Concurrent Message Processing using Transactional

Memory in the Actor ModelWTM 2013

Pascal Felber, Derin Harmanci, Yaroslav Hayduk and Anita Sobe

first.last@unine.ch

Yaroslav Hayduk 2

The Actor Model

• Hewitt & Baker - „Laws for Communicating Parallel Processes“

• Motivated by the prospect of highly parallel computing machines with many microprocessors + own local memory

Yaroslav Hayduk 3

The Actor Model – basic principles

• everything is an actor (VS an object)• asynchronous message passing• has access to its local state only • strong encapsulation• inherently concurrent

Yaroslav Hayduk 4

OOP and Actors: Communication

Object A Object B

Actor A Object B

ObjectB.publicField=10ObjectB.publicMethod()

ActorA [SendMessageTo] ActorB

VS

asynchronous message passing

direct access

5

Essential Actor properties

• Processes one message at a time– Yey! No race conditions or other hazards

• No (access to) shared state • Can send messages to other actors only• Switch its behavior

Yaroslav Hayduk

6

Behind the (Actor) scenes

Yaroslav Hayduk

Mailbox…Contains 2 Insert 3

2 5 6 8 9

Local Actor state

In progress:Contains 8

C

A

B

A “List” Actor

Remove 9

Yaroslav Hayduk 7

Main issues

1.The processing of one message at a time

Yaroslav Hayduk 8

Notable Related Work

• Scholliers et al. – Parallel Actor Monitors– does not use STM for processing messages

concurrently• Shams et al. – Habanero Scala– async-finish programming model– processes parts of one message (and not many

messages) concurrently

Yaroslav Hayduk 9

Idea:

Process multiple messages concurrently

• Problem: the list can be easily corrupted• Solution: Wrap each message processing in an

STM transaction. When conflicts happen, (roll back and) repeat the message processing

Yaroslav Hayduk 10

Concurrent message processing

Actor Mailbox…Contains 2 Insert 3

2 5 6 8 9

Local Actor state

A “List” Actor

Messages In progress

Contains 6

Insert 12

… Insert 12

Each message is processed in a transaction

11

Coordinated Transactions

Yaroslav Hayduk

A

A List Range (1..25) Actor

Global list sum

A List Range (100..125) Actor

……….

Yaroslav Hayduk 12

Local Actor Sum: Zoom-in

2 5 6 8 9

Local Actor state

A List Range Actor

Messages In progress

1. Calculate local sum2. Forward new partial sum3. Block (for consistency)

New partial sumPrev partial sum

Yaroslav Hayduk 13

Main issues

2.Extensive blocking of actors involved in a

coordinated transaction

Yaroslav Hayduk 14

Idea:

Remove blocking; process other messages speculatively

• Pre-commit the local coordinated transaction• Process other messages in a transaction

speculatively

Yaroslav Hayduk 15

Non-blocking coordinated transactions

2 5 6 8 9

Local Actor state

A List Range Actor

Messages In progress

Saved: Partially committed local coordinated transaction

Actor Mailbox…Contains 2 List sum

Speculatively process:Insert 3

Yaroslav Hayduk 16

Evaluation

50 100 250 500 750 10000

10000

20000

30000

40000

50000

60000

List Actors = 8, Writes = 98%, Sum = 1%

Regular Concurrent Coordinated

Request Actors

Mill

is

Yaroslav Hayduk 17

Evaluation

50 100 250 500 750 10000

10000

20000

30000

40000

50000

60000

70000

List Actors = 16, Writes = 98%, Sum = 1%

Regular Concurrent Coordinated

Request Actors

Mill

is

Yaroslav Hayduk 18

Summary

• By using speculation, we can achieve a higher message throughput in the Actor Model

• By using STM we guarantee that the Actor’s state is never corrupted

Thank you Questions?