Introduction to Distributed Systems EEE466 DA06

1
Introduction to Distributed SystemsEEE466DA06
2
Next Quiz

• Coulouris, et al
• 3.1
• 3.4
• 3.7

3
Review

• Deadlock
• Livelock

4
Interprocess Communication

• based on messages (data flow) between processes
• originally developed to allow communication
  between different processes running on the same
  computer
• later extended to allow distributed communication
• relies on protocols at the application level to
  coordinate communication and to give meaning to
  received messages

5
Architectural Model

• processes communicate with each other via data
  flow to named destinations, usually ports or
  sockets (or groups of these)
• a single process may have several ports/sockets
• processes must actively poll (read from) their
  ports/sockets to detect messages on them

unless we use asynchronous reads in the IPC API
6
Destinations

• a destination may be any one of
• process (V)
• port (Mach, Chorus, Amoeba)
• socket (Unix, Windows)
• process group (V, Amoeba)
• port group (Chorus)
• object (Clouds, Emerald)
• all these destinations except sockets are
  location and migration transparent
• a migration transparent destination allows the
  destination to be relocated from one machine to
  another
• a Unix-style socket consists of a hostname or IP
  address plus a numbered socket identifier

7
Fundamental Operations

• send(portId, message)
• sends a message to a given port
• in some systems, the portId is the destination
  port in others it is a local port which has been
  bound (attached) to a destination port
• receive(portId, message)
• receives a message on a local port

8
Synchronous or Asynchronous

• a queue is associated with each destination
• sending processes add messages to queues,
  receiving processes remove them from queues
• communication can be either synchronous or
  asynchronous
• assuming multiple threads in a single process,
  the asynchronous version with blocking receive
  has few disadvantages
• the non-blocking receive is a somewhat unusual
  programming model

9
Timeouts

• Normally a blocking receive will wait until data
  arrives on the port
• If the connection has gone down, or the sender
  has disappeared, data may never arrive
• One solution is to associate a timeout interval
  with blocking receives
• if the time expires, the receiving process can
  take appropriate action
• it is not always clear what this should be
  design decision
• the timeout interval must be chosen appropriately
• fairly long compared with time required to
  transmit a message
• message transmission time may not be known at
  design time

10
Reliability and Ordering

• Message delivery may be
• guaranteed (stream - TCP) or best effort
  (datagram UDP )
• ordered (stream) or unordered (datagram)
• duplicate-free (stream) or possibly duplicated
  (datagram)
• Perfectly reliable communication sometimes not
  needed and causes overhead
• need to store state at both ends
• retransmission
• latency
• Reliable system can be created from unreliable
  one using acknowledgements
• requires unique message identifiers, normally
  consist of request Id and sender identification

11
Reliability and Ordering

• Stream (TCP Transport Control Protocol )
• Connection oriented service
• Like telephone service
• Datagram (UDP User Datagram Protocol )
• Connectionless oriented service
• Like postal service

12
Message Contents

• Data in programs is data structures/objects data
  in messages is linear streams of bytes
• conversion from data structures to byte streams
  (and vice versa) is required
• if system is heterogeneous, an external data
  representation may be required
• inefficient if applied unnecessarily
• conversion process is referred to as marshalling/
  unmarshalling
• Marshalling and unmarshalling can be done
  automatically or by hand
• In Java, marshalling can be handled automatically
  (and may be tuned) called serialization

13
Marshalling Example

• A data structure has three fields, two
  variable-length strings and an integer Smith,
  London, 1934
• Marshalled by hand, could be converted to
  character string Smith London 1934
• requires reconstruction at other end
• generally wasteful of bandwidth
• Marshalled using Sun XDR (External Data
  Representation)
• sent in four-byte blocks here 5 Smit h---
  Lond on--- 1934
• for large binary data blocks, generally more
  space efficient (here, actually less efficient)
• Must either send or know data types

14
Marshalling

• Complex, non-linear data structures require
  flattening as part of marshalling
• Must include enough information to exactly
  reconstruct their original forms
• Examples
• arrays
• nested arrays
• linked lists
• trees
• hash tables
• maps