Storage Devices

Lecture Notes for CS 140
Winter 2013
John Ousterhout

  • Readings for this topic from Operating Systems: Principles and Practice: Section 12.1.

Magnetic Disk (Hard Drive)

  • Basic geometry:
    • 1-5 platters, magnetically coated on each surface
    • Platters spin at 5000-15000 RPM
    • Actuator arm positions heads, which can read and write data on the magnetic surfaces.
    • Overall size of disk package: 1-8 inches.
  • Organization of disk data:
    • Circular tracks corresponding to a particular position of the actuator arm.
    • Typical density today: 200,000 tracks per radial inch.
    • Tracks divided into 512-byte sectors. Typical tracks contain a few thousand sectors.
    • Typical total drive capacities: 100GB-2TB
      • 100GB ~ 50M double-spaced pages of text.
    • Disk technology is one of the most rapidly advancing technologies: capacities increasing faster than Moore's Law.
  • Reading and writing:
    • Seek: move actuator arm to position heads over desired track. Typical seek time: 2-10ms.
    • Select a particular head.
    • Rotational latency: wait for desired sector to pass under the head. One-half disk rotation on average (4ms @ 7500RPM)
    • Transfer: read or write data as it passes under the head. Typical transfer rates: 100-150 MBytes/sec.
    • Latency refers to the sum of seek time plus rotational latency; typically 5-10ms.
  • API for disks:
    read(startSector, sectorCount, buffer)
    write(startSector, sectorCount, buffer)
  • In the old days the track and surface structure of the disk was visible to software:
    read(track, sector, surface, sectorCount, buffer)
  • Nowadays the track structure is hidden inside the disk:
    • Inner tracks have fewer sectors than outer tracks
    • If some sectors are bad, disk software automatically remaps them to spare sectors.

Reel-to-Reel Magnetic Tape

  • Old technology, popular in the 1960s and 1970s.
  • Magnetic tape:
    • 9 tracks across tape (one byte plus parity),
    • 1/2" wide by 2400 feet long
    • Variable-length records (20-30000 bytes); read or write blocks, but cannot write in middle
    • Recording density up to 6250 bytes/inch (180 MBytes max for a tape)
    • Tape moves at 20-200 inches/sec. (up to a few Mbytes/sec.)
    • Interesting physical design, such as vacuum columns to buffer tape during fast start and stop.

Communicating with I/O Devices

  • Device registers:
    • Each device appears in the physical address space of the machine as a few words, called device registers.
    • The operating system reads and writes device registers to control the device.
    • Bits in device registers serve 3 purposes:
      • Parameters provided by CPU to device (e.g. number of first sector to read)
      • Status bits provided by device to CPU (e.g. "operation complete" or "error occurred").
      • Control bits set by CPU (e.g. "start disk read") to initiate operations.
    • Device registers don't behave like ordinary memory locations:
      • "Start operation" bit may always read as 0.
      • Bits may change without being written by CPU (e.g. "operation complete" bit).
    • On some machines special instructions are used to access device registers and initiate I/O, rather than reading and writing special memory locations.
  • Programmed I/O: all communication with I/O device occurs through the device registers.
    • CPU writes device registers to start operation (e.g., read)
    • CPU polls ready bit in device register
    • When operation is finished, device sets ready
    • CPU reads data from one or more device registers, writes data to memory.
    • Problems with this approach:
      • CPU wastes time waiting for data to become ready; can only keep one device busy at a time.
      • Expensive for CPU to mediate all data transfers; with some fast devices CPU can't keep up with device.
  • Interrupts: allow CPU to do other work while devices are operating.
    • CPU starts I/O operation, then works on other things.
    • When device needs attention (operation completes) it interrupts the CPU:
      • A forced procedure call to a particular address in the kernel.
    • Operating system figures out which device interrupted, services that device.
    • Operating system returns from the interrupt back to whatever it was working on.
    • Interrupts make the operating system much more efficient; for example, can keep many devices busy at the same time, while also running user code.
  • Direct Memory Access (DMA):
    • Device can copy data to and from memory, without help from the CPU.
    • CPU loads buffer address into a device register before starting operation (e.g., where to copy data read from disk).
    • Device moves data directly into memory.
    • When transfer complete, device issues an interrupt to the system.
    • Today DMA is the norm for I/O devices (controller hardware is cheap).