• 👨‍🏫 no, but low level, OS process might need to

Program structure

• Int[128,128] data;
• each row: stored in 1 page = 128 objects
• program 1 (1 page allocated)

  for (j = 0; j < 128; j++)
      for (i = 0; i < 128; i++)
          data[i,j] = 0;
  

  • picking different row each times
    • resulting in: $128\times 128=16,384$ page faults
  • d
• program 2

  for (i = 0; i < 128; i++)
      for (j = 0; j < 128; j++)
          data[i,j] = 0;
  

  • resulting in: $128$ page faults
    • once every row change

Moving head disk mechanism

• each disk platter: w/ flat circular shape
  • diameter of 1.8, 2.5, or 3.5 inches
• two surface of a platter: covered with a magnetic materials
  • stores information
• read-write head "files": just above each surface of every platter
  • heads: attacked to disk arm: moving all heads as a unit
  • 👨‍🏫 back and forth!
• platter surface: logically divided into circular tracks
  • then subdivided into hundreds of sectors
• sef of tracks at one arm position: makes a cylinder
  • thousands of concentric cylinders in a disk drive!
• 
• 
  • 👨‍🎓 lovely!

Mass storage structure overview

• magnetic disks: provide bulk of secondary storage of modern computers
  • drives: rotate at 60-250 times / second
  • transfer rate: rate of data flow between drive and computer
  • positioning time (random-access time): time to
    1. move disk arm to desired cylinder (seek time)
    2. time for desired sector to rotate under the disk head (rotational latency)
  • head crush: from disk head making (physical) contact w/ disk surface
    • 👨‍🏫 bad
• disks: can be removable
• drive: attache dto computer via IO bus
  • busses: vary
    • including EIDE, ATA, SATA, USB, Fibre Channel, SCSI, SAS, Firewire
    • host controller in computer: uses bus to talk to disk controller
      • disk controller: built into drive / storage array

Hard disk drive

• platters: range from /85" to 14" (historically)
  • commonly: 3.5", 2.5", and 1.8"
• range from: 30GB to 3TB per drive
• performance
  • transfer rate: (theoretical) 6 Gb / sec
  • effective transfer rate: (real) 1 Gb / sec
  • seek time: from 3 ms to 12 ms
    • average seek time: measured based on 1/3 of tracks
    • RPM: typically 5400, 7200, 10000, and 15000 rpm
    • latency: based on spindle speed
      • 1 / (RPM / 60) = 60 / RPM
  • average latency: 1/2 latency
  • e.g. 7200 RPM = 120 rps
    • average latency: 1/2*1/120 = 4.17 mini-seconds

HDD performance

• access latency / average seek time: average seek time + average latency
  • for fast disk: 3 ms + 2 ms = 5 ms
  • for slow disk: 9 ms + 5.56 ms = 14.56 ms
• average IO time: average access time + (amount to transfer / transfer rate) + controller overhead
  • e.g. 4KB block on a 7200 RPM disk, w/ 5 ms average seek time & 1 Gb/s transfer rate w/ 0.1 ms controller overhead:
    • 5 ms + 4.17 ms + 4 KB / (1 Gb / s) + 0.1 ms =
      • (notice: one is bytes, the other bits)
      • 4 KB / (1 Gb / s) = 2^{-15} s = 0.0305 ms
    • 9.27 ms + 0.0305 ms = 9.3 ms
    • taking average 9.3 ms to transfer 4 KB, w/ effective bandwidth of 4 KB / 9.3 ms ≈ 3.5 Mb / sec only
    • w/ transfer rate at 1 Gb/s given the overhead
  • huge gap in HDD between random & sequential workloads
    • a disk w/ 300 GB capacity, avg. seek time 4 ms, RPM 15,000 RPM = 250 RPS (avg. latency = 2 ms), transfer rate 125 MB/s, 4KB read at random location
      • average IO time = 4 ms + (4 / 125,000) s + 2 ms = 6.03 ms
      • effective bandwidth: 4 KB / 6 ms = 0.66 MB/s
    • w/ sequential access of a 100 MB file: one seek & rotation needed (ideally)
      • can yield: effective bandwidth / transfer rate close to 125 MB/s

Solid state disk

• SSD: a nonvolatile memory (NVM) used like a hard drive
  • many tech variations, e.g. from DRAM w/ battery to maintain state its state in power failure
    • w/ flash-memory technology
      • single-level cell (SLC)
      • multi-level cell (MLC)
  • SSDs: can be more reliable than HDD as no moving / mechanical parts
  • much faster: as there is no seek time / rotation latency
  • consumes less power: power efficiency
  • yet, more expensive per MB, w/ less capacity & shorter life span
• as it's much faster than magnetic disks: standard bus interface mgiit be too slow, being a bottleneck
  • some: connect directly to system bus (e.g. PCI)
  • some: use them as a new cache tier
    • moving data between magnetic disk, SSDs, and memory: for optimizing performance

Magnetic tape

• magnetic tape: an early secondary storage medium
  • relatively permanent & can hold large quantities of data
  • access time: slow
    • moving to correct spot might take minutes
  • random access: ≈ 1000 times slower than magnetic disk
    • not useful for secondary storage in modern computer systems
• mainly used for backup, storage of infrequently used data
  • or: medium of transferring information from one system to another
• tape capacities: vary greatly
  • depending on particular kind of tape drive
    • w/ current capacities exceeding several terabytes
  • typically between 200 GB and 1.5 TB

Disk structure

• drives: addressed as large 1d arrays of logical blocks
  • logical block: smallest unit of transfer
    • disk: represented as a no. of disk blocks
      • each block: w/ unique block number: disk address
    • size of a logical block: usually 512 bytes
    • low-level formatting: creates logical blocks on physical media
• 1d array of logical blocks: mapped into sectors of the disk (sequentially)
  • sector 0: first sector of first track
    • on the outermost cylinder
  • mapping proceeds in order through track, then rest of tracks in the cylinder
    • then rest of cylinders: from outermost to innermost
  • logical to physical address (consist of: cylinder number, track number in cylinder, sector no. in track): must be easy, except:
    • defective sectors: mapping: hides this by substituting spare sectors from elsewhere on disk
    • no. of sectors / track: may not by constant on some devices
      • non-constant no. of sectors / track: via constant angular velocity

Disk scheduling

• OS: responsible for using hardware efficiently
  • for disk drives: meas having fast access time & large disk bandwidth
• seek time: time for disk head arm to move the head to corresponding cylinder containing desired sector
  • measured by: seek distance in term of no. of cylinders / tracks
• rotational latency: additional time for the disk to rotate desired sector to disk head
• disk bandwidth: total no. of bytes transferred / total time between first service req. and completion of transfer
  • improve access time & bandwidth by: managing order of disk IO requests serviced
• many resources of disk IO requests
  • from OS, system processes, and user processes
  • IO requests: include
    • IO modes
    • disk address
    • memory address
    • no. of sectors to transfer
• OS: maintains a queue of requests / disk or device
  • in multiprogramming system w/ many processes: disk queue often has several pending requests
• idle disk: can immediately work on IO request
  • for busy disk: requests must be queued
  • optimization: only when the queue of request exists
• disk drive controllers: w/ small buffers & manage a queue of IO requests (of varying "depth")
  • which request to be completed & selected next?
  • disk scheduling
• now: illust scheduling algorithms w/ request queue, having 0-199 as cylinder numbers
  • e.g. 98, 183, 37, 122, 14, 124, 65, 67
  • current head position: 53

FCFS / FIFO

• intrinsically: fair, yet doesn't proved the fastest service
• resulting in: total head movement of 640 cylinders
• 

SSTF: Shortest seek time first

• selects req. w/ least seek time from current head position
  • i.e. pending request closest to current head position
• SSTF: a form of SJF scheduling (greedy)
  • may cause: starvation of some requests
• results in: 236 cylinders
• 

SCAN scheduling

• disk: starting at one end of the disk, moving towards the other end
  • on the way: servicing requests as it reaches each cylinder
• at the other end: direction of head movement: reversed, and continue the service
  • aka: elevator algorithm
• note: if requests are uniformly distributed across cylinders
  • heaviest density of requests: at other end of disk, wait the longest
  • must know: direction of head movement
• results in: 236 cylinders
• 

LOOK scheduling

• similar to SCAN, but disk arm only goes as far as the final request in each direction
  • NOT the end of the disk
• 👨‍🏫 guaranteed to be not worst than SCAN
• results in: 208 cylinders
• 

C-SCAN

• Circular-SCAN: provides a more uniform waiting time than SCAN
• head: moves from one end of the disk to the other, servicing requests as it goes
  • upon reaching the other end: immediately returns to the beginning of the disk
    • not serving any requests on the way back
  • treats:cylinders as a circular lists, wrapped around
    • from the last cylinder to the first one
• results in: 382 cylinders
• 

C-LOOK

• LOOK version of C-SCAN
• disk arm: goes far as the last request in each direction
  • then reversing the direction
• better than C-SCAN, for the very same reason
• results in: 322 (C-LOOK), 308 (LOOK) cylinders
• 

Selecting a disk-scheduling algorithm

• common solution: SSTF
  • increases performance of FCFS
  • SCAN, C-SCAN: performs better for systems w/ heavy load on disk
    • as: they are less likely to cause starvation problems
• scheduling performance: depends on the no. o& ype of requests
  • w/ 1 request: all scheduling algorithm must behave the same
• requests for disk services: greatly influenced by the file-allocation method (to be discussed)
  • contiguously allocated file: generate several requests, close together on the disk
    • resulting in limited head movement
  • linked or indexed file: may include blocks widely scattered on the disk
    • resulting in greater head movement
  • location of directories & index blocks: also important
    • as: they are accessed frequently
    • directory entry and file data on different cylinders
      • => cause excessive head movement
    • caching directory & index block in memory: helps
• disk-scheduling algorithm: written as a separate module of the OS
  • allowing to be replaced w/ different algorithm if necessary
  • either SSTF / LOOK: a reasonable choice as the default algorithm

RAID: improving reliability via redundancy

• by: Randy Kal & David Peterson, late 80-90s
  • 👨‍🏫 I know one of them very well
• RAID: Redundant Arrays of Independent Disks
  • in the past: composed of small, cheap disks
    • viewed as: a cost-effective alternatives to large & expensive disks
      • once called: redundant arrays of inexpensive disks
  • now: used for higher reliability via redundancy & higher data-transfer rate (access in parallel)
• increases: mean time to failure