Concurrent Revisions:A deterministic concurrency model.

Daan Leijen & Sebastian BurckhardtMicrosoft Research

(invited talk at FOOL 2010)

Algorithmic/Scientific

Special-Purpose Interactive/Reactive

Concurrent ProgrammingThreads & Locks,

Futures, Promises, Transactions, ...

?Our Focus.

Parallel Programming

TPL, Fortran, MPI, Cilk, StreamIt, X10, Cuda, ...

Operating System

Database

Data MiningBiology

ChemistryPhysics

MultimediaSignal Processing

Games

Desktop Applications

Application = Shared Data and Tasks

Shared Data

ReaderMutator

Mutator

Mutator

Mutator Reader

ReaderR R R R

Example: Office application• Save the document• React to keyboard input by the user• Perform a spellcheck in the background• Exchange updates with collaborating remote users

Spacewars!

Example 1: read-write conflict Render task reads position of all game objects Physics task updates position of all game objects

=> Render task needs to see consistent snapshotExample 2: write-write conflict Physics task updates position of all game objects Network task updates position of some objects

=> Network has priority over physics updates

Examples from SpaceWars Game

Conflicting tasks can not efficiently execute in parallel. pessimistic concurrency control• use locks to avoid parallelism where

there are (real or potential) conflicts optimistic concurrency control• speculate on absence of conflicts

rollback if there are real conflicts

either way: true conflicts kill parallel performance.

Conventional Concurrency Control

Our Proposed Programming Model: Revisions and Isolation Types

• Deterministic Conflict Resolution, never roll-back• Full concurrent reading and writing of shared data• No restrictions on tasks (can be long-running, do I/O)• Clean semantics• Fast and space-efficient runtime implementation

RevisionA logical unit of work

that is forked and joined

Isolation TypeA type which implements

automatic copying/merging of versions on write-write conflict

What’s new

• Isolation: side effects are only visible when the revision is joined.

• Deterministic execution!

int x = 0;task t = fork { x = 1;}assert(x==0 || x==1);join t;assert(x==1);

versioned<int> x = 0;revision r = rfork { x = 1;}assert(x==0);join r;assert(x==1);

Traditional Task Concurrent Revisions

Isolation types:declares shared data fork revision:

forks off a private copy of the shared state

join revision:waits for the revision to

terminate and writes back changes into the main revision

isolation:Concurrent modifications

are not seen by others

int x = 0;int y = 0;task t = fork { if (x==0) y++;}if (y==0) x++;join t;

int x = 0;int y = 0;task t = fork { atomic { if (x==0) y++; }}atomic { if (y==0) x++; }join t;

versioned<int> x = 0;versioned<int> y = 0;revision r = rfork { if (x==0) y++;}if (y==0) x++;join r;

Sequential Consistency

Transactional Memory

Concurrent Revisions

assert( (x==0 && y==1) || (x==1 && y==0) || (x==1 && y==1));

assert( (x==0 && y==1) || (x==1 && y==0));

assert(x==1 && y==1);

Conflict resolution

x = 0

x = 1 x = 2

assert( x==2 )

x = 0

x = 1 x = 0

assert( x==0 )

x = 0

x = 1

assert( x==1 )

By default, on a write-write conflict (only), the modification in the child revision wins.

Versioned<int> x;

Custom conflict resolution

x = 0

x += 1

x += 2

assert(x==3)

merge(1,2,0) 3

Cumulative<int,(main,join,orig).main + join – orig> x;

1

2

0

x = 0

x += 1

x += 2

merge(1,2,0) 3

x += 3

assert( x==6 )

merge(3,5,2) 6

1 2

0

3 5

2

A Software engineering perspective• Transactional memory:

Code centric: put “atomic” in the code Granularity: • too broad: too many conflicts and no parallel

speedup• too small: potential races and incorrect code

• Concurrent revisions: Data centric: put annotations on the data Granularity: group data that have mutual constraints

together, i.e. if (x + y > 0) should hold, then x and y should be versioned together.

• For each versioned object, maintain multiple copies Map revision ids to versions `mostly’ lock-free array

• New copies are allocated lazily Don’t copy on fork… copy on first write after fork

• Old copies are released on join No space leak

Current Implementation: C# library

Revision Value

1 0

40 2

45 7

Demo class Sample{ [Versioned] int i = 0;

public void Run() { var t1 = CurrentRevision.Fork(() => { i += 1; }); var t2 = CurrentRevision.Fork(() => { i += 2; }); i+=3;

CurrentRevision.Join(t1); CurrentRevision.Join(t2); Console.WriteLine("i = " + i); }}

Demo: Sandboxclass Sandbox{ [Versioned] int i = 0;

public void Run() { var r = CurrentRevision.Branch("FlakyCode"); try { r.Run(() => { i = 1; throw new Exception("Oops"); }); CurrentRevision.Merge(r); } catch { CurrentRevision.Abandon(r); } Console.WriteLine("\n i = " + i); }}

Fork a revision without forking an associated task/thread

Run code in a certain revision

Merge changes in a revision into the main one

Abandon a revision and don’t merge its changes.

By construction, there is no ‘global’ state: just local state for

each revision

State is simply a (partial) function from a location to a

value

Operational SemanticsFor some revision r, with snapshot and local modifications and an

expression context with hole (x.e) v

the state is a composition of the root snapshot and

local modifications

On a fork, the snapshot of the new

revision r’ is the current state: ::

On a join, the writes of the joinee r’ take priority over the writes of the current

revision: ::’

Custom merge: per location (type)

On a join, using a merge function.

No conflict if a location was not

written in the joinee

No conflict if a location was unmodified in the current revision, use

the value of the joinee

Conflict otherwise, use a location/type specific

merge function

Standard merges:

What is a conflict?• Merge is only called if:

(1) write in child, and (2) modification in main revision:

Cumulative<int> x = 0

x += 2

x += 3

assert( x = 5 )

0

2 5

2

No conflict (merge function

is not called)

No conflict (merge function

is not called)

0

2

Merging with failure

On fail, we just ignore any writes in the joinee

Snapshot isolation• Widely used in databases, for example Oracle

and Microsoft SQL • In essence, in snapshot isolation a concurrent

transaction can only complete in the absence of write-write conflicts.

• Our calculus generalizes snapshot isolation: We support arbitrary nesting We allow custom merge functions to resolve

write-write conflicts deterministically

Snapshot isolationWe can succinctly model snapshot isolation as:• Disallow nesting• Use the default merge:

Some versions of snapshot isolation do not treat silent writes in a transaction as a conflict:

Sequential merges• We can view each location as an abstract data

types (i.e. object) with certain operations (i.e. methods).

• If a merge function always behaves as if concurrent operations for those objects are sequential, we call it a sequential merge.

• Such objects always behave as if the operations in the joinee are all done sequentially at the join point.

Sequential merges• A merge is sequential if:

merge(uw1(o), uw2(o), u(o))=

uw1w2(o)

• And uw1w2(o)

u

w1 w2

merge(uw1(o), uw2(o), u(o))

x = o

Abelian merges• For any abstract data type that forms an

abelian group (associative, commutative, with inverses) with neutral element 0 and an operation , the following merge is sequential:

merge(v,v’,v0) = v v’ v0

• This holds for example for additive integers and additive sets.

SpaceWars Game

Shared State

Parallel Collision DetectionParallel Collision DetectionParallel Collision Detection

Graphics Card

Network Connection

Play Sounds

Render Screen

Process InputsAutosave

Send

Receive

Disk

Key-board

SimulatePhysics

Sequential Game Loop:

Revision Diagram for Parallelized Game Loop

Coll.

Det

. 1

Coll.

Det

. 2

Coll.

Det

. 3

Coll.

Det

. 4Re

nder

Phys

ics

netw

ork

auto

save

(lo

ng ru

nnin

g)

“Problem Example 1” is solved

Render task reads position of all game objects

Physics task updates position of all game objects

No interference! Coll.

Det

. 1

Coll.

Det

. 2

Coll.

Det

. 3

Coll.

Det

. 4Re

nder

Phys

ics

netw

ork

auto

save

(lo

ng ru

nnin

g)

“Problem Example 2” is solved.

Physics task updates position of all game objects

Network task updates position of some objects

Network updates have priority over physics updates

Order of joins establishes precedence!

Coll.

Det

. 1

Coll.

Det

. 2

Coll.

Det

. 3

Coll.

Det

. 4Re

nder

Phys

ics

netw

ork

auto

save

(lo

ng ru

nnin

g)

Autosave now perfectly unnoticeable in background Overall Speed-Up:

3.03x on four-core (almost completely limited by graphics card)

Results

Overhead:How much does all the copying and the indirection cost?

Revisions and Isolation Types simplify the parallelization of applications with tasks that Exhibit conflicting accesses to shared data Have unpredictable latency Have unpredictable data access pattern May perform I/O that can not be rolled back

Revisions and Isolation Types are easy to reason about (determinism, isolation) have low-enough overhead for many applications

Conclusion

Questions?• daan@microsoft.com• sburckha@microsoft.com

• Download available soon on CodePlex