I/O Systems

1
I/O Systems

• I/O Hardware
• Application I/O Interface
• Kernel I/O Subsystem
• Transforming I/O Requests to Hardware Operations
• Streams
• Performance

2
I/O Hardware

• Wide variety of I/O devices
• Common concepts
• port
• bus (daisy chain or shared single medium)
• controller (host adapter)
• I/O instructions control devices
• Devices have addresses, used by
• direct I/O instructions
• memory-mapped I/O

3
A Typical PC Bus Structure
4
Device I/O Port Locations on PCs (partial)
5
Polling

• Determines state of device
• command-ready
• busy
• error
• Busy-wait cycle to wait for I/O from device

6
Interrupts

• CPU interrupt request line triggered by I/O
  device
• Interrupt handler receives interrupts
• Maskable to ignore or delay some interrupts
• Interrupt vector to dispatch interrupt to correct
  handler
• based on priority
• some unmaskable
• Interrupt mechanism also used for exceptions,
  traps

7
Interrupt-Driven I/O Cycle
8
Intel Pentium Processor Event-Vector Table
9
Direct Memory Access (DMA)

• Avoid programmed I/O to move large data
• Requires DMA controller
• Bypass CPU
• transfer data directly I/O device ltgt memory

10
Six Step Process to Perform DMA Transfer
11
Application I/O Interface

• I/O system calls abstract device behaviors in
  generic classes
• Device-driver layer hides I/O-controller
  differences from kernel
• Devices vary in many dimensions
• character-stream or block
• sequential or random-access
• sharable or dedicated
• speed of operation
• read-write, read only, or write only

12
A Kernel I/O Structure
13
Characteristics of I/O Devices
14
Block and Character Devices

• Block devices include disk drive
• commands include read, write, seek
• raw I/O or file-system access
• file system maps location i onto block offset
• memory-mapped file access possible
• Character devices include keyboard, mouse, serial
  port
• commands include get, put
• libraries layered on top allow line editing

15
Network Devices

• Differ enough from block, character to have own
  interface
• Unix and Windows NT/9x/2000 have socket interface
• separates network protocol from network operation
• includes select functionality (eliminates
  polling)
• Wide variations, e.g.
• pipes, FIFOs, streams, queues, mailboxes

16
Clocks and Timers

• Provide current time, elapsed time, timer
• programmable interval timer for periodic
  interrupts
• UNIX system calls
• timer_create
• get_itimer
• Check their man pages

17
Blocking and Nonblocking I/O

• Blocking process suspended until I/O completed
• Easy to use and understand
• Insufficient for some needs
• Nonblocking I/O call returns as much as
  available
• user interface, data copy (buffered I/O)
• implemented via multi-threading code for I/O call
• returns quickly with count of bytes transferred
• Asynchronous process runs while I/O executes
• difficult to use
• I/O subsystem signals process when I/O completed
• e.g., callbacks pointer to completion code

18
Kernel I/O Subsystem

• Scheduling
• some I/O request ordering via per-device queue
• some OSs attempt fairness
• Buffering data in memory between transfers
• cope with device speed mismatch
• cope with device transfer size mismatch
• maintain copy semantics
• buffer guaranteed to be same as at start of I/O
  call

19
Sun Enterprise 6000 Device-Transfer Rates
20
Kernel I/O Subsystem

• Caching fast memory with copy of data
• must (eventually) update original
• key to performance
• Spooling hold output for a device
• if device can serve only one request at a time
• e.g., printing
• Device reservation exclusive access to device
• system calls for allocation, deallocation
• potential for deadlock

21
Error Handling

• OS can recover from disk read, device
  unavailable, transient write failures
• Most return error number or code if I/O request
  fails
• System error logs hold problem reports

22
Kernel Data Structures

• Kernel keeps state info for I/O components,
  including open file tables, network connections,
  character device state
• Many complex data structures to track buffers,
  memory allocation, dirty blocks
• Some use object-oriented methods and message
  passing to implement I/O

23
UNIX I/O Kernel Structure
24
I/O Requests to Hardware Operations

• Consider reading file from disk for process
• determine device holding file
• translate name to device representation
• read data from disk into buffer (slow)
• make data available to requesting process
• return control to process

25
Life Cycle of I/O Request
26
STREAMS

• STREAM full-duplex communication channel between
  user-level process and device
• STREAM consists of
• - STREAM head interfaces with user process
• - driver end interfaces with device- zero or
  more STREAM modules between them
• Each module contains read queue and write queue
• Message passing used to communicate between queues

27
The STREAMS Structure
28
Performance

• I/O major factor in system performance
• CPU executes device driver, kernel I/O code
• context switches due to interrupts
• data copying
• network traffic especially stresses system

29
Intercomputer Communications
30
Improving Performance

• Reduce number of context switches
• Reduce data copying
• Reduce interrupts by large transfers, smart
  controllers, polling
• Use DMA
• Balance CPU, memory, bus, and I/O performance for
  highest throughput

31
Device-Functionality Progression
32
Mass-Storage Systems

• Disk Structure
• Disk Scheduling
• Disk Management
• Swap-Space Management
• RAID Structure
• Disk Attachment
• Stable-Storage Implementation
• Tertiary Storage Devices
• OS Issues
• Performance Issues

33
Disk Structure

• Disk drives addressed as large 1-dimensional
  arrays of logical blocks (smallest transfer unit)
  
• 1-dimensional array of logical blocks mapped onto
  sectors of disk sequentially
• sector 0 1st sector of 1st track on outermost
  cylinder
• mapping in order through that track, then rest of
  tracks in that cylinder, then through rest of
  cylinders from outermost to innermost

34
Disk Scheduling

• OS responsible for using hardware efficiently
  for disk drives having low access time, high
  bandwidth
• Access times two major components
• seek time time for disk arm to move head to
  desired cylinder
• rotational latency time for disk to rotate
  desired sector to head
• Minimize seek time
• Seek time ? proportional to seek distance
• Disk bandwidth total number of bytes
  transferred, divided by total time from start to
  end of transfer
• Total transaction time access time bytes /
  bandwidth
• realistically add OS overheads, queuing delay,
  etc.

35
Disk Scheduling

• Several algorithms schedule disk requests
• Illustrate with request queue
• 200 blocks, numbered 0..199
• head initially at block 53
• 98, 183, 37, 122, 14, 124, 65, 67

36
First-Come, First-Served (FCFS)
total head movement 640 cylinders (add over all
moves)
37
Shortest Seek-Time First (SSTF)

• Selects request with minimum seek time from
  current head position
• SSTF scheduling form of SJF scheduling may cause
  starvation
• Next slide total head movement 236 cylinders
  (compare FCFS 640)
• not optimal can e.g. cut to 208 cylinders

38
SSTF example
39
SCAN

• Arm starts at end of disk, moves to other end,
  servicing requests up to far end head movement
  and service order reversed
• Sometimes called elevator algorithm
• Next slide total head movement 242 cylinders
• is this better than SSTF?
• why use this algorithm?

40
SCAN Example
41
C-SCAN

• Provides more uniform wait time than SCAN
• head moves from one end of disk to other
  servicing requests as it goes
• when reaches other end immediately returns to
  beginning without servicing requests on return
  trip
• Treats cylinders as circular list that wraps
  around from last to first
• whats total head movement now?

far end of disk more likely to have more new
requests
42
C-SCAN
43
C-LOOK

• Version of C-SCAN
• Arm only goes as far as last request in each
  direction, then right back to first request at
  other end of disk
• Is this a good idea?

44
C-LOOK (Cont.)
45
Selecting a Disk-Scheduling Algorithm

• SSTF common, natural appeal
• SCAN and C-SCAN perform better if heavy load on
  disk
• Performance depends on number and types of
  requests
• Requests for disk service influenced by
  file-allocation method
• Disk-scheduling should be separate module of OS,
  allowing replacement with different algorithm if
  necessary
• SSTF and LOOK both reasonable default choices

46
Disk Management

• Low-level formatting, or physical formatting
  divide disk into sectors that disk controller can
  read and write
• To use disk to hold files, OS needs to record own
  data structures on disk
• partition disk into 1 groups of cylinders
• logical formatting or making a file system
• Boot block to start up system
• bootstrap code in ROM
• bootstrap loader program minimum in ROM
• Methods such as sector sparing to handle bad
  blocks

47
MS-DOS Disk Layout
48
Swap-Space Management

• Swap-space VM uses disk to extend main memory
• Swap-space can be part of normal file system, or
  (traditionally) separate partition
• Swap-space management
• 4.3BSD allocates swap space when process starts
  holds text segment (code) and data segment
• kernel uses swap maps to track swap-space use
• Solaris 2 allocates swap space only when page
  forced out of physical memory

49
4.3 BSD Text-Segment Swap Map
50
4.3 BSD Data-Segment Swap Map

• Size can grow
• Efficient support for small and large processes
• each new block allocated 2x previous size

51
RAID Structure

• RAID multiple disk drives provide high
  availability via redundancy
• History
• large drives used to be cheaper per bit
• in 1980s small drives became cheaper per bit
• using many drives unreliable (more parts to fail)
• RAID arranged into 7 different levels

redundant array of inexpensive disks
52
RAID

• Multiple disks working cooperatively
• Disk striping uses group of disks as 1 unit
• data split across drives
• RAID schemes improve performance and availability
  (dependability) with redundant data
• mirroring or shadowing duplicates each disk
• block interleaved parity (ECC) much less
  redundancy

book refers to reliability not correct
reliable nothing breaks
53
RAID Levels
54
RAID (0 1) and (1 0)
55
Disk Attachment

• Disks may be attached two ways
• Host attached via I/O port
• Network attached via network connection

56
Network-Attached Storage
57
Storage-Area Network
58
Stable-Storage Implementation

• Write-ahead log requires stable storage
• To implement stable storage
• replicate information on gt 1 nonvolatile storage
  medium sequentially
• on recovery, only successful if both writes
  complete (i.e., pair of blocks equal)

59
Tertiary Storage Devices

• Low cost defining characteristic
• Generally uses removable media
• examples floppy disks, CD, DVD, tapes

60
Removable Disks

• Floppy disk flexible disk coated with magnetic
  material in protective plastic case
• most floppies about 1 MB similar technology for
  removable disks gt 1 GB
• removable magnetic disks can be nearly as fast as
  hard disks, but greater risk of damage

61
Removable Disks

• Magneto-optic disk records data on rigid platter
  coated with magnetic material
• laser heat amplifies large, weak magnetic field
  to record bit
• laser light also used to read data (Kerr effect)
• magneto-optic head much further from disk surface
  than magnetic disk head, and magnetic material
  covered with protective clear layer resistant to
  head crashes
• Optical disks do not use magnetism use materials
  altered by laser light examples
• CD-RW
• DVD-RW

62
WORM Disks etc.

• Data on read-write disks can be modified over and
  over
• WORM (Write Once, Read Many) disks written once
• thin aluminum film sandwiched between two clear
  layers
• to write bit drive uses laser light to burn small
  hole through aluminum recording can be destroyed
  not altered
• very durable and reliable
• Examples CD-R, DVD-R
• Read-only disks, e.g., CD-ROM, DVD, factory
  recorded

63
Tapes

• Compared to disk, tape less expensive, holds
  more, but random access much slower
• Tape economical medium if fast random access not
  needed, e.g., backup of disk, archive of huge
  data
• Large tape installations can use robotic tape
  changers move tapes between drives and slots in
  tape library
• stacker library holds few tapes
• silo library holds thousands
• Disk-resident file can be archived to tape for
  low cost storage computer can stage it back to
  disk for active use

64
OS Issues

• Major OS tasks manage physical devices, virtual
  machine abstraction for applications
• For hard disks, OS abstractions
• raw device array of data blocks
• file system names, seeks, reads, writes, etc.

65
Application Interface

• Most OSes removable disks almost exactly like
  fixed new cartridge formatted, empty file
  system generated
• Tapes presented as raw storage medium, i.e.,
  application does not not see files on tape, opens
  whole tape drive as raw device
• Usually drive reserved for exclusive use of 1
  application
• Since OS doesnt provide file system services,
  application must decide how to use array of
  blocks
• Since every application makes up own tape
  organization, tape generally only usable by
  program that created it
• Some standards, e.g., tar format (give or take
  variations)

66
Tape Drives

• Basic operations for tape drive differ from those
  of disk
• locate positions tape to specific logical block,
  not entire track (corresponds to seek)
• read position operation returns logical block
  number where tape head is
• space operation for relative motion
• Tapes append-only devices updating block in
  middle effectively erases everything beyond that
  block
• EOT mark placed after block that is written

67
File Naming

• Naming files on removable media especially
  difficult write data on removable cartridge on
  one computer, then use on another
• Current OSes generally leave name space problem
  up to applications, users to work out how to
  access, interpret data
• Some kinds of removable media (e.g., CD ISO
  format) standardized

68
Hierarchical Storage Management (HSM)

• Hierarchical storage system extends hierarchy
  beyond primary memory and secondary to tertiary
  often jukebox (tapes, removable disks)
• Incorporate tertiary storage extend file system
• small and frequently used files on disk
• large, old, inactive files archived to jukebox
• HSM usually found in supercomputing centers,
  other large installations with enormous data

69
Speed

• 2 aspects of tertiary storage speed
• Bandwidth
• latency
• Bandwidth measured in bytes per second
• sustained bandwidth average data rate during a
  large transfer of bytes/transfer timedata
  rate when data stream actually flowing
• effective bandwidth average over entire I/O
  time, including seek or locate, and cartridge
  switchingdrives overall data rate

70
Speed

• Access latency time to locate data
• access time for disk move arm to selected
  cylinder, wait for rotational latency lt 15
  milliseconds
• access on tape requires winding tape reels until
  selected block reaches tape head tens or
  hundreds of seconds
• generally random access within tape cartridge
  about 1,000 times slower than disk random access
• Low cost of tertiary storage is result of many
  cheap cartridges etc. sharing few expensive
  drives
• Removable library best to store infrequently used
  data library can relatively low rate of requests
  (per hour)

71
Reliability

• Fixed disk drive likely to be more reliable than
  removable disk or tape
• Optical disk likely to be more reliable than
  magnetic disk or tape
• Head crash in fixed hard disk generally destroys
  data, whereas failure of tape drive or optical
  disk drive often leaves data unharmed

72
Cost

• Main memory much more expensive than disk
  storage
• Cost / megabyte of hard disk competitive with
  tape if 1 tape per drive
• Cheapest tape drives and disk drives about same
  storage capacity historically
• Tertiary storage gives savings only when number
  of removable media gtgt number of drives

73
DRAM Price / MB 1981 2000
cheaper x 4 every 3 years recently x 2 every 3
years
74
Hard Disk Price / MB 1981 2000
cheaper x 2 every 3 years until c.1985 since x 4
every 3 years
75
Tape Drive Price per MB 1984 2000
of most significance if small number of tapes
76
Summary

• OS hides variations in I/O devices where
  possible
• I/O interface critical to performance
• Dependability
• RAID (high availability, fault recovery)
• stable storage (crash recovery)
• Idea of HSM under threat
• price trend favours using disks increasingly
• Long-term trends useful to watch could be new
  breakthrough