SLIDE 1
EECS 591 DISTRIBUTED SYSTEMS
Manos Kapritsos Fall 2020
EECS 591 D ISTRIBUTED S YSTEMS Manos Kapritsos Fall 2020 S TATE M - - PowerPoint PPT Presentation
EECS 591 D ISTRIBUTED S YSTEMS Manos Kapritsos Fall 2020 S TATE M - - PowerPoint PPT Presentation
EECS 591 D ISTRIBUTED S YSTEMS Manos Kapritsos Fall 2020 S TATE M ACHINE R EPLICATION M ODELING FAULTS Mean Time To Failure/Mean Time To Recover used mostly for disks of questionable value in expressing reliability Threshold: out of makes
SLIDE 2
SLIDE 3
MODELING FAULTS
Mean Time To Failure/Mean Time To Recover used mostly for disks
- f questionable value in expressing reliability
Threshold: out of makes condition for correct operation explicit measures fault-tolerance of the architecture, not
- f individual components
Enumerate failure scenarios
SLIDE 4
A HIERARCHY OF FAILURE MODELS
Crash Fail-stop Send omission Receive omission General omission Arbitrary (Byzantine) failures = benign failures
SLIDE 5
A HIERARCHY OF FAILURE MODELS
crash
SLIDE 6
FAULT TOLERANCE: THE PROBLEM
Clients Server Solution: replicate the server
SLIDE 7
REPLICATION IN TIME
When a server fails, restart it or replace it Failures are detected, not masked Lower maintenance, lower availability Tolerates only benign failures
SLIDE 8
REPLICATION IN SPACE
Run multiple copies of a server (replicas) Vote on replica output Failures are masked High availability and can tolerate arbitrary failures but at high cost
SLIDE 9
THE ENEMY: NON-DETERMINISM
An event is non-deterministic if its output is not uniquely determined by its input The problem with non-determinism: Replication in time: must reproduce the original
- utcome of all non-deterministic events
Replication in space: each replica must handle non- deterministic events identically
SLIDE 10
THE SOLUTION: STATE MACHINES
Design the server as a deterministic state machine 1 3 4 2 a b c d e f
SLIDE 11
THE SOLUTION: STATE MACHINES
State machine example: a switch
- ff
- n
click click
SLIDE 12
Ingredients: a server
- 1. Make server deterministic (state machine)
- 2. Replicate server
- 3. Ensure that all replicas go through the same
sequence of state transitions
STATE MACHINE REPLICATION
= x = 1 x=2
- 4. Vote on replica outputs
SLIDE 13
Ingredients: a server
- 1. Make server deterministic (state machine)
- 2. Replicate server
- 3. Ensure that all replicas go through the same
sequence of state transitions
STATE MACHINE REPLICATION
x = 1 x=2
- 4. Vote on replica outputs
All state machines receive all commands in the same order
SLIDE 14
Ingredients: a server
- 1. Make server deterministic (state machine)
- 2. Replicate server
- 3. Ensure that all replicas go through the same
sequence of state transitions
STATE MACHINE REPLICATION
- 4. Vote on replica outputs
SLIDE 15
…
’
When in trouble, cheat!
Voter and client share fate!
- 4. Vote on replica outputs
SLIDE 16
ADMINISTRIVIA
Send me your paper preferences by tonight Send me your group declaration preferences by Oct 1 Homework #2 will be sent out later today due Monday, Oct 12, before class Implementation project will be out next Monday due Monday October 26, by end of day Research project topics due next Thursday, 10/08
SLIDE 17
PRIMARY-BACKUP
SLIDE 18
THE MODEL
Failure model: crash Network model: synchrony All messages are delivered within time Reliable, FIFO channels Tolerates crash failures
SLIDE 19
THE IDEA
Clients communicate with a single replica (primary) Primary: sequences and processes clients’ requests updates other replicas (backups) Backups use timeouts to detect failure of primary On primary failure, a backup becomes the new primary
SLIDE 20
A SIMPLE PRIMARY-BACKUP PROTOCOL
request reply sync new primary
Passive replication: sync = state update Active replication: sync = client request(s)
SLIDE 21
WEAKENING THE MODEL
Failure model: crash Network model: synchrony Unreliable, FIFO channels Channels may drop messages All messages are delivered within time (looks paradoxical) Tolerates crash failures
SLIDE 22
A SLIGHTLY DIFFERENT PRIMARY-BACKUP PROTOCOL
request reply sync new primary ack
SLIDE 23
GENERALIZING TO MORE BACKUPS
Primary backups
SLIDE 24
GENERALIZING TO MORE BACKUPS
Primary backups
update
SLIDE 25
GENERALIZING TO MORE BACKUPS
Primary backups
update
SLIDE 26
GENERALIZING TO MORE BACKUPS
(active updates) Primary backups
SLIDE 27
GENERALIZING TO MORE BACKUPS
(passive updates) Primary backups
SLIDE 28
GENERALIZING TO MORE BACKUPS
(passive updates) Primary backups
SLIDE 29
GENERALIZING TO MORE BACKUPS
ack ack ack ack
Primary backups
ack
SLIDE 30
GENERALIZING TO MORE BACKUPS
Primary backups
reply
SLIDE 31
HANDLING QUERIES
Primary backups
query
SLIDE 32
HANDLING QUERIES
Primary backups
SLIDE 33
HANDLING QUERIES
Primary backups
reply
However…
SLIDE 34
HANDLING QUERIES
Primary backups
query
SLIDE 35
HANDLING QUERIES
ack ack ack ack
Primary backups
The primary cannot respond until it has received all acks for prior updates
query ack