Design of Operating Systems that support Distributed Multimedia Applications

1
Design of Operating Systems that support
Distributed Multimedia Applications

• Vijay Gondi
• Dhigha D Sekaran

2
What is Multimedia OS ?

• A Multimedia OS extends functionality of
  traditional OS's to accommodate multimedia data
  manipulation, so as to provide an environment for
  Real-Time Support, while simultaneously running
  traditional applications efficiently.
• The architecture of the MM OS is optimized for
  dealing with real-time, high-bandwidth data types
  such as audio and video, and for handling a wide
  array of communications capabilities.

3
How to Build Multimedia OS ?

• Extension of traditional OS.
• Constructing from scratch using micro-kernel
  architecture( Ex. CHORUS, Real-Time Mach).

4
Technology Thrust

• Multimedia workstations with audio and video
  processing capabilities
• Tremendous improvement in CPU processing power
• Special purpose compression engines for video and
  audio
• High capacity, high performance storage devices
• Availability of magnetic disks with several
  gigabyte capacity, tremendous increase in data
  transfer for large optical storage servers
  evolving rapidly
• High speed fibre optic networks and fast packet
  switching technology
• Networks with 100-200 Mbits/s already in place,
  prototype gigabit networks also being deployed

5
Service Vision and Applications

• Residential services
• Switching access television, video on demand
• Video phone/conferencing systems
• Video advertising and shopping (e.g, multimedia
  catalogue classified by category, product
  demonstrations, real estate sales,..) Residential
  services
• Business services
• Corporate education
• Desktop MM conferencing and multimedia e-mail
• Education services
• MM repository of available classes, videos,
  books,..
• Access to digital MM library over high speed
  networks
• Science and Technology
• Computational prototyping and scientific
  visualization
• Astronomy and environmental science studies
  efficient access to large number of satellite
  images
• Medicine
• Diagnosis and treatment- support queries like
  where was the tumor found, what happened to the
  tumor and assessment of therapeutic response(I.e
  , temporal analysis, etc

6
Hierarchy of Distributed Multimedia Applications

• Class 1 Information Retrieval
• Distance Learning
• Class 2 Information retrieval and processing
• Digital library of satellite imagery, subsurface
  modeling
• Desktop MM conferencing and multimedia e-mail
• Class 3 Interactive systems
• Education and training for fire fighters, command
  and control

7
Design Difficulties.

• Real-time performance requirements
• Large sizes of information objects

8
Real-Time requirements for Multimedia

• The fault-tolerance requirements of multimedia
  systems are less strict than those of real-time
  systems
• - Such as a delay in delivering
  video-on-demand, will not lead to the destruction
  of technical equipment
• Missing a deadline in a multimedia system is not
  a severe failure
• - If an uncompressed video frame is not
  prepared on time, it can simply be omitted
• A sequence of digital continuous-media data
  results from periodically sampling a sound or
  image signal
• - Scheduling periodic task is much easier than
  scheduling sporadic ones
• The bandwidth demand of continuous media is not
  always that stringent
• - If not enough bandwidth is available for full
  quality, the application can accept a reduced
  quality

9
Main Characteristics of Multimedia OS

• CPU Management
• Memory Management
• I/O Management
• File-System Management
• Resource Management
• Synchronization Issues

10
Resource Management

• In MM Operating System the given timing
  guaranties for the processing of continuous media
  must be adhered to by every H/W and S/W component
  along the data path.
• In a Distributed MM system several applications
  compete for resources
• Shortage of resources requires careful
  allocation.
• Resource types
• Active/Passive
• Active Resource like CPU, Network Adaptor
  provides a service
• Passive Resource like Main Memory, Communication
  Bandwidth, File System etc denotes system
  capability required by active resources
• Exclusive or Shared
• Single or Multiple A resource type that
  exists only once in the system is single,
  otherwise multiple

11
Phases of Resource Management

• Schedulability test The Resource Manager checks
  whether, given the Qos Parameters-Throughput and
  Reliability, for e.g.-there is enough resource
  capacity to handle the additional request
• Quality-of-service calculation The Resource
  Manager Calculates the best possible performance.
• Resource reservation The Resource Manager
  allocates the capacity required to meet the Qos
  guarantees for each request.
• Resource scheduling The incoming messages are
  Scheduled according the given Qos guarantees. In
  Process Management, for instance, the Scheduler
  allocates resources a the moment the data arrives
  for processing.

12
Resource Management

• QOS requirements
• - Throughput It is Determined by the
• data rate a connection needs to satisfy the
  application requirements, as well as the size of
  the data units
• - delay(local or global) Local Delay is the max
  time a resource takes to complete a certain task.
  Global Delay is the total delay for a data unit
  traveling from its source to its destination
• - Jitter Maximum allowed variance in the arrival
  of data at the destination
• - Reliability Error Detection and Error
  Correction Mechanisms used for the transmission
  and processing of MM tasks

13
Resource Management

• Allocation Schemes
• Pessimistic or guaranteed Approach avoids
  resource conflicts by making reservations for the
  worst case that is reserving resource bandwidth
  for the longest processing time and the highest
  rate a task might ever need. Leads to
  Underutilization of resources.
• Optimistic or statistical Reserves resources
  according to the average workload only. If
  Resources are highly utilization this may lead to
  failure

14
CPU Management

• Real time Processing can be achieved through
  efficient real time scheduling.
• The Challenges of MM Scheduling is due to two
  conflicting goals
• Non Real Time Processes should not suffer from
  the execution of real time processes
• Real Time Processes should be allowed to preempt
  non real time processes and other other real time
  processes with lower priorities.
• The Most important real time scheduling
  approaches include EDF and RMS

15
EDF(Earliest Deadline First)

• At Any Interval of a new task, EDF immediately
  computes a new order that is, it preempts the
  running task and schedules the new tasks
  according to its deadlines.
• EDF not only periodic tasks but also tasks with
  arbitrary requests, deadlines and service
  executions
• Upper bound of processor utilization in EDF is
  100.
• EDF assigns priorities according to deadlines.
  The highest priority is assigned to task with
  earliest deadlines and lowest to the one with
  furthest
• Extension divides every task into a mandatory and
  an optional part
• The optional parts are processed if the resource
  capacity is not fully utilized

16
Rate Monotonic Algorithm(RMS)

• Each Task is Preemptive and is assigned a
  priority according to the request rate.
• The Highest priority is assigned to the job with
  the highest rate
• It is an optimal, static, priority-driven
  algorithm for preemptive, periodic jobs
• No other static algorithm can schedule a task set
  that set that the RM cannot schedule
• It assigns static priorities to tasks at the
  connection setup phase according to their
  requests
• Each task is processed with no further
  rearrangement of priorities required
• The following 5 assumptions are prerequisites of
  RMS
• The requests for all tasks with deadlines are
  periodic
• The Processing of Single task must finish before
  the processing of the next task in the same data
  stream
• All tasks are independent
• Runtime for each request of a task is constant
• Any non-periodic task in the system has no
  required deadline.

17
EDF and RMS Scheduling Example
18
EDF vs RMS

• EDF is more dynamic. Since it has to be executed
  frequently it incurs higher scheduling overheads.
  RMS is a static algorithm
• Since priorities are assigned according to the
  request rate, more context switches occur in RMS
  than in EDF
• EDF can achieve processor utilization of up to
  100RMS has an worst case upper bound of about
  69
• Because it has no scheduling overhead and is
  optimal for periodic jobs, RMS is suitable for
  continuous media applications

19
EDF vs RMS
20
MM Support in various Operating Systems
21
MM Applications in Operating Systems
22
Comparison of MM support in Windows and Mac
23
File Management

• Distributed MM applications impose major
  challenges in the design of MM servers including
• Massive storage of huge amount of data
• Real time storage and retrieval performance at
  specified Qos Guarantees.
• High Data Throughput to support large number of
  simultaneous user requests.
• The internals of how the System represents
  information in files and how it accesses those
  files in secondary storage are important to us in
  the design and implementation of the file system.

24
File Management

• File System Manages Disk Bandwidth, Storage
  Space
• Disk Bandwidth Disk Scheduling Algorithms
• Storage Space Placement Techniques.
• Real-time Storage and retrieval CM recording
  devices( such as video cameras) generates a
  continuous stream of media quanta that should be
  stored in real time. CM playback is essentially
  the reverse. ( Need for the same timing sequence)
  Any deviation leads to jerkiness in video
  motion, pops in Audio, or possibly complete
  unintelligibility. Further more need for
  synchronization between two media.
• High Data transfer rate and large storage space
  Digital video and audio playback, require
  efficient store, retrieve, and manipulate data in
  large quantities at high speeds.

25
Approaches to supporting continuous media in file
systems

• a. Organization of the files on disk remains as
  it is, with the necessary real-time support
  provided through special disk-scheduling
  algorithms and enough buffer capacity to avoid
  jitter.
• b. Second approach is to optimize the
  organization of audio and video files, especially
  on distributed hierarchical storage like disk
  arrays. The basic idea being to increase the
  throughput and capacity by storing the data of
  each audio and video file on several volumes.

26
Storage methods

• Contiguous Placement
• Continuous storage of data, requiring just one
  seek to position the disk head at the start of
  the data. Suitable for VOD Servers (WORM
  philosophy).
• Disadvantages Subject to Fragmentation, and
  require enormous copying overheads during
  insertions and deletions to maintain continuity.
• Constrained Placement
• Constrained placement allows the successive
  segments of media stream to be scattered as long
  as the inter-segment separation is bounded by the
  constraint of playback continuity.
• Disadvantages Require elaborate algorithms to
  assure that separation between blocks conforms to
  the required constraints.

27
Storage Methods

• Log-Structured Placement.
• To reduce disk seeks, when modifying blocks of
  data, log-structured systems do not store
  modified blocks in their original positions.
  Instead, all writes are performed sequentially in
  a large contiguous free space. (Suitable for
  extensive editing systems)
• Disadvantage Playback performance is more
  complex to implement.
• Data striping
• Interleave the storage of each media stream among
  disks
• Stripe Unit maximum amount logically continuous
  data that is stored on a single disk.
• Degree of Striping Number of disks across which
  a media stream is striped.
• RAID (redundant array of inexpensive disks)

28
Data placement schemes for multimedia storage
(a) contiguous placement, (b) constrained
placement, and (c) log-structured placement
29
Striped data accessed in parallel.(RAID
Architecture).
30
Disk Scheduling

• Disk Fundamentals
• - Seek Time
• - Rotational Latency
• - Transfer Rate

31
Disk Scheduling Algorithms

• EDF, Earliest Deadline First
• block of stream with the nearest streams are
  read first
• Most often applied as a preemptive scheduling
  scheme
• EDF must be adapted or combined with file system
  strategies
• SCAN-EDF
• Combination of Scan and EDF Mechanisms
• Seek optimization of Scan and the real-time
  guarantees of EDF are combined the request with
  the earliest deadline is always served first, if
  the deadline are same the specific one that is
  first according to the scan direction is served
  first
• SCAN-EDF can be easily implemented

32
Disk Scheduling Algorithms .

• Group Sweeping Scheduling, GSS
• Requests are served in cycles, in round-robin
  manner
• To reduce disk arm movements, the set of n
  streams is divided in to g groups. Groups are
  served in fixed order
• Data must be buffered
• Play out can be started at the end of the group
• Buffer can be reused of each group
• GSS is a trade-off between the optimization of
  buffer space and arm movements
• Mixed Strategy
• Based on the shortest seek( also called greedy
  strategy) and the balanced strategy
• The goal is to maximize transfer efficiency by
  minimizing seek time and latency and to serve
  process requirements with a limited buffer space
• Crucial part is to decide among the two
  strategies

33
Disk Scheduling Algorithms .
34
Memory Management

• Traditional virtual memory with demand paging is
  inappropriate for such real-time processes since
  page faults incur too much swapping time between
  main memory and disk and may effect real time
  performance.
• One solution is to avoid swapping and to lock
  continuous media data in memory during the
  process.
• Another method, is to pre-fetch such data in a
  timely fashion, taking advantage of their
  periodicity.
• Other important practical implementation
  technique include using scatter buffers and
  passing pointers.
• With scatter buffers, the address space of a
  process is loaded into possibly discontiguous
  regions of memory. By passing pointers, objects
  are passed by reference rather than having to
  pass the objects themselves.

35
Device Management.

• device management and the actual access to a
  device allows the operating system to integrate
  all hardware components
• the physical device is represented by an abstract
  device driver
• the physical characteristics of devices are
  hidden
• existing operating system extensions for
  multimedia usually provide one common system-wide
  interface for the control and management of data
  streams and devices
• the multimedia extensions of Microsoft Windows,
  for example, provide the following classes of
  function calls
• 1) system commands are not forwarded to a single
  device driver (MCI driver) they are served by a
  central instance
• 2) each device driver must be able to process
  compulsory commands
• 3) basic commands refer to characteristics that
  all devices have in common
• 4) extended commands may refer to both device
  types and special single devices
• for example camera control calls

36
System Architecture

• - the employment of continuous media in
  multimedia systems also imposes additional new
  requirements to the system architecture
• - a typical multimedia application does not
  require processing of audio and video to be
  performed by the application itself
• - data are obtained from a source and are
  forwarded to a sink
• - in such a case the best solution is "The
  shortest possible path" through the system, i.e.,
  to copy data directly from adapter to adapter
• - in multimedia systems, such an adapter to
  adapter connection is defined by the capabilities
  of the two involved adapters and the bus
  performance
• - most of today's multimedia systems must coexist
  with conventional data processing they share
  hardware and software components
• - protocol processing is the bottleneck because
  it cannot provide the necessary throughput
• - the paradigm of streaming from source to sink
  is an appropriate way of doing this
• - the multimedia application opens devices,
  establishes a connection between them, starts the
  data flow and returns to other duties
• - the most dominant characteristic of multimedia
  applications is to preserve the temporal
  requirement at the presentation time

37
Example of Distributed Multimedia System
38
Synchronization in Distributed Multimedia Systems

• Stream Synchronization
• This refers to specific temporal relation between
  the source and sink of a single communication.
• In the above example, for each individual video
  and audio play out it is necessary to maintain an
  internal stream synchronization between the
  sender and the receiver.
• Continuous Synchronization
• This refers to an ongoing commitment to a
  repetitive pattern of temporal linkage between
  two or more communications.
• In the above example, it is required to maintain
  an ongoing continuous synchronization
  relationship, between the associated video and
  audio play outs.
• Event based Synchronization
• This refers to the act of notifying that a
  relevant event or a set of events has taken
  place, and then causing an associated action or
  actions to take place. This must be done in a
  timely manner.
• In the above example, it is required to display
  text captions and voice annotations at the
  correct time, with respect to progress of the
  video display.

39
Future Research Ideas

• Moving Toward a Multimedia Web OS Integrating
  multimedia environment on the Web that includes
  new audio, video and 3D technologies.
• Integration of real-time processing and non-real
  time processing in the Native System Kernel.
• Improving Scheduling Algorithms to support
  Multimedia Objects in conjunction with
  traditional applications more efficiently.
• Synchronization Techniques of Multimedia In a
  Distributed Environment.
• Integrating Database Operations In Multimedia
  Operating Sytems Future Search Engines
  Would deal not only text , but also - Audio,
  Image, Video

40
References..

• http//mig.msu.edu/Followups/OS/breakdown.html
• http//www.beos.com
• http//www.computer.org/multimedia/mu1995/u1068abs
  .htm
• http//www-inf.int-evry.fr/bernard/distributed_sy
  stems/node5.html
• http//citeseer.nj.nec.com/details/leslie93pegasus
  .html
• http//www.cl.cam.ac.uk/Research/SRG/netos/pegasus
  /papers/usenix/usenix.html
• http//www.cl.cam.ac.uk/Research/SRG/nossdav98/
• http//citeseer.nj.nec.com/did/4545
• --P. Goyal, X. Guo, and H. M. Vin. A hierarchical
  CPU scheduler for
• multimedia operating systems. In Proceedings of
  the Second Symposium on
• Operating Systems Design and Implementation
  (OSDI). USENIX, October 1996
• --Ralf Steinmetz, "Analyzing the Multimedia
  Operating Systems," IEEE
• Multimedia, Vol. 2, No. 1, Spring 1995, pp.
  68-84.
• --K. Nahrstedt, R. Steinmetz, "Resource
  Management in Networked
• Multimedia Operating Systems," IEEE Computer, May
  1995, pp 40--49.
• --D.J. Gemmell, H.M. Vin, D.D. Kandlur, P.V.
  Rangan, and L.A. Rowe.
• "Multimedia Storage Servers A Tutorial". IEEE
  Computer, 28(5)40--49, May
• 1995.
• --V. O. Li and W. Liao, "Distributed multimedia
  systems," Proceedings of

41
Questions and Answers..