Modern Operating Systems

1
Modern Operating Systems

• Lecture 5

2
Lecture 5 Distributed System Design

• Objectives Lecture 5
• To consider distributed system architecture
• To identify basic concepts
• To investigate design goals

3
Lecture 5 Distributed System Design

• We have seen that Operating System provides an
  environment to execute programs.
• What are distributed systems ?
• They could have a distributed OS
• They could have centralised OS with distributed
  resources
• Or they can have a mixture of both!

4
Lecture 5 Distributed System Design.

CPU
CPU
CPU
CPU
CPU
CPU
Loosely Coupled
Tightly Coupled Systems
Separated but joined by a Network
In same box shared BUS
5
Lecture 5 Distributed System Design.

• What are the design goal for a distributed
  system?
• Economic - Idle workstation CPUs can share
  possible heavy workloads!
• Reliability - if one of ten workstations crash
  only 10 of the system has failed
• Resource Share
• Performance
• Incremental growth

6
Lecture 5 Distributed System Design.

• Two different types of OS have been developed for
  Distributed Systems
• Most Common - attach file system from remote
  system to a local workstation
• Runs totally on the Server side
• Another type of NOS that starts to blur the
  boundary between machines
• System manages the resources but were the
  resources are physically does not matter.

7
Lecture 5 Distributed System Design.

• Hybrid Systems
• Uses a fully distributed approach not yet that
  common.
• CORBA - Common Object Request Broker
  Architecture.
• COM DCOM NTs Common Object Model.

8
Lecture 5 Distributed System Design.

• SOCKETS
• Any form of distributed system relies on the
  ability of the different systems to communicate
  effectively.
• Sockets is a set of facilities to HIDE the
  network protocols etc. and offer a uniform
  approach to the interworking of two or more
  machines.

9
Lecture 5 Distributed System Design.

• RPC - Remote Procedure Calls
• Sockets provide a mechanism to transmit data
  between machines I.e. across the address space
  boundary.
• RPC provides the mechanism to use this data in a
  form that makes a function call appear
  transparent I.e. the application has no idea if
  the function being used is local or remote.

10
Lecture 5 Distributed System Design.

• Mutual Exclusion
• With Sockets and RPC we can communicate between
  machines in a distributed system but what about
  unique access to a specific resource when we are
  dealing across machine boundaries.

11
Lecture 5 Distributed System Design.

• Critical Regions for Process and thread access on
  a single machine lead to the concepts of
  Semaphores, Critical Regions etc.
• These tend to rely on sharing data variables
  within the OS and hence do not map to a
  distributed system.

12
Lecture 5 Distributed System Design.

• Centralised algorithms
• In this design technique the distributed system
  has a single process that controls entry to
  critical or exclusive resources.
• Such systems work until there is a problem with
  either a failure or message loss in the
  connection philosophy.
• Techniques must be available to detect that the
  control process is still functioning.

13
Lecture 5 Distributed System Design.

• Distributed algorithms
• Various techniques have been developed that take
  away the centralised algorithmic control
• Most use a variant of a time stamp to work out
  who got the request in first and whether to grant
  the request.
• The disadvantage is that a great deal of message
  passing is required to determine what is going
  on.
• Suitable for small, stable sets of co-operating
  processes

14
Lecture 5 Distributed System Design.

• DeadLock
• single standalone systems can reach deadlock.
• When looking at a system of co-operating machines
  the situation becomes even more complicated as do
  the solutions
• Prevention and avoidance algorithms associated
  with single OS machines are rare in distributed
  systems

15
Lecture 5 Distributed System Design.

• Prevention and Avoidance
• Such algorithms require advanced knowledge of the
  resources required.
• Not possible on a co-operative system
• Deadlock detection and recovery
• Requires complex algorithms in eaxh machine
  looking at resource utilisation locally, when a
  pattern of unison between machines is found
  recovery of resource must start.

16
Lecture 5 Distributed System Design.

• Distributed Shared Memory
• Current OS designs are allowing memory for a
  process on one system to be shared with a process
  on another.
• Advantage of this is that standalone applications
  could be migrated to distributed systems with
  minimal effort.
• The detail of the run time support structures for
  such scheme is complex and CORBA and COM offer
  more viable alternatives.

17
Lecture 5 Distributed System Design.

• File Systems
• Distributed File Systems will be dealt with in
  specific lectures.

18
Lecture 5 Distributed System Design.

• Scaling Issues
• Any implementation of a distributed system must
  offer scaling not just to small, meduim or large
  but to world-wide services.
• This tends to rule out all centralised solutions
  to problems such as deadlock.
• Emphasis must be placed on fully distributed
  algorithms to overcome the identified problems

19
Lecture 5 Distributed System Design.

• Fault Tolerance
• Since we are dealing with multiple machines and a
  communications infrastructure which could be
  spread over a large area
• System must be prone to more faults than a
  standalone machine
• These same characteristics can be used to ensure
  the system is more fault resistant than any
  standalone system but at a price!

20
Lecture 5 Distributed System Design.

• Security issues beyond those found on standalone
  systems
• Authentication of the users and machines that are
  part of the system
• Protection of the messages that pass arround the
  system

21
Lecture 5 Distributed System Design.

• Authentication
• One way Authentication
• Two way Authentication
• More details in a later lecture

22
Lecture 5 Distributed System Design.

• Cryptography
• Data on the infrastructure is vulnerable to
  attack
• Masquerading as a legitimate source is a method
  of obtaining unauthorised information
• Interception and tampering offers a means of
  making servers or services perform unauthorised
  actions

23
Lecture 5 Distributed System Design.

• Design Techniques
• Pre partition
• Design the system with the distributed
  architecture already defined
• Prone to errors if the architecture has to
  altered later
• Post partitioning
• Design the system possibly start the
  implementation before the manner of the
  distributed architecture is defined

24
Lecture 5 Distributed System Design.

• Post partition
• Pros
• No constraint on the initial design
• Stand alone design methodologies can be used
• Cons
• Each element of the system once distributed will
  need to be able to map all other elements
• Leads to the idea of the Virtual Node (ADA)

25
Lecture 5 Summary

• Distributed systems are not simply a collection
  of standalone systems bolted together.
• Specific problems are unique to distributed
  design
• Design goals need to be carefully examined before
  a design can be undertaken