator algorithm
  • C-SCAN
    • Scan but with fly-back.
    • This makes wait time more uniform.
  • The book says there are many other algorithms.
  • Old Linux
    • Processes are more likely to block on read then write.
    • Linux implements a queue for each.
    • Then C-SCAN is used.
    • Requests are sent in batches.
    • And queue selection is modified to prevent starvation.
  • New(er) linux
    • Three queues (rt, normal, batch)
    • There can be starvation
    • Anticipates a process' need based on historic data and prioritizes heavy I/O processes to minimize seek time.
  • There may be newer, but I did not look.
• NVM Scheduling
  • No moving parts, so this is different.
  • And the NVM devices have their own practices at a lower level to prevent burn-out.
  • This is a very short section, I bet they are working on/have better algorithms now.
  • A quick google search suggests FCFS