Better concurrency with a.NET language for isolation and asynchrony Niklas Gustafsson Parallel...

Better concurrency with a .NET language for isolation and asynchrony

Niklas GustafssonParallel Computing PlatformMicrosoft Corporation

[email protected]© 2009 Microsoft Corporation 1

Take-aways

• Concepts– Domains and agents– Channels and protocols– Schema– Dataflow networks

• Examples

© 2009 Microsoft Corporation

Layers of a Concurrent Program

• Asynchronous Agents– Isolation,

asynchrony• Concurrent

Collections– Parallelism,

decomposition• Interacting

Infrastructure– Consistency

Interacting Infrastructure

Asynchronous Agents

Concurrent Collections


Maestro Value Propositions

• Reduce complexity

• Provide a declarative model for serializing access to state

• Leverage latency

• Components default to concurrency

• Provide an application model that is inherently concurrent and responsive


Incubation Objectives

• Is language the better way to go?• Does our model reduce complexity?• Is it feasible and productive to incrementally adopt

Maestro for legacy apps?• Are we building it right?• Syntax and semantics• Missing concepts• Conceptual consistency and composition• Productivity• Platform fit


Maestro from 50,000 Feet

• Imperative• Actor-oriented• Special-purpose• Object-aware• Concurrent by default


• Ada

• Erlang

• SR (U of Arizona)

• Web-based Application Patterns


Maestro Inspiration

Is a parallel system that:Runs on hundreds of millions of coresIs known to scaleIs built by tens of thousands of programmers of

average skill working in thousands of organizationsIs remarkably robust

Maybe there are some things we could learn from how it's programmed?


The Web

Programming model aspects of relevance to manycore:

Loosely coupled

Strictly isolated components

Globally concurrent, locally sequential

Message-passing with simple protocols

Partial failure is expected and common


Web Programming

Sources Of Concurrency

• Asynchronous Agents– An agent has a separate “thread” of control

• Dataflow networks– Multiple nodes “light up” concurrently


Maestro: Core ConceptsDomain Agent

Channel

SchemaPattern


Maestro: Application Structure

Agent:ReaderAgent:Reader

Shared StateShared State

Agent:ReaderAgent:ReaderReader

AgentsReader Agents

WriterAgentWriterAgent


Domain Domain

Shared StateShared State

Reader Agent

Reader Agent




SchemaSchema

SchemaSchema

Channel w/ protocolChannel w/ protocol


Domain

• Passive construct• Scope of global state• Isolation boundary

domain A{ int j; int func( int k ) { } agent X : Channel1 { } agent Y : Channel2 { }}

domain B{ double m; string func( ) { } agent Z : Channel1 { }}


Agent

• Thread of control• Communicates with

other agents:– Intra-domain– Inter-domain

• Inheritance model• Reader/Writer

declaration

domain A{ int j; int func( int k ) { } writer agent X : Channel1 { } reader agent Y : Channel2 { }}

domain B{ double m; string func( ) { } agent Z : Channel1 { }}


Channel

• Agent proxy• Duplex message-

passing via ports• Enforces value passing

semantics• Bi-lateral channels only,

not a general "open mailbox" concept

channel ChAddition{ input int Add; output int Sum;}


Duality Of Ports


Num1

Channel

Num2

Sum

Num1

Num2

Sum

Agent Implside

Usingside

Each port is both a source and a target, depending on which side of the channel you are using it from, the channel perspective.

Pattern

• Enforces protocol rules• Special states: start

and end• For channels and

networks

channel ChAddition{ input int Add; output int Sum;

Start: { Add -> S1; } S1: { Sum -> End; }}


Schema

• Fields• Functions• Constants• Rules• Serializable• Strong name identity

schema A { required String Name; optional Int32 Age = 0;

rules { require String.IsNullOrEmpty(Name); require Age >= 0; } }

schema B { optional String Name; } B b = ...;A a = b; A a = {Name = “John”, Age = 39};


Dataflow Networks

• Think of data flowing through network– Not objects performing actions!

• Get data, do something with it, pass on to next stage(s)


Network Declarations

• Encapsulates DF layout• Exposes ports• Implementation details

hidden in private methods and fields

network N1{

input string RequestByKey; input int RequestBySize; output ResultRecord Result;

public N1() { … }

layout { RequestByKey ==> Search ==> Result; RequestBySize ==> szTbl.Find ==> Result; }

private Map<int,ResultRecord> szTbl;}


Dataflow Examples

• Compiler:– Characters flow through lexer turning into tokens;– Tokens go through grammar turning into AST;– AST goes through several passes

• Spreadsheet:– Change in a cell triggers reevaluation of

dependent cells;– Independent cells can change concurrently


Example: Image Processor


• Read JPG, Sharpen, Reduce “red eye” effect, Shrink, Save as 85% quality JPG

• Simple pipeline

Network n1 = Read ==> Sharpen ==> RedEye ==> Shrink ==> Save;

Example: Image Processor (2)


• Simple branching

Save

Network n1 = (Read ==> Sharpen ==> RedEye);n1 -<< { Save, Shrink ==> Save};

Example: Image Processor (3)


• Simple branching

Network n1 = (Sharpen ==> RedEye ==> Shrink);Network n2 = (Sharpen ==> RedEye ==> Shrink); // or: n1.Clone();Network n3 = (Sharpen ==> RedEye ==> Shrink); // or: n1.Clone();

Read -<: { n1, n2, n3 } >>- Save;

Sharpen Red Eye Shrink

Sharpen Red Eye Shrink

Concurrency in Network

• Same model: one writer or multiple readers– Reader/writer of what?– Notation: methods write, true functions read


Domain

ReaderAgent1

method1

method2

function1

function2

m1m2f1f2

WriterAgent2

m1m2f1f2

No AccessAgent3

m1m2f1f2


Example: Backup Utility

Example: Backup Utility

GUIController

Profiles

Backup

FolderMonitor

ConfigurationLog

Backup Sched


We're planning a early-access public release through Microsoft DevLabs in the spring.

Download itPlay with itBuild something with itTry to break itTell us what you thinkSpread the word


How to join the band…


Thank You

© 2009 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market

conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

Better concurrency with a.NET language for isolation and asynchrony Niklas Gustafsson Parallel...

Documents

Transcript of Better concurrency with a.NET language for isolation and asynchrony Niklas Gustafsson Parallel...