Introduction and Architectures

1
Introduction and Architectures
INF SERV Media Storage and Distribution Systems

• 29/8 - 2002

2
Overview

• Intro
• about the course
• multimedia applications and challenges
• Architectures
• Media (Video) on Demand
• Machine internals
• Video server structures
• Examples

3
INFSERVThe Course
4
Content of INFSERV I
architectures
file systems
media data
distribution
protocols
resource scheduling
topologies
5
Content of INFSERV II

• System architectures (server and system
  designs)
• Media data (wire and file formats, codecs)
• Media characteristics and user behavior
  (processing multimedia data, user
  interactivity)
• Resource management (CPU and memory management)

6
Content of INFSERV III

• Protocols with Quality of Service (QoS)
  (specific and generic QoS approaches)
• Adaptation and mobility (various clients,
  different amount of resources)
• Storage systems (management of multimedia files,
  retrieval)
• Distribution(use of caches and proxy servers)

7
Content of INFSERV IV

• Student assignment
• either
• theoretical (literature) exercise find, read
  and present papers
• or
• practical exercise programming and software
  benchmarking
• assignment must be reported and presented to the
  class at the end of the course independent of the
  type chosen

8
Pensum INFSERV

• Approved presentation of student assignment
• Oral exam (possibly 28-29/11)
• all transparencies from lectures
• content of own student assignment

9
Applications and Challenges
10
Applications

• Multimedia enriches the user interface giving new
  applications
• Broadcast server is only VCR substitute (e.g.,
  MTV Europe)
• Media-on-Demand (MoD)
• Video-on-Demand (VoD) (e.g., Bell Atlantic)
• News-on-Demand (NoD) (e.g., CNN, BBC)
• Learning-on-Demand (LoD)
• Virtual worlds
• video conferencing (e.g., USITs electronic
  classroom)
• games (e.g., Quake, )

11
Requirements

• Application
• QoS time sensitivity
• resource capabilities support interactive
  streaming of multimedia content
• Business
• scalability
• reliability
• Architectural
• topology
• cost performance

12
Technical Challenges

• User end system
• real-time processing of streams (1000 MIPS for
  an MPEG-2 decoder)
• request/response delay (lt 150 ms for
  videophones)
• Storage
• real-time retrieval of contiguous media streams,
  e.g.
• 4000 movies 90 minutes 15 Mbps (HDTV)
  40.5 TB
• 2000 CDs 74 minutes 1.4 Mbps 1.4 TB
• Network
• real-time transport of contiguous media data

13
Media-on-Demand (MoD) Systems

• Classification parameters
• structure of movies
• interaction
• presentation form
• Common directions
• analog ? digital media
• distribution ? interaction
• media broadcasting ? media multicasting ?
  personalized media
• linear movies ? branched movies ? variable
  movies
• Evolution
• broadcast - traditional, no user control
• pay-per-view - limited interactivity
• quasi MoD - distinction into user groups, limited
  timely control
• near MoD - same media distributed in intervals
• true MoD - full user control, VCR capabilities,
  bidirectional

14
Television (Broadcast)
time
channels
sender
receiver(s)

• analog or digital
• one program per channel
• linear movies

15
Near Video-on-Demand (NVoD)
time
channels
sender
receiver(s)

• analog or digital broadcasting
• one program over multiple channels
• time-slotted emission of the program
• linear movies

16
(True) Video-on-Demand (VoD)
time
movies
sender
receiver(s)

• digital uni- or multicasting
• control channels
• linear movies

17
Comparison TVoD vs. NVoD
18
Interactive Vision
time
movie
sender
receiver(s)

• digital uni- or multicasting
• control channels
• fixed non-linear movies

19
Cyber Vision
time
sender
receiver(s)

• digital uni- or multicasting
• control channels
• variable non-linear movies, e.g.,
• - games, virtual reality,

20
Application Classification Overview
Cyber Vision
variable
Interactive Vision
movie structure
branched
HDTV
VoD
digital
TV NVoD
linear
presentationform
unidirectional
VIDEO
interaction
analog
bidirectional
21
Media (Video) on Demand
22
Challenges

• VoD in LANs is solved OVERPROVISIONING works
• established in studio business
• established in hotel/hospital/plane/ business
• VoD in WANs
• goals
• network-based distribution of media content to
  consumers
• bring control to users
• assumptions
• overprovisioning of resources will NOT work
• no central control of delivery system
• programs
• need for interoperability not from a single
  source
• need for co-operative distribution systems
• amount of data
• estimated 65000 movies made in 1995 ? 260 TB
  MPEG-2 data
• additionally, data from TV-series, sport clips,
  news,
• historically
• much attention as interactive TV (ITV) some
  years ago
• many (not successful) field trials

23
ITV Network Architechture Approaches

• WAN backbones
• SONET
• ATM
• Local distribution network
• ADSL (asymmetric digital subscriber line)
• FTTC (fiber to the curb)
• FTTH (fiber to the home)
• HFC (hybrid fiber coax)
• Internet WAN
• Diffserv over MPLS (multi-protocol layer
  switching)
• point-to-point Gbps ethernet
• Internet local
• IP over the old distribution networks

ATM / SONETbackbone network
wireless
Internet based systems
ATM
telephone
cable
24
Concerns Internet-Based VoD Systems

• Can technical problems be mastered?
• broadband communication to every home
• user-friendly end systems
• server technology
• Market success?
• what prices will consumers accept?
• high equipment (HW SW) costs
• data costs
• will it be competitive?
• to existing TV programs
• to video rentals
• what is the consequence if ITV field trials in
  USA and Europe?
• no big success
• only a few private consumers willing to
  participate
• trials were cancelled or switched from TV-based
  to PC-based platforms

25
Driving Forces

• Hardware/software (IT) companies
• computer (e.g., IBM, HP, Sun, Microsoft, )
• consumer electronics (e.g., Sony, Philips, )
• Network companies
• telephone (e.g., Telenor, Telia, BT, ATT, )
• cable TV (e.g., UPC, Time-Warner Cable, )
• Content companies
• media
• movies (e.g., Time-Warner, Disney, Paramount, Leo
  Kirch, )
• TV programs (e.g., NRK, TV2, TV3, RTL, )
• hyper media information bases (e.g., Springer,
  Bertelsmann, )
• home-shopping (e.g., Quelle, OTTO, )
• video games (e.g., Nintendo, Sega, Microsoft, )

26
VoD Deployment Status I

• Digital Video Broadcast (DVB)
• no VoD
• cable, antenna, or satellite broadcast
• some NVoD scheduling approaches (e.g., TV 1000)
• Digital Audio-Visual Council (DAVIC)
• defines interfaces only
• no standardization of algorithms for
  interoperation
• Internet Engineering Task Force (IETF)
• currently no large-scale, wide-area video
  distribution
• sporadic use of cooperative web caching
• starting AV caching considerations
• defines protocols and inspires interoperability
  testing

Broadcast world suited for Near
VoD Broadcast world suited for True
VoDInternet world suited for True VoD
MPEG 4 suited for Interactive Vision
27
VoD Deployment Status II
28
VoD System Architecture
local distribution network
backbone network
local distribution network
local distribution network
29
VoD Storage Hierarchy
completeness of available content

• Popularity of moviesnot all are equally popular
  most request directed to only a few (Zipf
  distribution)
• Use hierarchies
• Straight forward hierarchy
• popular videos replicated and kept close to
  clients
• locality vs. communication vs. server costs

30
VoD Components

• Servers
• Networks
• backbone
• local networks
• Intermediate nodes
• routers
• proxy cache servers
• replica servers
• End-systems
• PCs
• TV sets with set-top boxes

31
Traditional Server Machine Internals
32
General Operating System Structure and Data Path
user space kernel space
33
Example Intel Hub Architecture (850 Chipset)
I
Intel D850MD Motherboard
RDRAM connectors
CPU socket
RDRAM interface
system bus
hub interface
PCI bus
Memory Controller Hub
I/O Controller Hub
PCI connectors
34
Example Intel Hub Architecture (850 Chipset)
II
Notethese transfers only show data movement
between sub-systems. Additionally, data touching
operations within a sub-system will require that
data is moved from memory and to the CPU, e.g.
- checksum calculation - encryption - data
encoding - forward error correction
Pentium 4 Processor
registers
cache(s)
RDRAM
RDRAM
RDRAM
RDRAM
PCI slots
PCI slots
PCI slots
35
Example IBM POWER 4
POWER 4 chip
core interface switch
L2
fabric controller
file system
L3
memory controller
GXcontroller
L3 controller
communication system
application
network card
PCI slots
remote I/O(RIO) bridge
PCI host bridge
PCI-PCI bridge
disk
PCI slots
PCI host bridge
PCI-PCI bridge
NoteAgain, data touching operations add
movement operations
36
Memory Hierarchies

• We cant access the disk each time we need data
• Typical computer systems therefore have several
  different components where data may be stored
• different capacities
• different speeds
• less capacity gives faster access and higher
  cost per byte
• Lower levels have a copy of data in higher
  levels
• A typical memory hierarchy

tertiary storage (tapes)
price
speed
secondary storage (disks)
main memory
capacity
cache(s)
37
Storage Costs Access Time vs Capacity
typical capacity (bytes)
access time (sec)
from Gray Reuter
38
Storage Costs Access Time vs Price
dollars/Mbytes
access time (sec)
from Gray Reuter
39
Internal Server Design

• Stream retrieval from disk and push to network
• buffer requirements
• bus transfers
• CPU usage
• concurrent streams (can be merged??)
• storage (disk) system
• scheduling ensure that data is available in
  time
• block placement contiguous, interleaving,
  striping
• Stable operations
• redundant HW
• multiple nodes
• Much more, e.g., caching/prefetching, admission
  control,

40
Video Server Structure
41
Video ServerServer Components Switches
Tetzlaff Flynn 94

• Internal content directory

• External content directory

42
Video Server Components
incomingresolve request
incomingdata request
delivered resolution
delivered data
network attachment
content directory
control/application server
data server
43
Video ServerSimple General Server Architecture
Sitaram Dan 00

• Storage subsystem
• stores data
• different devices

• Network subsystem
• transmit MM data

• Processor subsystem
• executing part
• management and operations

storagesubsystem
network subsystem
processor subsystem
control server
data server
clients
application server

• Application server
• user interface
• billing
• content database
• user database
• service gateways

• Data server
• data delivery
• specialized file system
• buffer manager
• data importer/exporter

• Control server
• administrator
• admission control
• optimization

44
Video ServerDirectory Access Data Retrieval

• Two-step retrieval
• problemresource management
• Request redirection
• problemclient gets data from another machine

45
Video ServerDirectory Access Data Retrieval

• Pull model
• client sends several requests
• deliver only small part of data
• fine-grained client control
• favors high interactivity
• suited for editing, searching, etc.
• Push model
• client sends one request
• streaming delivery
• favors capacity planning
• suited for retrieval, download, playback, etc.

46
Video ServerServer Topology I

• Single server
• easy to implement
• scales poorly
• Partitioned server
• users divided into groups
• content assumes equal groups
• location store all data on all servers
• load imbalance

47
Video ServerServer Topology II

• Externally switched servers
• use network to make server pool
• manages load imbalance(control server directs
  requests)
• still data replication problems
• (control server doesnt need to be a physical box
  - distributed process)
• Fully switched server
• server pool
• storage device pool
• additional hardware costs
• e.g., Oracle, Intel, IBM

48
Video Distribution ServerTypical In the
Internet Today

• Push systems(pull in video editing/database
  systems)
• Traditional (specialized) file systems not
  databases for data storage
• No in-band control (control and data information
  in separate streams)
• External directory services for data
  location(RTSP/control server data pump)
• Request redirection for access control
• Single stand-alone servers ? (fully) switched
  servers

49
Server Examples
50
Video Server Research Status
Härtig,Sitaram,Dan,Nahrstedt,Steinmetz, Klas,
Shulzrinne,Coulson,Seltzer,Rangan,Zhang,Hutch
inson,
Research covering allcomponents is rare
51
Video Server Product Status
1) Real server, VXtreme, Starlight, VDO, Netscape
Media Server,MS Media Server, Apple Darwin
1) Real server, VXtreme, Starlight, VDO, Netscape
Media Server,MS Media Server, Apple Darwin
2) IBM Mediastreamer, Oracle Video
Cartridge, N-Cube
2) IBM Mediastreamer, Oracle Video
Cartridge, N-Cube
3) SGI/Kassena Media Base, SUN Media Center,
IBM Video Charger
3) SGI/Kassena Media Base, SUN Media Center,
IBM Video Charger
52
Real Server

• User space implementation
• one control server
• several protocols (1-3)
• several versions of data in same file
• adapts to resources
• Two formats
• Reals own
• MPEG-2 version with stream thinning(dropped
  with REAL ?)
• Does not support
• management
• Quality-of-Service
• load leveling

1
2
3
RTP/ RTCP
Reals protocol

UDP
TCP
IP
53
IBM Video Charger

• May consist of one machine only, or
• several IBMs Advanced Interactive eXecutive
  (AIX) machines
• Servers
• control
• data
• Lightly modified existing components
• OS AIX4
• virtual shared disks (VSD)(guaranteed disk I/Os)
• Special components
• TigerShark MMFS(buffers, data rate, prefetching,
  codec, ...)
• stream filters, control server, APIs, ...

specificcontrol server
RTSP
video stream API
mlib API
RTP
encrypt
filter

TigerShark MMFS
UDP
VSD
IP
54
IBM Mediastreamer

• Version of Video Charger
• failed project to guarantee MPEG-2 over ATM and
  Cable
• similar machine setup(machine cluster)
• special HW Mowgli
• SCSI controller
• MPEG-2
• ATM or analog cable out
• moved to Video Charger
• Unlike Video Charger, Mediastreamer runs on old
  IBM machines due to special HW
• Special components
• TigerShark
• control server
• APIs
• special board

55
nCUBE

• One server scales from 1 to 256 machines, n ? 0,
  8
• Special components
• board with SCSI, connectors, HAVOC vector
  processor, etc.
• TRANSIT operating system
• Interface modules
• ATM
• digital video broadcast (DVB)
• ethernet
• QAM
• Real Networks agreed to integrate nCUBE's n4x
  with RealSystem 8 ??

8 hypercube connectors
configurable interface
memory
PCI bus
vector processor
SCSI ports
56
Small Comparison
57
Summary

• Multimedia applications and challenges
• Media (Video) on Demand
• Machine internals
• Video server structures
• Video server examples

58
Some References

• Sitaram, D., Dan, A. Multimedia Servers
  Applications, Environments, and Design, Morgan
  Kaufmann Publishers, 2000
• Tendler, J.M., Dodson, S., Fields, S. IBM
  e-server POWER 4 System Microarchitecture,
  Technical white paper, 2001
• Intel, http//www.intel.com
• nCUBE, http//ncube.com

59
Assignments

• Practical assignmentsEach
  practical assignment is performed by up to two
  persons (and morewhere noted). Each person will
  present his work before the class thiswill
  typically be achieved by presenting the approach
  to the practicalsolution in a first talk, and
  the results in a second. Starting a practical
  assignment comprises understanding of the
  assignmentin sufficient detail, getting used to
  the appropriate developmentenvironment for the
  investigation, and acquiring the required
  softwaretools.
• Testing IXP cards This task is pending the
  arrival of the IXP cards. It is meant to provide
  insights into the use of the Intel IXP cards
  capabilities, programming interfaces, testing
  tools and so on.
• Testing TCP implementations Linux has
  implemented Reno TCP, which includes Fast
  Retransmit and isgenerally more efficient than
  BSD's Tahoe TCP because it does not enterslow
  start for all packet losses. In newer Linux
  revisions, the SACK(selective acknowledgement)
  and FACK (forward acknowledgement) are
  alsosupported as runtime-configurable options.
  Test the behavior of these implementations in
  different loss/delaysituations. Consider
  streaming applications such as progressive
  downloadin the design of your tests.

60
Assignments

• Implementation of memory caching Usually,
  caching is not necessary in applications
  involving videotransfers due to high data rate,
  i.e., data is replaced before it can bereused.
  However, for very popular data, caching can
  beneficial. In theINSTANCE OS enhancements, we
  have implemented a buffer manager
  withoutcaching. In this assignment, one should
  implement a simple cachingmechanisms and test
  the system with a video server workload (Zipf)
  withand without the caching mechanism.
• Performance of IPv4 vs. IPv6 Both IPv4 and
  IPv6 are implemented in Linux network stacks.
  Thecapabilities of IPv6 are far beyond those of
  IPv4, not only in terms ofaddress range, but
  there are also preparations for mobility
  support,automatic configuration and so on.
  Investigate the efficiency of both stacks by
  adding probes to thenetwork stack. What is the
  processing time for a package in IPv4 and
  IPv6under various conditions. Consider the Raw,
  TCP and UDP performance.

61
Assignments

• Usefulness of replacement strategies in komssys
  Data replacement is the central part of all
  caching strategies, becauseit decides not which
  data to store, but which to remove from storage
  whenthe available space gets filled up. Data
  replacement for high rate video data must be
  different than fortraditional low volume,
  discrete data. In this assignment, one
  shouldimplement several algorithms in KOMSSYS
  and monitor the number ofreplacement on a video
  server workload. To be tested with a Zipf
  workload.
• Implementation of multicast/broadcasting schemes
  in KOMSSYS In contrast to the typical on-demand
  video servers, near video-on-demandsystems have
  actually been deployed in vast numbers (e.g. TV
  1000). Theschedules of these existing services
  are very simple and not efficient fordelivery
  over the Internet, and research approaches have
  been developedfor years, starting with
  batching. In this thesis, our example code for
  batching should be replaced withone of the more
  advanced strategies (pyramid broadcasting,
  harmonicbroadcasting, skyscraper, patching,
  HMSM, ...). The performance (in termsof packets
  that are actually sent to a "dummy network" card)
  of the twoshould be compared with a standard
  Zipf workload.
• Packet scheduling Linux networking support
  provides several optional packet schedulers
  askernel modules that can be used to support
  load levelling for routers, butmay also be
  feasible to use in servers. Test several of
  these schedulers, and compare them with
  severalworkloads that can be generated by a
  streaming media server. Workloadsshould
  generally consist of long-lived streams of
  various bitrates.

62
Assignments

• RTSP in C The Real-Time Streaming Protocol is
  a meant to control streamingoperations such as
  start, pause, stop or seek, and looks similar
  toHTTP. It is the established standard for this
  task in audio and videoservers, but their are
  few reusable open source implementations. Port
  our existing C RTSP parser to C, to make it
  available as alibrary for general use.
• Compare and test Real vs Darwin servers Real
  has recently released a source code version of
  its servers, calledthe Helix server (trial
  version), and Apple's Quicktime server is
  partlydeveloped in the open source under the
  name Darwin. Both servers supportstandard-complia
  nt protocols such as RTSP and RTP. Install both
  systems, compare their structure, and compare
  theirresource efficiency and latency in
  answering user requests.

63
Assignments

• Linux file system benchmarking Identify
  different benchmarks for file systems, and choose
  one or more to test different file systems (e.g.,
  ext2, ext3, reiser, jfs, ) under Linux.
  Install and test the file systems with respect to
  multimedia server performance varying different
  parameters

64
Assignments

• Theoretical assignmentsEach
  theoretical assignment is performed by a single
  person.Theoretical assignments begin with a
  paper in the topic area that willintroduce the
  relevant problem and probably propose a solution.
  This papermay be early work, or an older
  snapshot of the state-of-the-art. It isnot meant
  to contain the entire answer to the assignment.
  Rather,literature study is meant to be an
  important element of theassignment. More recent
  papers and their results are supposed to be
  foundand read by the student before the own
  presentation and delivering therequired report.
  During the course of the lecture, the time
  1100-1200 will in someweeks be assigned to
  presentations and discussions of the assignments.
  Wewill discuss issues such as literature search
  and structuring of researchreports. We offer
  students the opportunity to discuss papers that
  they havediscovered in their search for related
  work concerning theirappropriateness for the
  assignment.gt -memory caching ??gt -compare
  zero-copy approaches ??gt -compare scheduling
  approaches ?? (disk/CPU/network)gt -hvordan
  routes MM traffic ??

65
Assignments

• Single Machine Storage Hierarchies Computers
  tend to have different bottlenecks at different
  times of thedevelopment cycle. Right now,
  network bandwidth is a minor issue formultimedia
  applications, as is raw computing power. The most
  problematicare currently RAM access speed and
  bus speed, but since expensive,proprietary
  solutions exist, we can expect a new shift of the
  bottleneckin the near future. The cost of
  storage was a major problem in early
  videoservers, and third-level storage such as
  tape drives and (then slow)optical storage were
  taken into account for the design of
  scalablesystems. We can expect these
  multi-hierarchy servers to come backconsider
  RAM, Flash-RAM disks, hard disks, and network
  attached storage asa possible 4-step hierarchy.
  Look at Doganata and Tantawi's book
  chapter"Storage Hierarchy in Multimedia Servers"
  for a start and explain theirinvestigation of
  the internal hierarchy. What is special for
  their case? Which later investigations were
  made and what were the results? Why havetapes
  vanished from media servers?
• Capacity Analysis Video-on-demand in the early
  days held no concept of efficientdistribution
  mechanisms. With the dream of ATM to the desktop
  andfiber-in-the-curb, sufficient bandwidth for
  the delivery of individualstreams for each
  request was expected. Nussbaumer, Patel and
  Schaffaanalysed the server requirements in 1994
  in a straight-forward manner, in"Multimedia
  Delivery on Demand Capacity Analysis and
  Implications". Thispaper has motivated some
  serious research on the internal scalability
  ofsingle servers. Which conclusions did
  researchers draw from this study? In which
  directions did their research go?

66
Assignments

• Compare different commercial systems A number
  of commercial systems have been produced since
  the late 90s,and some have survived. Others
  haven't but still provide insight into
  theapproaches that were taken to make multimedia
  servers (or, in most cases,streaming video
  servers) feasible and scalable. Some examples
  arepresented briefly during the lecture. Find
  the existing/remainingtechnical white papers and
  research papers that document these
  servers,classify them. Write down and present
  their features and a feturecomparison.
  Examples - Kassena Mediabase (was SGI
  Mediabase) - Apple Streaming Server (same as
  Darwin Open Source Streaming Server) - Real
  Server (also Real Helix Server) - NCube - IBM
  VideoCharger (also Ultimedia Server,
  Mediastreamer-0) - Oracle Video Server - SUN
  Media Center (withdrawn) - Microsoft Tiger
  (never sold)
• Guaranteed QoS Guaranteed QoS is a phantom
  idea in Internet networking. It is to
  beeverything from harmful to fairness and
  equality among users, or vital forcommercial use
  of the Internet. Whatever the case, to achieve
  guaranteedQoS, it must be supported in LANs, and
  unmanaged standard Ethernet is mostcertainly
  unfit to provide it. Nagarajan and Vogt did the
  math foradmission control in Token Ring networks
  in "Guaranteed PerformanceTransport of
  Multimedia Traffic over the Token Ring". How
  does their model work? Which other approaches
  have been devised to guarantee QoS for
  otheraccess network technologies (examples are
  Ethernet, FDDI, FDDI-2, QPSX,DQDB, IP over ATM,
  Frame Relay)?