Tuesday, May 7, 2013
INTRODUCTION
Transmission of multimedia data over a packet-switched
network typically requires resource reservation to guarantee an acceptable
level of performance (e.g., throughput or delay). In this article we address
the problem of how to make such real-time communication reliable. First of all,
it is essential to bound the duration of service disruption caused by failures
to a reasonably little value. Considering the large volume of multimedia data,
minimizing the fault- tolerance overhead is also important. Furthermore as more
applications with different dependability requirements share the same network,
the level of dependability for a given application should be “customizable”,
depending on the criticality of the application. We first survey the existing
approaches, and then present our scheme which is developed in accordance with
three design goals: fast failure recovery, low fault tolerance overhead, and
per-connection reliability guarantee. Our scheme provides and integrated
solution covering such issues as connection establishment, failure detecting,
runtime failure recovery, and resource reconfiguration.
REAL – TIME COMMUNICATIONS
Real time transport of continuous media (video and audio)
is achieved through circuit switching in telephone services or by broadcasting
over shared media in television services. The end-to-end performance is
necessary to achieve required functionally of these application (in real time
applications) is often called end-to-end quality of services (QOS). Today’s
representative computer network, Internet also lacks QOS support for continuous
media applications.
However, many protocols such
as RTP [1] XTP [1] and IP multicast which are deployed on Internet to achieve
QOS. But these protocols do not meet the true multimedia requirements because
they only support a best effort service model. As the demand for real-time
communication services in recent years, numerous QOS models were developed
ranging from constant bit rate (CBR) services, which resembles telephony
service to the “controlled-load” service which mimics the best-effort service
in unloaded network. As QOS is our main concern n real-time communications,
they rely on some form of resource reservation and admission controls. They
share three common properties.
QOS
contracted
Connection-oriented
Reservation-based
A contract between a client and the network is established
before the client’s messages are actually transferred. The client must first
specify his input traffic behaviors and required QOS. The network then computes
the resource needs (e.g. link and cpu bandwidth, buffer space) from this
information, selects a path, and reserves necessary resource along the path. If
there are not enough resources to meet the client’s QOS requirements., the
request is rejected. The client’s data messages are transported only via the
selected path with the resources reserved, and this virtual circuit is often
called a “Real-Time Channel”.
NETWORK DEPENDABILITY
Some application requires both dependable and timely
communication services. Example for such real-time multimedia communications
include remote medical service, Collaborative scientific research, business net
meeting, battle field command/control etc. in some applications network
availability, ie. Probability of connection being available at any given time
is they dependability QOS measure.
Network failures can cause even larger-scale social
disasters. Example, a fire at an unmanned tall office building Illinois caused
3.5 million telephone calls to be blocked in 1988. in 1990’s several similar
accidents have been reported for various reasons, such as damage of fiber cable
caused by construction , earthquake, network overhead etc. thus even though
failures occur rarely, the consequences
of mishandling failures could be
devasting, thus making network reliability concern.
The current Internet with data gram services has
successfully dealt with two types of network failures: Transient and
persistent. An example of transient failure is temporary losses due to network
congestion or data corruption. There persistent failure includes the breakdown
of network components. Transient failures and dealt by TCP which can handle
transient loss of packets by acknowledgement and retransmission. On the other
hand persistent failures are dealt by IP protocol along faulty networks by
routing the packet. But retransmission is unlikely in real-time systems because
there is usually not enough time o deletes and retransmits the lost real-time
message before its dead line expires. And also as reserving resources on a
fixed path and transporting real-time messages via that path realize QOS
guarantee, unlike data gram messages it cannot be derouted.
Hence the persistent failure also causes serious failure.
The prevalence of optical fibers affects network dependability. The probability
of transmission errors in optical links becomes negligible. Hence the chance of
packet loss due to transmission errors is very rare. Hence for real-time
systems transient failures are relatively less important because
cohesion-induced packet losses can be avoided by resource reservation. But in
persistent failures a single failure or link lads to the loss of large number
of connection. Not only link failure but also node failures are to be carefully
dealt with. Moreover computer networks are more vulnerable to viruses or
hacking. So, development of effective mechanism is must to cope up with network
failure is a must.
DESIRABLE FEATURES
To design fault-tolerant service, one must define the modal
of failure to be tolerated. Some applications can tolerate slow failure
recovery but require reliable delivery of messages even if it takes long time
such as e-mail and file transfer. Some applications require fast failure
recovery but loss can be tolerated. Real-time multimedia applications fall to
this category, as loss of couple of frames in video/voice data streams may be
acceptable. Let us assume transient packet losses are acceptable to
applications and are dealt by FEC (forward error checking), and focus
effectively on how to handle persistent failures. We shall discuss about
“channel failure”.
A real-time channel is said to have failed if the rate of
correct (content and timing) message delivery within a certain time interval is
below threshold specified by the application. There are 5 criteria that
characterize a good solution.
1.pre-connection dependability
guarantee: The network should provide guarantee on dependability for each
connection, so that successful recovery is guaranteed as long as failure occurrences
do not exceeded the fault tolerance capability of the connection.
2. Fast-failure recovery: The service
disruption time of a connection caused by failures should be bounded to
reasonable small value.
3. Small fault-tolerance overhead: The additional
resources overhead for fault-tolerance should be acceptably low.
4. Robust failure handle: Failures should
be handled robustly even though failure occurrences may exceeder assumed
failure hypothesis.
5. Interoperability/scalability: The
failure recovery scheme must be interoperable with various existing and future
real-time channel protocols.
EXISTING APPROACHES
The work has done on
real-time communications and approaches have been developed. Some of them are:
5.1 Reactive method: Simplest way of recovering
real-time channel failure is to be establishing a new real-time channel, which
includes failure components. This scheme relies on the broadcast of all
component failures to the entire network so that all hosts can maintain a
consistent view of the current topology.
Advantage: No fault-tolerance
overhead in the absence of failure recovery.
Disadvantage: The channel
re-establishment attempt can fail due to resource shortage at a particular
time.
5.2 Failure masking: Here multiple copies of
message are sent simultaneously over disjoint paths.
Advantage: The method attempts to
achieve both timely and reliable delivery of message at the same time. Both
persistent and transient failures can be handled.
Disadvantage: It is very expensive due
to multiple copies of same message. Instead of transmitting entire message,
each message is broken into equal size sub-message, which is then transmitted
over different paths for FEC.
5.3 Single-Failure Immune (SFI): In this approach,
cold-standby resources are reserved for fault tolerance.
Advantage:
1. Guarantees failure recovery. In this additional
resources are reserved in vicinity of each real-time channel at the time of
channel establishment.
2. The advantage of this
cold standby approach is that although additional resources need to be
reserved, the resources reserved for fault tolerance can be utilized by best
effort in the absence of failure.
TELEPHONE NETWORKS
In old telephone network, a
true electric circuit through electro-mechanical or pure electrical exchanges
connected two phones. Now, telephone networks are very close to computer
network. A modern switching node in telephone networks is almost a
general-purpose computer equipped with high fault-tolerance capability and
powerful I/O capability. Techniques for telephone service resemble those for
real-time communication services in packet switched network in both services rely
on similar principles such as dedicated resources and static routing. Whenever
a telephone connection is broken down it is detoured. Failure recovery should
be fast so that users hardly notice disruption caused by the failure. The
successful fault recovery is also important. if no enough resources are
available for re-routing al affected connections, some of them be dropped . To
avoid resource shortage by rerouting spare resources are reserved in advance.
For rerouting there are 2 strategies.
6.1 Span Restoration (Local
Rerouting): This is used in synchronous transfer mode
networks. Here a “maximal flow” model is used to find the optimal placement of
spare resources under deterministic failure hypothesis, typically a single link
failure. A drawback of the local rerouting is the resource usage becomes
inefficient after failure recovery because channel paths tend to be lengthened
by local detouring.
6.2 Path Resolution
(End-to-End Rerouting): There are two variations in this strategy depending o whether the
failure recovery paths are pre computed before failure occurrence or
determining after failures actually occur. In the former approach, the
pre-routed recovery path should be disjoint with the original connection path,
while the later the recovery paths can use the healthy components of their
original connection paths. The former has an advantage over the secondary in
terms of dependability guarantee.
COMPARISION OF
EXISTING APPROACHES
The latest approach uses end-to-end re-routing with
pre-computed recovery paths. We set up one or more backup channels in advance
in addition to each primary channel. Upon failure of primary channel, one of
its backups is prompted to a new primary channel. There two main differences
between path restoration and latest approach.
|
Recovery Method
|
Recovery
overhead
|
Recovery
delay
|
Recovery guidance
|
Reactive
|
No
|
Long
|
No
|
|
SFI
|
High
|
Shorter
|
Deterministic
|
|
Multicopy
|
VeryHigh
|
No
|
Flexible
|
|
Span restoration
|
Low
|
Shorter
|
Deterministic
|
|
Path restoration
|
Lower
|
Short
|
Deterministic
|
|
Our approach
|
Lower
|
Short
|
Flexible
|
1. All connections are treated
equally under the same failure model in path restoration and in contrast the
latest approach allows per-connection fault-tolerance.
2. In path restorations connection demands are
known at the time of network design and change very rarely. Hence this method
cannot be applied to an environment where short-lived channels are setup and
torn-down frequently. In contrast the latest approach needs only the
information that can easily be obtained at run-time i.e.
(a) Any algorithm may select a
backup path.
(b) Space resource allocation
may be done with the given routing results.
CONNECTION ESTABLISHMENT
A backup channel does not consume any bandwidth in normal
situation, as it does not carry any data until it is activated. However a
backup channel is not free since it requires the same amount of resources to be
reserved as its primary channel in order to provide the same quality of service
upon its activation. But backup channels are too expensive to be useful for
multimedia networking.
This resource sharing
technique called backup multiplexing was developed. By this we reserve only a
very small fraction of link resources needed for all channels going thro the
link. With backup multiplexing backup channels are over booked by a Meta
admission, test, in which some existing backup channels are not accounted for
in the admission test of new backup channel. Our strategy is to multiplex those
backups, which are less likely to be, activated backup channel Bit. Bit of two
different connections i.e. the probability of simultaneous failures of their
respective primary channels is given by S (Bi, Bj). Bi & Bj are multiplexed
if S (Bi, Bj) is smaller than a certain threshold V, called the multiplexing
degree, which is specific to each backup. More accurately if S (Bi, Bj) <
Vi, Bj can be multiplexed with Bi. The smaller the V of a backup, the higher
fault tolerance will result, since, fewer backups will be multiplexed with it.
FAILURE DETECTION
Effective failure detection with high coverage and low
latency is essential for failure recovery. Instead of adopting expensive
failure reduction techniques. We use behavior-based detection techniques that
do not require special hard work support and hence can be used in any network.
1. End-to-End method and
2. Neighbor detection method.
1. End-to-end detection
involved both the source and destination notes of a real time channel. The
source node regularly injects “Channel hear beats” i.e. a sort of real time
message into the channel message stream and the intermediate modes on a
channels do not discriminate channel heart beats from data messages. The
destination note can monitor the no. Of data messages lost as the heart beat
contains sequence no. of data message lost. If the message rate exceeds
threshold the destination node declares the channel has failed.
2. Neighbor detection resembles
the gateway failure protocol in the Internet.
Adjacent nodes periodically exchange node heart beats (“ I am alive”).
If the node does not receive heartbeat from one of its neighbors for a certain
period, it declares all the channels going through the silent Neighbor as fail.
FAILURE REPORTING AND CHANNEL SWITCHNG
The
node that detects the failure of a channel should report to the node
responsible for channel switching.
a. Failure reports are sent
from the failure detecting nodes to the end nodes of failed channels.
b. Failed reports are delivered
through healthy segments of the failed channels paths.
c. Each failure report contains
the channel-id of the failure channel.
The latest approach handles multiple simultaneous failures
very naturally and easily. If multiple failures occur to a channel, only one
failure report will reach its end nodes; all other reports will be lost due to
the failures themselves or discarded by the intermediate nodes.
When an end node of an real-time communication channel
receives a failure report on a primary channel, it selects one of this healthy
backups and sends an activation message along the path of the selected backup.
The transmission of failure reports and activation messages
is time critical, because their delays directly affect the service disruption
time. To achieve delay-bounded and robust transmission of time-critical control
messages, we transmit them over special purpose real-time channels, called
RCCs, one in each direction, are established on every link of the network. If
the capacity of the RCC on each link is large enough to accommodate all
time-critical control messages on the link, timely delivery of such messages
can be guaranteed.
RESOURCE RECONFIGURATION
In a normal situation, the dependability QoS of a
connection is maintained by limiting the admission of new connections not to
impair the QoS of existing connections. Upon occurrence of a failure, more
explicit actions (i.e. resource reconfiguration) need to be taken to preserve
the QoS of the connections, which are directly or indirectly affected by the
failure.
Even when a connection is not directly inflicted with
failures, its dependability QoS can be affected by the failure recovery for
other connections. This is because multiple backups share spare resources, and
activation of a backup will reduce the spare resources on its path, and as a
result the remaining backups on this path may not receive their original QoS.
At such links, more spare resource has to be allocated to maintain the same QoS
for the remaining backups. Here the network has to take care of a situation
where there are not enough resources available at a link to match the need for
additional spare resources. The network can resolve such situations by moving
some of the remaining backups to different paths or by QoS degradation.
PERFORMANCE ANALYSIS
As a metric of the fault-tolerance level achieved by each backup
configuration, the ratio of fast recovery to the number of failed primary
channels was used. For instance, a 90 percent fast recovery ratio means that 90
percent of the connections whose primary failed were recovered by using their
backup channels.
The multiplexing backup was even further improved by using
the double was achievable with significantly less spare resources in the double
backup configuration.
In
the mesh network, the reduction of spare resources by multiplexing is not as
great as in the tours network. This is because the absence of wrapped links in
the mesh network makes the primary channel paths more concentrated on the
central region of the network, those discouraging multiplexing among their
backups.
