Distributed Systems 2006

1
Distributed Systems 2006

• Course Summary
• About Exam

2
Course Summary

• Summary of
• Fundamental Concepts
• Communication
• Remote Procedure Call
• Client/Server Computing
• Web Services
• Distributed Time and Commit
• Static and Dynamic Membership
• Group Communication
• Virtual Synchrony
• Retrofitting Reliability
• Executive summary style
• One slide per topic...

3
Aim of the course

• The main aim of the course is to introduce
  fundamental concepts and techniques for
  distributed systems
• The course gives the students prerequisites to
  analyse, design, and implement distributed
  systems.

4
Fundamental Concepts

• Definitions
• Process a program running on a computer
• Message arbitrary-sized data used to communicate
  between processes
• Network a collection of computers interconnected
  by hardware that supports message passing
• Distributed system processes that coordinate
  actions through message exchange, usually over a
  network
• Protocol an algorithm by which processes
  cooperate through message exchange
• Motivation
• Resource sharing, collaboration, ...
• Needs to need for performance, availability,
  fault tolerance, ...
• Characteristics
• Concurrent processes
• Message exchange
• No global clock
• Independent failures
• Reliability components of reliability, types of
  failures
• Models
• Real-world networks
• Asynchronous model
• Synchronous model

5
Communication

• Open Systems Interconnection (OSI) stack
• Messages travel conceptually between layers on
  the same level
• OSI Details
• Physical layer electrical and physical means for
  communicating (e.g, parts of Ethernet)
• Data link layer transit of data across physical
  link
• Network layer logical network layout
• Transport layer segments data
• Session layer coordinates request and response
  sessions between applications
• Presentation layer encoding and decoding of
  application data
• Application layer
• Networking hardware
• Ethernet, routers, switches/bridges
• Protocols
• The end-to-end argument
• Addressing
• IP, UDP, TCP, HTTP

6
Remote Procedure Call

• Remote Procedure Call (RPC)
• Mask distributed communication using transparent
  mechanism akin to local procedure calls
• But new kinds of exceptions
• Handling of complex data
• RPC protocols
• Basic, assuming no failures
• Reliable RPC
• Message loss, disorder, duplication
• Machine failure
• RPC semantics
• At most once
• Exactly once (impossible)
• At least once
• Java RMI
• Basic mechanisms
• RMI wire protocol
• Threading and RPC

7
Client/Server Computing

• Characteristics
• Clients use resources of server
• Little communication among clients
• Client to server ratio high
• Central concepts
• Caching stale/consistent, in-coherent/coherent
• Stateless/stateful
• Archetypical examples
• Distributed file systems
• Mimicking local file system
• SUN NFS as stateless example
• Transactional databases
• Transaction model ACID
• Serialization
• Concurrency control

8
Web Services

• Web architecture
• Applications, proxies, servers
• Relevant web technologies
• XML
• Well-formedness, validity, ...
• HTTP
• GET, HEAD, POST, PUT, ...
• Web services
• Interoperability as main motivation
• RPC via
• WSDL interface description via service
  definition and service binding
• SOAP defining XML messages
• HTTP transport protocol
• Performance
• XML vs binary encodings
• FastWS as a potential speed-up
• Reliability
• Web security
• WS_RELIABILITY (including WS_TRANSACTIONS) via
  intermediaries

9
Distributed Time and Commit

• Distributed time
• No concept of global time possible
• (Can come close with GPS time, though)
• Need logical clocks in many protocols
• Lamports happens-before relation
• Can order locally for process
• Send(a) happens-before receive(a)
• Consistent cuts now
• Supporting logical clocks
• Logical timestamp increment integer and send it
  with messages
• Vector timestamps increment integer per process
• Distributed commit
• Two-phase commit (2PC)
• 1) Transaction manager asks data managers for
  vote
• 2) Commit if all ready, abort if not
• Garbage collection, handling failure
• Three-phase commit (3PC)
• 2PC has bad state transaction manager
  process fails just after process learns commit
  what should others do?
• Add extra prepared to commit state all
  processes may learn outcome of vote before actual
  commit

10
Static and Dynamic Membership

• Which processes are available in a distributed
  system?
• Static membership
• Resolve liveness of potential list of members on
  a per-operation basis
• Slow, simple
• Dynamic membership
• Use group membership protocol to track members in
  the form of views
• Performant, complex
• Static quorum update and read
• A quorum read/write should intersect prior quorum
  write at at least one process to ensure that
  latest value is read/written
• Qr Qw gt N, Qw Qw gt N
• Uses a 2PC protocol for update
• Dynamic Group Membership Service (GMS)
• Tracks a set of processes join, leave, shun
  (accept false positives)
• GMS service itself needs to be fault tolerant
• Designate leader (oldest process) running 2PC
  protocol for view change
• Run 3PC if leader fails
• JGroups as an example of a Java toolkit
  supporting a GMS

11
Group Communication

• Multicast primitive
• Deliver message to members of a process group
• Unreliable multicast
• E.g., using UDP multicast
• Reliable multicast
• Non-uniform, failure-atomic
• Use acks to guide retransmits
• Use GMS for failure handling
• Dynamically uniform, failure-atomic
• Additional round for ensuring all have message
• Flush
• Ensure all members have seen same messages when
  view changes
• Ordered multicast
• fbcast/FIFO
• Number multicasts, handle failure
• cbcast
• Deliver in causal order, use vector timestamps
• abcast
• E.g., using a token to designate a sequencer of
  messages

12
Virtual Synchrony

• Elements of virtual synchrony
• Process groups with identical views and ranking
• Ordered reliable, multicast
• View-synchronous delivery
• Gap-freedom
• State transfer when joining
• Main insights
• Execution looks synchronous/state machine-like,
  but free to optimize
• Correlated failures less likely
• Uses of virtual synchrony
• Election choose oldest process as leader
• Consensus leader multicasts question, all reply,
  leader chooses (compare FLP)
• Consistent snapshot start multicast, done
  multicast use cbcast
• Replicated data read, update, lock operations
  abcast vs cbcast
• Load-balancing group decides vs client decides
• Fault tolerance e.g., primary-backup group of
  servers

13
Retrofitting Reliability

• Toolkit approach
• Provide reliability tools via a toolkit directly
  to programmers
• JGroups supporting virtual synchrony as example
• Mixed experiences
• Tricky to work with
• E.g., JGroups hidden in JBoss
• Wrappers
• Transparent wrapping of reliability tools
• Wrapping services, data, communication channels
• Examples
• CORBA fault tolerance
• Unbreakable TCP
• Reliable Distributed Shared Memory

14
Meta summary

• (not the complete picture, though...)

Robust Web Services
Toolkits and wrappers
Ordered multicast
Fault-tolerant multicast
Group Membership
2PC and 3PC
Communication
15
About Exam

• Curriculum
• Exam questions
• Trivial advice
• Where and when

16
Curriculum

• Birman, 2005
• Chapters 1, 2, 3, 4, 5, 8, 9, 10, 12, 13, 14
• Chapter 15 except Sections 15.2 and 15.3
• Chapter 16 except Section 16.2.5
• Chapter 18 except Section 18.2
• Chapter 20
• Slides from the lectures
• Solutions to hand-ins A1 to A5

17
Exam Questions

• Fundamental Concepts
• Communication
• Remote Procedure Call
• Client/Server Computing
• Web Services
• Distributed Time and Commit
• Static and Dynamic Membership
• Group Communication
• Virtual Synchrony
• Retrofitting Reliability

18
Exam

• Exam form
• Oral
• No preparation
• 13-grade scale
• Internal censor (Jonas Thomsen)
• Process at exam
• Pick question
• Present what you know about the question (for ca.
  12 minutes)
• Discuss with examiner and censor
• Max 20 minutes altogether

19
Trivial advice before the exam

• Read and practice ?
• Create outlines for each exam question
• (The summaries given here are not necessarily
  good ones)
• Map literature to questions
• Practice exam questions in groups
• Pick questions, hold test exams
• Fidelity, fix-points
• Go watch other exams if possible

20
Trivial advice during the exam

• Pick question, dont panic, look at your outline,
  put it away
• Use whiteboard as much as possible
• Structures your presentation and helps the
  subsequent evaluation
• Use transparencies of diagrams (max 2) for
  figures
• Listen to (and think about) questions from
  examiner and censor before answering
• Dont drink fizzy drinks ?

21
Trivial advice after the exam

• Unacceptable (00, 03, 5)
• Unable to present particular question
• Very limited knowledge and operational skills in
  topic
• Unable to distinguish relevant and irrelevant
• Insecure use of knowledge and skills on known
  problems
• Acceptable (6)
• Able to present particular question
• Knowledge and operational skills in topic
• Understanding of essence of course
• Mean (7, 8, 9)
• Good presentation of particular question
• Able to cover (and argue for) most relevant parts
  of particular questions
• Good knowledge and operational skills in topic
• Very good (10, 11, 13)
• Very good presentation of particular question
• Extensive, thorough, and secure knowledge and
  operational skills in topic
• Able to distinguish relevant from irrelevant
• Able to combine concepts, methods, information
  from different parts of the topic

22
Where and when

• Shannon-157, Finlandsgade 24 A
• Week 13
• Tuesday 2006-03-28 Friday 2006-03-31
• Starting at 900
• Exam lists are at Grædemuren in Ny Munkegade

23
Questions?