[PPT] - Networked Embedded Systems Broadcast Applications Heinz Deinhart PowerPoint Presentation

SLIDE 1

1

Networked Embedded Systems Broadcast Applications

Heinz Deinhart <heinz@ecs.tuwien.ac.at>

Networked Embedded Systems Broadcast Applications

SLIDE 2

2

Overview This presentation consists of three parts:

Short introduction
Classifications of protocol implementations
Real world examples

Networked Embedded Systems Broadcast Applications

SLIDE 3

3

Part 1: Introduction

Networked Embedded Systems Broadcast Applications

SLIDE 4

4

Conventional broadcast in the wild First, lets take a look where plain broadcast (no order, reliability, etc.) is used in LANs:

RARP: address resolving
BOOTP/DHCP: address and metainfo resolving
SMB: Server Message Block, windows fileserver
GAMES: fun on the LAN
NOVELL: a long time ago..

Basically where no info is available or a complex setup is not needed.

Networked Embedded Systems Broadcast Applications

SLIDE 5

5

What does LAN mean?

Some years ago LANs were a usually on shared mediums (thinwire

ethernet, token ring) separated by routers.

Nearly all of them in production envionment are now replaced by

some kind of switched infrastructure

Some of them are even that smart and convert broadcast

communication into point of point where possible (with application layer sniffing)

That means: Broadcast is not what it used to be
But: recently shared medium celebrated comeback: wireless

communication

Networked Embedded Systems Broadcast Applications

SLIDE 6

6

What about the Internet?

internet wide broadcast is a bit resource intensive
there is some feature called directed broadcast
but its disabled on nearly all routers out there for security (DoS)

reasons

applications that use broadcast like features can use multicast
ip multicast requires some setup (IGMP, ..)
or implement broadcast features at application level, but this is not the

direction we want to look further into

Networked Embedded Systems Broadcast Applications

SLIDE 7

7

Back to our domain: Networked embedded systems

Embedded sytems are usually small or/and cheap
To connect some in a network its a good idea to use a simple

(=shared) medium

Imagine lots of ultra small sensors that are connected with switched

wired ethernet: not that clever

Its obvious that using are shared medium, maybe even air is the way

to go

Using shared medium makes “real” broadcast an important feature

again

Networked Embedded Systems Broadcast Applications

SLIDE 8

8

Part 2: Classification of Protocol Implementations

Networked Embedded Systems Broadcast Applications

SLIDE 9

9

Introduction

There are many implementations of different broadcast algorithms
To few time to look into implementation details
Real world examples in part 3
Here we take a look how implementations can be classified
What design decisions have to be made for a new protocol?
To save time we have to focus on the most important characteristics

Networked Embedded Systems Broadcast Applications

SLIDE 10

10

Basic System Model: Synchrony Two important parameters:

relative speed of processes = speed ratio between slowest and fastest

process

communication delay

Only if both parameters have a known upper bound which always holds the system is called synchonous. Otherwise it is called an asynchronous system Hardly any implementation is 100% synchronous.

Networked Embedded Systems Broadcast Applications

SLIDE 11

11

Basic System Model: Failure Mode The general class of failures are:

Crash faiures: When a process crashes it ceases functioning forever.
Omission failures: Process omits performing some actions such as

sending or receiving a message

Timing failures: Process violates one of the synchrony assumptions
Byzantine failures: Arbitrary fault

A correct process is one that never expresses any of the faulty behaviors.

Networked Embedded Systems Broadcast Applications

SLIDE 12

12

Oracles Synchrony does not solve all Problems. Some are only solvable with an

rcale, which is a distributed component a process can query.

Oracles can be:

Physical Clocks: Gives information about physical time
Failure Detectors: Process can query status of other processes

(crashed or not)

Coin Flip: Generates random values. Not that trivial as it might seem.

Networked Embedded Systems Broadcast Applications

SLIDE 13

13

Agreement Problems Processes have to reach some common decision.

Consensus
Byzantine Agreement
Reliable Broadcast
Atomic Broadcast

Networked Embedded Systems Broadcast Applications

SLIDE 14

14

Process Controlled Crash

gives processes the ability to kill other processes
.. or to commit suicide
that allows the transformation of failures into less severe ones
e.g. detect a byzantine or omission fault and transform it into crash
has bad sides: reduces the fault tolerance of system

Networked Embedded Systems Broadcast Applications

SLIDE 15

15

Group Membership Service .. is a distributed service that is responsible for managing groups of

processes. Groups are important for multicast.

It can support:

Join/Leave: way for a process to dynamically join/leave a group
Exclusion: when process crashed it is removed from group

Networked Embedded Systems Broadcast Applications

SLIDE 16

16

Part 3: Finally, some real world examples

Networked Embedded Systems Broadcast Applications

SLIDE 19

19

Broadcast Applications - Whats that?

okay, sorry, still no examples yet, but on next slide (promise)
lets take a short break and think about the title of this presentation
a broadcast application is everything that somehow applicates

broadcasting

can be algorithms (you could even think of reliable broadcast as a

broadcast application)

can be implementations of algorithms
can be an application that uses such an implementation
can be an invocation of an application by some user (who may or may

not know what he is really doing)

okay, lets proceed in this order..

Networked Embedded Systems Broadcast Applications

SLIDE 20

20

Using Broadcast Primitives in Replicated Databases [3]

databases are a perfect example
everyone (except the very lucky ones) know about it
access is done using transactions
broadcasting can be helpful for replication
database replicas should stay in sync
it is obvious that best effort broadcast is not sufficient
operations can be performed on a higher layer

Networked Embedded Systems Broadcast Applications

SLIDE 21

21

Replicated databases with reliable broadcast

reliable broadcast is simple and straight forward to implement
since it guarantees eventual delivery it removes need for explicit ack
since transaction are not blocked for global synchronisation there are

no deadlocks

remaining problem: because order is not guaranteed write operations

may arrive in different orders

solution: two phase locking must be used to circumvent this problem
interesting property: read only transactions are never aborted

Networked Embedded Systems Broadcast Applications

SLIDE 22

22

Replicated databases with causal broadcast

causal broadcast is a more expensive primitive than reliable broadcast
it imposes more ordering restrictions on message delivery
that avoids the need for two phase commits
it uses the causal delivery properties to implicitly collect the

information used in a two phase commit

Networked Embedded Systems Broadcast Applications

SLIDE 23

23

Replicated databases with atomic broadcast

using atomic (and causal) broadcast is even more expensive
ensures total order of commit requests
benefit: eliminates need for acks during commits
drawback: to work, the protocol needs versioning with numbers

Networked Embedded Systems Broadcast Applications

SLIDE 24

24

Replicated databases conclusion

conclusion: choosing proper broadcast primitive is important
it changes the logical layer on which things are done
choosing the proper primitives regarding to the used medium is a

good idea

Networked Embedded Systems Broadcast Applications

SLIDE 25

25

Transis [1] Overview

Transis project currently under development at the Hebrew University
f Jerusalem
it was designed to supply reliable and efficient transport layer services

for distributes services

it provides fast and reliable message multicast between all currently

connected processors (despite an unreliable network)

it recovers msgs that were lost due to ommission failures
handles processor crashes and dynamic partitions of the network
it achieves good performance by using the underlying networks

unreliable broadcast service

Networked Embedded Systems Broadcast Applications

SLIDE 26

26

Transis Services Transis offers the following set of services:

Membership: maintains the membership of processors
Basic multicast: guarantees deilvery of the message at all active sites.

Delivers messages immediatly to upper level

Causal multicast: guarantees that delivery order preserves causality
Agreed multicast: delivers messages in the same order at all sites.

Various protocols are supported for achieving agreed order.

Safe multicast: delivers a message after all active processors have

acknowledged its reception

Networked Embedded Systems Broadcast Applications

SLIDE 27

27

Transis Mempership Protocol

automatically maintains membership of connected processors
this is necessary for constructing fault-tolerant distributed applications
it guarantees virtual synchrony,
it is symmetric and does not disturb the regular flow of messages
it does not allow indefinite blocking of processors and may (rarely)

remove live but inactive processors unjustly

inadvertently removed processors can rejoin immediatley

Networked Embedded Systems Broadcast Applications

SLIDE 28

28

Transis Partitining and Merging

environment is dynamic: processors can crash and restart
whole (sub)network can partition and reconnect
partitioning and merging is the greatest challenge transis has to handle
they can be handled becuase of the symmetric design
all previous systems deal only with join of single processors
or even stop execution during network partition
transis can do better - how, thats beyond our scope today

Networked Embedded Systems Broadcast Applications

SLIDE 29

29

Transis LAN Performance

original prototype of transis was testet on 10Mbit Ethernet
using UDP in most requiring conditions it achives throughput of 160

1k messages per second

in comparision: TCP peer to peer transfare rate is about 350k/s
warning: those numbers are from last centory, probaly about 1991

Networked Embedded Systems Broadcast Applications

SLIDE 30

30

A replicated mail server using transis [?]

a typical replicated application in a distributed system
was a final example in university course
domain: a mail service that is replicated among several mail servers
each client can send read or delete mail when connected to any server

Networked Embedded Systems Broadcast Applications

SLIDE 31

31

Replicated Mail Service using Transis [?]

a typical replicated application in a distributed system
was a final example in university course
domain: a mail service that is replicated among several mail servers
each client can send read or delete mail when connected to any server
makes use of safe messages provided by transis
so no need to manually care about acks, omissions or processor faults

Networked Embedded Systems Broadcast Applications

SLIDE 32

32

Replicated Mail Service Architecture

Networked Embedded Systems Broadcast Applications

SLIDE 33

33

The algorithm

mail service implements own algorithm on top of transis environment
all transaction are divided into white, green and red
white are stable and ordered
green are ordered but not locally stable
red are neither ordered nor stable (yet)
stable at p is when p knows that alles targets received transaction and

stored it to disk

you can see that transis needs more than just a “configuration”, it

needs some logic on top of it

Networked Embedded Systems Broadcast Applications

SLIDE 34

34

Mail service: Transis compared to other scenarios

Centralized Server: easy to implement but no fault tolerance
One Server with Several NFS Copies: Lots of overhead since

transaction must be simulated, inconsistences if network partitiones

Replicated Servers with DB using two phase commit: Does not fit the

mail concept very vell, won’t perform very well

Conclusio: ransis provides three major properties that are important

for mail service: multicasting, message ordering and membership management

If your base system lack some of them they are hard to emulate.

Networked Embedded Systems Broadcast Applications

SLIDE 35

35

BMW: Broadcast Medium Window [4]

reliable broadcast in ad hoc networks (802.11)?
802.11 use collision avoidance along with RTS/CTS/ACK control

frames to tramsmit unicast packages

for broadcast no control frames are used
that means broadcast packats are sent blindly without consideration of

hidden terminals and channel noise

BWM is a new protocol to support reliable multicast in ad hoc

networks

fundamental idea: transmit each package to each neighbor in a round

robin fashion

Networked Embedded Systems Broadcast Applications

SLIDE 36

36

BMW: How it works

each node is required to maintain three lists: neighbors, send buffer

and receive buffer

nodes in neighbors list are touched when node hears from them
nodes are removed from it when time out since last touch occurs
the send buffer contains copies of frames that might be needed for

resend later

a copy is removed when all neighbours got it
thus the size of send buffer must be at least the max number of

neighbours

internal a queue of frames that weren’t sent at all is maintained
when a node receives a new node then its seq-nr added to the receive

buffer so it can be compared later

Networked Embedded Systems Broadcast Applications

SLIDE 37

37

BMW: How it works 2

of course the whole algorithm is a bit more complex and implements a

kind of state machine

the key point is that the broadcasting handshake is done at mac layer
when a node wants to send it goes to a CSMD/CA phase similar to

802.11

then it RTS to one of its neighbours which becomes the receiver for

this send and declares what seq nrs are missing

the sending node then broadcasts new and missing frames
all other neighbours can use the frames as well, if they need it
of course the receiver neighbour is rotated

Networked Embedded Systems Broadcast Applications

SLIDE 38

38

BWM Conclusio

a concept like this could be implemented on a highler layer
the performance will not be that good
especially if there is high traffic which means lots of advoided

collision

lets take a look at the results of a simulated experiment
compared is the performance of ODMRP with and without BMW
ODMRP is the On-Demand Multicast Routing Protocol

Networked Embedded Systems Broadcast Applications

SLIDE 39

39 Networked Embedded Systems Broadcast Applications

SLIDE 40

40

Thanks!

Networked Embedded Systems Broadcast Applications

SLIDE 41

41

Bibiliography

References

[1] Y. Amir, D. Dolev, S. Kramer, and D. Malki. Transis: A communication subsystem for high

availability. In FTCS-22: 22nd International Symposium on Fault Tolerant Computing, pages

76–84, Boston, Massachusetts, 1992. IEEE Computer Society Press. [2] X. Defago, A. Schiper, and P. Urban. Totally ordered broadcast and multicast algorithms: a comprehensive survey, 2000. [3] Ioana Stanoi, Divyakant Agrawal, and Amr El Abbadi. Using broadcast primitives in replicated databases. In International Conference on Distributed Computing Systems, pages 148–155, 1998. [4] K. Tang and M. Gerla. Mac reliable broadcast in ad hoc networks. In Communications for Network-Centric Operations: Creating the Information Force. IEEE Military Communications Conference, volume 2, pages 1008–1013 vol.2, 2001.

Networked Embedded Systems Broadcast Applications