3.

Synchronization in Distributed
Systems
 In a centralized system: all processes reside
on the same system utilize the same clock.

 In a distributed system: like synchronize

everyone’s watch in the classroom.
Global Time
 Global Time is utilized to provide
timestamps for processes and data.
  Physical clock: concerned with
“People” time
  Logical clock: concerned with
relative time and maintain logical
consistency
Physical Clocks
 There are two aspects:

 Obtaining an accurate value for physical time

 Synchronizing the concept of physical time
throughout the distributed system

These can be implemented using centralized

algorithms or distributed algorithms
Obtaining an Accurate
Physical Time
 A physical time server is needed to access the
current time from a universal time coordinator
(UTC).
 Two sources for UTC:

 WWV shortwave radio station in Ft. Collins,

Colorado
 Geostationary Operational Environmental
Satellites (GEOS)
Synchronizing Physical Time
 The difference in time between two clocks
due to drifting is defined as clock skew. As
long as any and every two clocks differ by
a value less than the maximum skew value,
the time service is considered to be
maintaining synchronization.
How to synchronize two clocks
in A and B?
 The information necessary to read the value must
be communicated across the network to location
B.
 B’s clock value must be read.
 B’s clock value is communicated back to location
A.
 B’s clock value is adjusted to reflect the time
necessary to travel across the network.
 B’s clock value is compared to A’s clock value.
Centralized Physical Time
Services

 Broadcast Based

 Request Driven
Broadcast Based – first
approach
 The centralized time server’s action:
The physical time service broadcasts periodically the current time to
members of the distributed systems.
 The participants’ action:
 If a given participant’s clock is ahead of the time server’s clock, the
participant slows down its clock so that it will continually move
closer to the accurate time.
 If a participant’s clock is behind the time server’s clock, the
participant moves its clock forward. Alternatives do include gradually
speeding up the clock.
For example
Location A Time server
Current time Current time=740
= 720 Broadcast based
Delay of 10

Location A
Current time=720
Adjusted current time=750
New current time=750
 Broadcast Based – second
approach (Berkeley algorithm)
Location A Time Server Location B
1 Current time=740 1
2 Adjusted location A 2 Current time=732
Current time=720 =730 Delay=6
Delay=10
Adjusted location B Slow clock down to
4 =738 5 accommodate 2
Move forward=6 Average and the new
current time=736

1. Current time = 740

2. My current time = 720
3. My current time = 732
4. Adjust forward = 6
5. Adjust slowdown to accommodate 2
Request Driven
Location A Timer Server
Current time=730 Request for
current time Current time=740

Current time=740
Adjusted time=750 Delay=10

New current time=750

Distributed Physical Time
Service
 Each location broadcasts its current time at predefined set
intervals. Once a location has broadcast its time, it starts a
timer. It then collects time messages that it receives. Each
time message that arrives is stamped with the local current
time. This process continues until the timer expires. Upon
the expiration of the timer, each message is adjusted to
reflect the network delay time estimated for the message
source. At this stage, the participant calculates the average
time according to one of the following approaches:
 Calculate the average of all messages
Adjusted received times

720
724
726
718
722
723
 Delete the times that are above the
threshold and then average the rest.
Adjusted received times

The numbers besides X are deleted.

760 X The rest are averaged.
724
726
718
702 X
723
 Discard the highest x and the lowest x
values and then average
Adjusted received times

760 X
724
726
718
702 X
723
703 X
765 X
Logical Clocks
 Why Logical Clocks?
It is difficult to utilize physical clocks to
order events uniquely in distributed
systems.
 The essence of logical clocks is based on
the happened-before relationship presented
by Lamport.
Happen-Before Relationship
 If two events, a and b, occurred at the same
process, they occurred in the order of which they
were observed. That is, a > b.
 If a sends a message to b, then a > b. That is, you
cannot receive something before it is sent. This
relationship holds regardless of where events a
and b occur.

 The happen-before relationship is transitive. If a happens before b

and b happens before c, then a happens before c. That is, if a > b and
b > c, then a > c.
Logical Ordering
 If T(a) is the timestamp for event a, the following relationships must hold in a
distributed system utilizing logical ordering.

 If two events, a and b, occurred at the same

process, they occurred in the order of which they
were observed. That is T(a) > T(b).
 If a sends a message to b, then T(a) > T(b).
 If a happens before b and b happens before c,
T(a) > T(b), T(b) > T(c), and T(a) > T(c).
 For example
E F
Process 3

C D
Process 2

A B
Process 1

A>B>C>D>F E
Lamport’s Algorithm
 Each process increments its clock counter
between every two consecutive events.
 If a sends a message to b, then the message must
include T(a). Upon receiving a and T(a), the
receiving process must set its clock to the greater
of [T(a)+d, Current Clock]. That is, if the
recipient’s clock is behind, it must be advanced to
preserve the happen-before relationship. Usually
d=1.
 For example
E(1) F(5)
Process 3

C(3) D(4)
Process 2

A(1) B(2)
Process 1
 Total Ordering with Logical
Clocks
E(1.3) F(5.3)
Process 3

C(3.2) D(4.2)
Process 2

A(1.1) B(2.1)
Process 1

A>E>B>C>D>F
Mutual Exclusion
 In single-processor systems, critical regions
are protected using semaphores, monitors,
and similar constructs.
 In distributed systems, since there is no
shared memory, these methods cannot be
used.
A Centralized Algorithm
coordinator
process Request
Grant

Enter crical Release

section

Exit

 Advantages: It is fair, easy to implement, and requires only three

messages per use of a critical region (request, grant, release).
 Disadvantages: single point of failure.
Distributed Algorithm
OK OK
REQ REQ

REQ REQ
Token Ring Algorithm
 A Comparison of the Three
Algorithms
Algorithm Messages Delay Problems
per entry/exit before entry

Centralized 3 2 Coordinator crash

Distributed 2(n-1) 2(n-1) Crash of any process

Token ring 1 to ∞ 0 to n-1 Lost token, process

crash
Election Algorithm
 The bully algorithm
 When a process notices that the coordinator is no
longer responding to requests, it initiates an
election. A process, P, holds an election as
follows:
 P sends an ELECTION message to all processes
with higher numbers.
 If no one responds, P wins the election and
becomes coordinator.
 If one of the higher-ups answers, it takes over.
P’s job is done.
For example
1 1 1
7 7 7 2
2 2
Election
6 6 6 3
Election 3 3
5 4 5 4 5 4
Ok

1 1
7 2 7 2
Coordinator
6 Ok 3 6 3
5 4 5 4
 A Ring Algorithm
234561
1
2 6
2
7

3
23
23456 6

4
5

2345 234
Atomic Transactions
 All the synchronization techniques we have
studied so far are essentially low level, like
semaphores.
 What we would really like is a much
higher-level abstraction such as atomic
transaction.
For example
 Atomic bank transactions:
1. Withdraw(amount, account1)
2. Deposit(amount, account2)
Stable Storage
 Stable storage is designed to survive
anything except major calamities such as
floods and earthquakes.
 Stable storage can be implemented with a
pair of ordinary disks.
 Stable storage is well suited to applications
that require a high degree of fault
tolerance, such as atomic transactions.
 Stable storage Stable storage Stable storage

Drive 1 s a s a’ s a
o h o h o h
t f t f t f
b w b w b w

s a s a s a
Drive 2

t f t f t f
b w b w b w

(a) Stable storage (b) Crash after drive 1 is updated © Bad spot
Transaction Primitives
 1 BEGIN_TRANSACTION: Mark the start of a transaction.
 2 END_TRANSACTION: Terminate the transaction and try to commit.
 3 ABORT_TRANSACTION: Kill the transaction; restore the old values.
 4 READ: Read data from a file (or other object).
 5 WRITE: Write data to a file (or other object).

 For example,
 BEGIN_TRANSACTION
 reserve Austin-Houston;
 reserve Houston-Los Angeles;
 reserve Los Angeles-Seatle;
 END_TRANSCATION
Properties of Transactions
 1 Atomic: To the outside world, the
transaction happens indivisibly.
 2 Consistent: The transaction does not
violate system invariants.
 3 Isolated: Concurrent transactions do
not interfere with each other.
4 Durable: Once a transaction commits,
the changes are permanent.
Isolated or serializable
 Isolated or serializable means that if two or
more transactions are running at the same
time, to each of them and to other
processes, the final result looks as though
all transactions ran sequentially in some
(system dependent) order.
An example
 BEGIN_TRANACATION
 X = 0;
 X=X+1;
 END_TRANSACTION
 (a)
 BEGIN_TRANSACTION
 X=0;
 X= X+2;
 END_TRANSACTION
 (b)
 BEGIN_TRANSACTION
 X=0;
 X=X+3;
 END_TRANSACTION
 (c )
Schedule 1 x=0; x=x+1; x=0; x=x+2; x=0; x=x+3; legal

Schedule 2 x=0; x=0; x=x+1; x=x+2; x=0; x=x+3; legal

Schedule 3 x=0; x=0; x=x+1; x=0; x=x+2; x=x+3; illegal

Nest Transactions
 Transactions may contain subtransactions,
often called nested transactions.

 If the subtransaction commits and the

parent transaction aborts, the permanence
applies only to top-level transactions.
Implementation
Private workspace
 Private Workspace
Index
Index Index 0’ 0
0 1 1
0
1 2 2
1
2 2 3’ 3

1 2 0 1 2 0 1 2

0’ 3’ 0 3
 Writeahead log
x=0;
y=0;
BEGIN_TRANSACTION
x=x+1; log: x=0/;
y=y+2; log: x=0/1; y=0/2;
x=y * y; log: x=0/1; y=0/2; x=1/4;
END_TRANSACTION
 Achieving atomic commit in a
distributed system
 Two-Phase Commit Protocol
Coordinator Subordinates
Write “Prepare” in the log
Send “Prepare” message
Phase 1 Write “Ready” in the log
Send “Ready” message

Collect all replies

Write log record (if all are ready, commit; if not, abort)
Send “Commit” message
Phase 2 Write “Commit” in the log
Commit
Send “Finished” message
Concurrency Control
 When multiple transactions are executing
simultaneously in different processes, some
mechanism is needed to keep them out of
each other’s way. That mechanism is called
a concurrency control algorithm.
Concurrency control
algorithms
 Locking
 In the simplest form, when a process needs to
read or write a file (or other object) as part of a
transaction, it first locks the file.
 Distinguishing read locks from write locks.
 The unit of locking can be an individual record
or page, a file, or a larger item.
 Two-phase locking
 The process first acquires all the locks it needs
during the growing phase, then releases them
during the shrinking phase.
 In many systems, the shrinking phase does not
take place until the transaction has finished
running and has either committed or aborted. This
policy is called strict two-phase locking.
 Two-phase locking
Lock point

Growing phase Shrinking phase

Number
of locks

Time
 Optimistic Concurrency Control
A second approach to handling multiple
transactions at the same time is optimistic
concurrency control. The idea is simple:
just go ahead and do whatever you want to,
without paying attention to what anybody
else is doing. If there is a problem, worry
about it later.
  Timestamps
Write
TRD TWR T T TRD
() () () () ()
Do tenative Abort
write
TWR TRD T T TWR
() () () () ()
Abort
Do tentative
write
 Read

TWR T T TWR
() () () ()
Ok Abort

TWR TTENT T T TTENT

() () () () ()
Wait Abort