8/3/2019 M S 13 DD R Performance Barrier Mechnisms

1/14

Parallel and Distributed Simulation

Deadlock Detection & Recovery: Performance

Barrier Mechanisms

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

2/14

Outline

Deadlock Detection and Recovery Algorithm

Empirical performance measurements

Synchronous Algorithms

Barrier mechanisms

Centralized Barriers Tree Barrier

Butterfly Barrier

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

3/14

PerformanceT = arrival time of jobQ = waiting time in queue

S = service time

Example: Tandem first-come-first-serve queues

Classical approach: lookahead?

LP1 LP2

arrivalevent

departureevent

arrivalevent

T

T+Q

T+Q+S

begin service

Optimized to exploit lookahead

LP1 LP2

arrivalevent

arrivalevent

T

T+Q

T+Q+S

Maintain variable indicating departure time of previous job

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

4/14

Efficiency of Queueing Network Simulation

1

10

100

1000

10000

1 2 4 8 16 32 64 128 256

Number of Jobs in Network

MessagesperDeadlock Optimized(deterministic service

time)Optimized (exponentialservice time)

Classical (deterministic

service time)

Classical (exponentialservice time)

Parallel Simulationof a Central ServerQueueing Network

Deadlock Detection and

Recovery Algorithm(5 processors)

merge fork

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

5/14

Speedup of Queueing Network Simulation

0

1

2

3

4

1 2 4 8 16 32 64 128 256

Number of Jobs in Network

Speedup

Optimized(deterministic servicetime)Optimized (exponentialservice time)

Classical (deterministicservice time)

Classical (exponentialservice time)

Deadlock Detection and Recovery Algorithm(5 processors)

Exploiting lookahead is essential to obtain good performance

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

6/14

Synchronous ExecutionBasic idea: each process cycles through the

following steps: Determine the events that are safe to process

Process events, exchange messages

Global synchronization (barrier)

Messages generated in one cycle are not eligiblefor processing until the next cycle

Issues

Barrier mechanism, transient messages

Determining safe events

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

7/14

Barrier Synchronization

Barrier Synchronization: when a process invokes the barrierprimitive, it will block until all other processors have alsoinvoked the barrier primitive.

When the last process invokes the barrier, all processes can

execute forward

- barrier -

- barrier -- barrier -

waitwait

wait

- barrier -

process 1 process 2 process 3 process 4

wallclocktime

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

8/14

Barrier ImplementationCentralized Message-Passing Approach

Central controller used to implement barrier 2 step process

Determine when barrier reached

Broadcast message to release processes from the barrier

Barrier primitive for non-controller processes: Send a message to central controller

Wait for a reply

Barrier primitive for controller process

Receive barrier messages from other processes When a message is received from each process, broadcast

message to release barrier

Performance

Controller must send and receive N-1 messages

Potential bottleneck

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

9/14

Broadcast Barrier

1 step approach

Each process broadcasts message when it reaches barrier

Wait until a message is received from each other process

N (N-1) messages

0 1 2 3

0 1 2 3

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

10/14

Tree Barrier

Organize processes into a tree

A process sends a message to its parent process when

The process has reached the barrier point, and

A message has been received from each of its children processes

Root detects completion of barrier, broadcast message torelease processes (e.g., send messages down tree)

2 log N time if all processes reach barrier at same time

3

87

4

109

5

1211

6

13

1 2

0

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

11/14

Butterfly Barrier N processes (here, assume N is a power of 2)

Sequence of log2 N pairwise barriers (let k = log2 N) Pairwise barrier:

Send message to partner process

Wait until message is received from that process

Process p: bkbk-1 b1 = binary representation of p

Step i: perform barrier with process bk bi b1(complement ith bit of the binary representation)

Example: Process 3 (011) Step 1: pairwise barrier with process 2 (010)

Step 2: pairwise barrier with process 1 (001)

Step 3: pairwise barrier with process 7 (111)

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

12/14

Butterfly Barrier Example

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7

step 1

step 2

step 3

Wallclocktime

0 1 2 3 4 5 6 7

0,1 2,3 4,5 6,7

0-3 4-7

0-7

step 1

step 2

step 3

The communication pattern forms a tree from the perspective of any process

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

13/14

0 1 2 3 4 5 6 7

0 2 4 6

0 4

0

1 3 5 7

1 5

1

2 6

2

3 7

3 4 5 6 7

Butterfly: Superimpose Trees

After log2 N steps each process is notified that the

barrier operation has completed

8/3/2019 M S 13 DD R Performance Barrier Mechnisms

14/14

Summary

Deadlock detection and recovery algorithm

Performance critically dependent on lookahead

Barrier mechanisms

Simple barriers using broadcast or central controller

OK for small number of processors Tree or butterfly give more scalable performance