Distributed Systems

Chapter 2 - Communication
 Objectives of the Chapter
 Review of how processes communicate in a network (the
rules or the protocols) and their structures
 Introduce the four widely used communication models for
distributed systems:
 Remote Procedure Call (RPC)
 Remote Method Invocation (RMI)
 Message-Oriented Middleware (MOM)
 Streams
 Introduction
 Interprocess communication is at the heart of all distributed
systems
 communication in distributed systems is based on message
passing as offered by the underlying network as opposed to
using shared memory
 Modern distributed systems consist of thousands of
processes scattered across an unreliable network such as
the Internet
 Unless the primitive communication facilities of the network
are replaced by more advanced ones, development of large
scale Distributed Systems becomes extremely difficult.
 2.1 Layered Protocols
 Two computers, possibly from different manufacturers, must
be able to talk to each other
 for such a communication, there has to be a standard
 The ISO OSI (Open Systems Interconnection) Reference
Model is one of such standards - 7 layers
 TCP/IP protocol suite is the other; has 4 or 5 layers
 OSI
 Open – to connect open systems or systems that are open
for communication with other open systems using standard
rules that govern the format, contents, and meaning of the
messages sent and received
 these rules are called protocols
 Two types of protocols: connection-oriented and
connectionless
layers, interfaces, and protocols in the OSI model
 Media (lower) Layers
 Physical: Physical characteristics of the media
 Data Link: Reliable data delivery across the link
 Network: Managing connections across the network
or routing
 Transport: End-to-end connection and reliability
(handles
lost packets); TCP (connection-oriented),
UDP (connectionless), etc.
 Session: Managing sessions between applications
(dialog control and synchronization); rarely
supported
 Presentation: Data presentation to applications; concerned
with the syntax and semantics of the
information transmitted
 Application: Network services to applications; contains
protocols that are commonly needed by
users;
Host (upper) FTP, HTTP, SMTP, ...
Layers
a typical message as it appears on the network
 A conversation occurs between a sender and a receiver at
each layer
 e.g., at the data link layer

discussion between a receiver and a sender in the data link layer

 Transport Protocols: Client-Server TCP

assuming no messages are lost,

 The client initiates a setup
connection using a three-way
handshake (1-3)
 The client sends its request (4)
 It then sends a message to
close the connection (5)
 The server acknowledges
receipt and informs the client
that the connection will be
closed down (6)
 Then sends the answer (7)
followed by a request to close
the connection (8)
 The client responds with an
normal operation of TCP ack to finish conversation (9)
 much of the overhead in TCP is for managing the connection
 combine connection setup with
request and closing connection
with answer
 such protocol is called TCP for
Transactions (T/TCP)
 the client sends a single
message consisting of a setup
request, service request, and
information to the server that
the connection will be closed
down immediately after
receiving the answer (1)
 the server sends acceptance of
connection request, the
answer, and a connection
release (2)
 the client acknowledges tear
down of the connection (3) transactional TCP
 2.2 Remote Procedure Call(RPC)
 The first distributed systems were based on explicit
message exchange between processes through the use of
explicit send and receive procedures; but do not allow
access transparency
 In 1984, Birrel and Nelson introduced a different way of
handling communication: RPC
 It allows a program to call a procedure located on another
machine
 Simple and elegant, but there are implementation problems
 the calling and called procedures run in different address
spaces
 parameters and results have to be exchanged; what if the
machines are not identical?
 what happens if both machines crash?
 2.2.1 Basic RPC Operation
 Conventional Procedure Call, i.e., on a single machine
 e.g. count = read (fd, buf, bytes); a C like statement, where
fd is an integer indicating a file
buf is an array of characters into which data are read
bytes is the number of bytes to be read
Stack pointer

Stack pointer

parameter passing in a local procedure the stack while the called

call: the stack before the call to read procedure is active
 parameters can be call-by-value (fd and bytes) or call-by
reference (buf) or in some languages call-by-copy/restore
 Client and Server Stubs
 RPC would like to make a remote procedure call look the
same as a local one; it should be transparent, i.e., the calling
procedure should not know that the called procedure is
executing on a different machine or vice versa

principle of RPC between a client and server program

 When a program is compiled, it uses different versions of
library functions called client stubs
 a server stub is the server-side equivalent of a client stub
 RPC Model
 A Server defines the server’s interface using an interface
definition language (IDL)
 The IDL specifies the names, parameters, and types for all client-
callable server procedures.
 A stub compiler reads the IDL and produces two stub
procedures for client and server:
 The server program implements the server procedures and links them
with the server-side stubs.
 The client program implements the client program and links it with the
client-side stubs.
 The stubs are responsible for managing all details of the remote
communication between client and server.
 Steps of a Remote Procedure Call
1. Client procedure calls client stub in the normal way
2. Client stub builds a message and calls the local OS
(packing parameters into a message is called parameter
marshaling)
3. Client's OS sends the message to the remote OS
4. Remote OS gives the message to the server stub
5. Server stub unpacks the parameters and calls the server
6. Server does the work and returns the result to the stub
7. Server stub packs it in a message and calls the local OS
8. Server's OS sends the message to the client's OS
9. Client's OS gives the message to the client stub
10. Stub unpacks the result and returns to client
 hence, for the client remote services are accessed by making
ordinary (local) procedure calls; not by calling send and
receive
 server machine vs server process; client machine vs client process
2.2.2 Parameter Passing
o The function of the client stub is to take its parameters, pack them into a mes
sage, and send them to the server stub. While this sounds straightforward, it is
not quite as simple as it at first appears.
1. Passing Value Parameters
 e.g., consider a remote procedure add(i, j), where i and j are
integer parameters

steps involved in doing remote computation through RPC

 the above discussion applies if the server and the client
machines are identical
 but that is not the case in large distributed systems
 the machines may differ in data representation (e.g., IBM
mainframes use EBCDIC whereas IBM PCs use ASCII)
 there are also differences in representing integers(1’s
complement or 2’s complement) and floating-point numbers
 byte numbering may be different (from right to left in Pentium
called little endian and left to right in SPARC, big endian)
 e.g.
 consider a procedure with two parameters, an integer and a
four-character string; each one 32-bit word (5, “JILL”)
 the sender is Intel and the receiver is SPARC
 original message on the Pentium
(the numbers in boxes indicate the address of each byte)

the message after receipt on the SPARC; wrong integer (5*224) 83886080, but
correct string
 one approach is to invert the bytes of each word after
receipt

the message after being inverted (correct integer but wrong string)

 additional information is required to tell which is an

integer and which is a string
 Solution: use a standard representation
Example: external data representation (XDR)
2. Passing Reference Parameters
 assume the parameter is a pointer to an array
 copy the array into the message and send it to the server
 the server stub can then call the server with a pointer to this
array
 the server then makes any changes to the array and sends it
back to the client stub which copies it to the client
 this is in effect call-by-copy/restore
 optimization of the method
 one of the copy operations can be eliminated if the stub
knows whether the parameter is input or output to the
server
 if it is an input to the server (e.g., in a call to write), it need
not be copied back
 if it is an output, it need not be sent over in the first place.
 the above procedure can handle pointers to simple arrays
and structures, but difficult to generalize it to an arbitrary
data structure.
 Parameter Specification and Stub Generation
 the caller and the callee need to use the same protocol
(format of messages) and the same steps; with such rules the
client and server stubs can assemble, communicate, and
interpret messages correctly
 consider the following example; the procedure foobar has 3
parameters: a character, a floating point number, and an array
of 5 integers

 assume a word is 4 bytes

 one possibility is to transmit the character
in the rightmost byte, a float as a whole
word, and an array as a group of words
equal to the array length preceded by a
word giving the length
 this way both client stub and server stub
can understand outgoing and incoming the corresponding message
messages
2.2.3 Extended RPC Models
 to solve some of the shortcomings of the original model
 no need of network communication if server and client are
on the same machine
 no need of blocking for the client in some cases

a. Doors
 the original RPC model assumes that the caller and the
callee can communicate only by means of passing
messages over a network; what if they are collocated on
the same machine?
 a door is a generic name for a procedure in the address
space of a server process that can be called by a process
collocated with the server
 support from the local OS is required
1. the server process registers a door before it can be called
(door_create) and a name is attached to it
1. a client calls a door by a system call (door_call) including
all parameters
2. results are returned by the system call door_return

the principle of using doors as IPC mechanism

 benefit: it allows the use of a single mechanism (procedure
calls) for communication
 disadv: application developers have to be aware of where a
procedure is located; is it
 local within the current process
 local to a different process on the same machine
 a remote process
b. Asynchronous RPC
 if there is no need to block the client until it gets a reply
 two cases
1. if there is no result to be returned
 e.g., adding entries in a database, ...
 the server immediately sends an ack promising that it
will carryout the request
 the client can now proceed without blocking

a) the interconnection between client and server in a traditional RPC

b) the interaction using asynchronous RPC
2. if the result can be collected later
 e.g., prefetching network addresses of a set of hosts, ...
 the server immediately sends an acknowledgment
promising that it will carryout the request
 the client can now proceed without blocking
 the server later sends the result

a client and server interacting through two asynchronous RPCs

  the above method combines two asynchronous RPCs
and is sometimes called deferred synchronous RPC
 variants of asynchronous RPC
 let the client continue without waiting even for an ack,
called one-way RPC
 problem: if reliability of communication is not
guaranteed
2.2.4 DCE (Distributed Computing Environment) RPC
 a middleware and an example RPC system developed by
OSF (Open Software Foundation), now The Open Group
 it is designed to execute as a layer of abstraction between
existing OSs and distributed applications
 the Open Group sells the source code and vendors integrate
it into their systems
 it uses the client-server programming model and
communication is by means of RPCs
 services
 distributed file service: a worldwide file system that
provides a transparent way of accessing files
 directory service: to keep track of the location of all
resources in the system (machines, printers, data,
servers, ...); a process can ask for a resource without
knowing its location
 security service: for protecting resources; access is only
through authorization
  distributed time service: to maintain clocks on different
machines synchronized (clock synchronization is covered
in Chapter 5)
 Steps in writing a Client and a Server in DCE RPC
 the system consists of languages, libraries, daemons,
utility programs, ... for writing clients and servers
 IDL (Interface Definition Language) is the interface
language - the glue that holds everything together
 it contains type definitions, constant declarations
and what the procedures do (only their syntax)
 Uuidgen generates a prototype IDL file with a globally unique interface
identifier
 the IDL file is edited (filling the names of procedures and parameters) and
the IDL compiler is called to generate 3 files
 the application writer writes the client and server codes and are then
compiled and linked together with the stubs
 Binding a Client to a Server in DCE RPC
 for a client to call a server, the server must be registered (1 &
2)
 the registration allows the client to locate the server and
bind to it
 the DCE daemon maintains a table (server, endpoint) and the
protocols the server uses
 the directory server maintains the locations of all resources
in the system (machines, servers, data,, ...)
 two steps for server location
 locate the server’s machine (3)
 locate the server process on that machine (which has
what is called an endpoint or port) (4)
 2.3 Remote Object (Method) Invocation (RMI)
 resulted from object-based technology that has proven its
value in developing nondistributed applications
 it is an expansion of the RPC mechanisms
 it enhances distribution transparency as a consequence of
an object that hides its internal from the outside world by
means of a well-defined interface
 Distributed Objects
 an object encapsulates data, called the state, and the
operations on those data, called methods
 methods are made available through interfaces
 the state of an object can be manipulated only by invoking
methods
 this allows an interface to be placed on one machine while
the object itself resides on another machine; such an
organization is referred to as a distributed object
 the state of an object is not distributed, only the interfaces
are; such objects are also referred to as remote objects
 the implementation of an object’s interface is called a proxy
(analogous to a client stub in RPC systems)
 it is loaded into the client’s address space when a client
binds to a distributed object
 tasks: a proxy marshals method invocation into messages
and unmarshals reply messages to return the result of the
method invocation to the client
 a server stub, called a skeleton, unmarshals messages and
marshals replies
common organization of a remote object with client-side proxy
 Binding a Client to an Object
 a process must first bind to an object before invoking its
methods, which results in a proxy being placed in the
process’s address space
 binding can be implicit (directly invoke methods using
only a reference to an object) or explicit (by calling a
special function)
 an object reference could contain
 network address of the machine where the object
resides
 endpoint of the server
 an identification of which object
 the protocol used
 ...
RMI Architecture
The General RMI Architecture
Remote Machine

 The server must first bind bind

its name to the registry RMI Server

 The client lookup the Registry

server name in the registry skeleton

to establish remote
references. return call lookup
 The Stub serializing the
parameters to skeleton,
stub
the skeleton invoking the
remote method and
RMI Client
serializing the result back
to the stub.
Local Machine
The Stub and Skeleton
call

skeleton
Stub
RMI Client RMI Server
return

 A client invokes a remote method, the call is first

forwarded to stub.
 The stub is responsible for sending the remote call
over to the server-side skeleton
 The stub opening a socket to the remote server,
marshaling the object parameters and forwarding the
data stream to the skeleton.
 A skeleton contains a method that receives the remote
calls, unmarshals the parameters, and invokes the
actual remote object implementation.
 Parameter Passing
 there are two situations when invoking a method with
object reference as parameter; is the object local or
remote to the client?
 remote object: copy and pass the reference of the object
as a value parameter; this means the object is passed by
reference
 local object: a copy of the object is passed; this means the
object is passed by value
Figure: the situation when passing an object by reference or by
value

42
 2.4 Message Oriented Communication
 RPCs and RMIs are not adequate for all distributed system
applications
 The provision of access transparency may be good but
they have semantics that is not adequate for all
applications
 Example problems
 They assume that the receiving side is running at the
time of communication
 a client is blocked until its request has been processed
 2.4.1 Persistence and Synchronicity in
Communication
 assume the communication system is organized as a
computer network shown below

general organization of a communication system in which hosts are connected

through a network
 communication can be
 persistent or transient
 asynchronous or synchronous
 persistent: a message that has been submitted for
transmission is stored by the communication system as long
as it takes to deliver it to the receiver
 e.g., email delivery, snail mail delivery

persistent communication of letters back in the days of the Pony Express

 transient: a message that has been submitted for transmission
is stored by the communication system only as long as the
sending and receiving applications are executing
 asynchronous: a sender continues immediately after it has
submitted its message for transmission
 synchronous: the sender is blocked until its message is
stored in a local buffer at the receiving host or delivered to the
receiver
 the different types of communication can be combined
 persistent asynchronous: e.g., email
 transient asynchronous: e.g., UDP, asynchronous RPC
 in general there are six possibilities

Persistent Transient

Asynchronous  

Synchronous  message-oriented; three forms

persistent asynchronous communication persistent synchronous communication
transient asynchronous communication receipt-based transient synchronous
communication

 weakest form; the sender is

blocked until the message is
stored in a local buffer at the
receiving host
delivery-based transient synchronous response-based transient synchronous
communication at message delivery communication
 the sender is blocked until the  strongest form; the sender is
message is delivered to the blocked until it receives a reply
receiver for further processing; message from the receiver
e.g., asynchronous RPC
2.4.2 Message-Oriented Transient Communication
 many applications are built on top of the simple message-
oriented model offered by the transport layer
 standardizing the interface of the transport layer by
providing a set of primitives allows programmers to use
messaging protocols
 they also allow porting applications

1. Berkley Sockets
 an example is the socket interface as used in Berkley
UNIX
 a socket is a communication endpoint to which an
application can write data that are to be sent over the
network, and from which incoming data can be read.
Primitive Meaning Executed by
Socket Create a new communication endpoint; also both
reserve resources to send and receive messages
Bind Attach a local address to a socket; e.g., IP
address with a known port number
Listen Announce willingness to accept connections; for
connection-oriented communication
Accept Block caller until a connection request arrives
servers
Connect Actively attempt to establish a connection; the
client is blocked until connection is set up
Send Send some data over the connection
Receive Receive some data over the connection
Close Release the connection
Socket primitives for TCP/IP
connection-oriented communication pattern using sockets
2. The Message-Passing Interface (MPI)
 sockets were designed to communicate across networks
using general-purpose protocol stacks such as TCP/IP
 they were not designed for proprietary protocols
developed for high-speed interconnection networks; of
course portability will suffer
 MPI is designed for parallel applications and tailored for
transient communication
 MPI assumes communication takes place within a known
group of processes, where each group is assigned an
identifier (groupID)
 each process within a group is also assigned an identifier
(processID)
 a (groupID, processID) identifies the source or destination
of a message, and is used instead of a transport-level
address
Primitive Meaning

MPI_bsend Append outgoing message to a local send buffer; to support

transient asynchronous communication
Send a message and wait until copied to local or remote
MPI_send buffer (to support receipt-based transient synchronous
communication)

MPI_ssend Send a message and wait until receipt starts (to support
delivery-based transient synchronous communication)

MPI_sendrecv Send a message and wait for reply (to support response-
based transient synchronous communication)

MPI_isend Pass reference to outgoing message, and continue (a

variant of MPI_send)

MPI_issend Pass reference to outgoing message, and wait until receipt

starts (a variant of MPI_ssend)
MPI_recv Receive a message; block if there are none
MPI_irecv Check if there is an incoming message, but do not block
some of the most intuitive message-passing primitives of MPI
2.4.3 Message-Oriented Persistent Communication
 there are message-oriented middleware services, called
message-queuing systems or Message-Oriented Middleware
(MOM)
 they support persistent asynchronous communication
 they have intermediate-term storage capacity for messages,
without requiring the sender or the receiver to be active
during message transmission
 unlike Berkley sockets and MPI, message transfer may take
minutes instead of seconds or milliseconds
Message-Queuing Model
 applications communicate by inserting messages in
specific queues
 it permits loosely-coupled communication
 the sender may or may not be running; similarly the
receiver may or may not be running, giving four possible
combinations
four combinations for loosely-coupled communications using queues
Primitive Meaning

Put Append a message to a specified queue; by the sender

and is nonblocking

Get Block until the specified queue is nonempty, and remove

the first message

Poll Check a specified queue for messages, and remove the

first. Never block

Notify Install a handler to be called when a message is put into

the specified queue; usually a daemon
basic interface to a queue in a message-queuing system
 General Architecture of a Message-Queuing System
 messages can be put only into queues that are local to the
sender (same machine or on a nearby machine on a LAN)
 such a queue is called the source queue
 messages can also be read only from local queues
 a message put into a local queue must contain the
specification of the destination queue; hence a message-
queuing system must maintain a mapping of queues to network
locations; like in DNS

the relationship between queue-level addressing and network-level addressing

 messages are managed by queue managers
 they generally interact with the application that sends and
receives messages
 some also serve as routers or relays, i.e., they forward
incoming messages to other queue managers
 however, each queue manager needs a copy of the queue-
to-location mapping, leading to network management
problems for large-scale queuing systems
 the solution is to use a few routers that know about the
network topology
the general organization of a message-queuing system with routers
 Message Brokers
 how can applications understand the messages they receive
 each receiver can not be made to understand message formats
of new applications
 hence, in a message-queuing system conversations are
handled by message brokers
 a message broker converts incoming messages to a format
that can be understood by the destination application based on
a set of rules

the general organization of a message broker in a message-queuing system

 2.4 Stream Oriented Communication
 until now, we focused on exchanging independent and complete
units of information
 time has no effect on correctness; a system can be slow or fast
 however, there are communications where time has a critical role
 Multimedia
 media
 storage, transmission, interchange, presentation,
representation and perception of different data types:
 text, graphics, images, voice, audio, video, animation, ...
 movie: video + audio + …
 multimedia: handling of a variety of representation media
 end user pull
 information overload and starvation
 technology push
 emerging technology to integrate media
 The Challenge
 new applications
 multimedia will be pervasive in few years (as graphics)
 storage and transmission
 e.g., 2 hours uncompressed HDTV (1920×1080) movie:
1.02 TB (1920×1080x3x25x60x60x2)
 videos are extremely large, even compressed
 continuous delivery
 e.g., 30 frames/s (NTSC), 25 frames/s (PAL) for video
 guaranteed Quality of Service
 admission control
 search
 can we look at 100… videos to find the proper one?

 National Television Standards Committee

 Types of Media
 two types
 discrete media: text, executable code, graphics, images;
temporal relationships between data items are not
fundamental to correctly interpret the data.
 continuous media: video, audio, animation; temporal
relationships between data items are fundamental to
correctly interpret the data.
 a data stream is a sequence of data units and can be applied
to discrete as well as continuous media
 stream-oriented communication provides facilities for the
exchange of time-dependent information (continuous media)
such as audio and video streams
 timing in transmission modes
 asynchronous transmission mode: data items are
transmitted one after the other, but no timing constraints;
e.g. text transfer
 synchronous transmission mode: a maximum end-to-end
delay defined for each data unit; it is possible that data can
be transmitted faster than the maximum delay, but not slower
 isochronous transmission mode: maximum and minimum
end-to-end delay are defined; also called bounded delay
jitter; applicable for distributed multimedia systems
 a continuous data stream can be simple or complex
 simple stream: consists of a single sequence of data; e.g.,
mono audio, video only
 complex stream: consists of several related simple streams
that must be synchronized; e.g., stereo audio, video
consisting of audio and video (may also contain subtitles,
translation to other languages, ...)
movie as a set of simple streams
 a stream can be considered as a virtual connection between a
source and a sink
 the source or the sink could be a process or a device

setting up a stream between two processes across a network

setting up a stream directly between two devices

 the data stream can also be multicasted to several receivers
 if devices and the underlying networks have different
capabilities, the stream may be filtered, generally called
adaptation (filtering?, transcoding?)

an example of multicasting a stream to several receivers

 Quality of Service (QoS)
 QoS requirements describe what is needed from the
underlying distributed system and network to ensure
acceptable delivery; e.g. viewing experience of a user
 for continuous data, the concerns are
 timeliness: data must be delivered in time
 volume: the required throughput must be met
 reliability: a given level of loss of data must not be
exceeded
 quality of perception; highly subjective
 QoS Dimensions
 timeliness dimensions
 latency (maximum delay between consecutive frames)
 start-up latency (maximum delay before starting a
presentation)
 jitter (delay variance)
 volume dimensions
 throughput in frames/sec or bits/sec or bytes/sec
 reliability dimensions
 MTBF (Mean Time Between Failure) of disks
 MTTR (Mean Time To Repair)
 error rates on the telecommunication lines
 QoS Requirements
 deterministic
 precise values or ranges
 e.g., latency must be between 45 and 55 ms
 probabilistic
 probability of the required QoS
 e.g., latency should be < 50 ms for 95% of the frames
 stochastic distributions
 e.g., frame arrival should follow normal distribution with
mean interval-time of 40 ms and 5 ms variance
 classes
 e.g., guaranteed and best effort
 QoS Management
 can be static or dynamic
 Static QoS Management Functions
 specification
 e.g., deterministic range for timeliness, volume and
reliability categories
 negotiation
 the application may accept lower level of QoS for
lower cost
 admission control
 if this test is passed, the system has to guarantee the
promised QoS
 resource reservation
 may be necessary to provide guaranteed QoS
 Dynamic QoS Management Functions
 monitoring
 notices deviation from QoS level
 at a certain level of granularity (e.g., every 100 ms)
 policing
 detect participants not keeping themselves to the contract
 e.g., source sends faster than negotiated (e.g., 25 fps)
 maintenance
 sustaining the negotiated QoS
 e.g., the system requires more resources
 renegotiation
 client tries to adapt – may be can accept lower QoS
 QoS requirements can be specified using flow specification
containing bandwidth requirements, transmission rates,
delays, ...
 e.g. by Partridge (1992)
 it uses the token bucket algorithm which specifies how the
stream will shape its network traffic (in fact the leaky
bucket, as used in networking)
 the idea is to shape bursty traffic into fixed-rate traffic by
averaging the data rate
 packets may be dropped if the bucket is full
 the input rate may vary, but the output rate remains
constant
the principle of a token bucket algorithm
  problem in flow specification
 an application may not know its requirements
 how can a user (human) specify quality using the various
parameters? usually very difficult
 may be provide defaults for various streams as high,
medium, low quality
 Setting up a Stream
 resources such as bandwidth, buffers, processing power
must be reserved once a flow specification is made
 on such protocol is RSVP - Resource reSerVation Protocol
 it is a transport-level protocol for enabling resource
reservation in network routers
the basic organization of RSVP for resource reservation in a distributed system
 Stream Synchronization
 how to maintain temporal relations between streams, e.g., lip
synchronization
 two approaches
1. explicitly by operating on the data units of simple
streams; the responsibility of the application

the principle of explicit synchronization on the level of data units

2. through a multimedia middleware that offers a collection of
interfaces for controlling audio and video streams as well as
devices such as monitors, cameras, microphones, ...

the principle of synchronization as supported by high-level interfaces