SUPPORTING LOCK-FREE COMPOSITION OF CONCURRENT DATA OBJECTS

Daniel Cederman and Philippas Tsigas

Overview

SUPPORTING OFCONCURRENT DATA

OBJECTS

COMPOSITION

LOCK-FREE

Concurrent Data Objects

Data structures shared between processes

and other …

Trees Skiplists

Queues

HashtablesFor example …

Synchronization Synchronization is required for

concurrency Mutual exclusion

Locks limits concurrency Busy waiting – repeated checks to see if

lock has been released or not Convoying – processes stack up before

locks A better approach is to use data

structuresthat are …

OverviewCONCURRENT DATA

OBJECTS

COMPOSITION

LOCK-FREE

Lock-free Lock-freedom is a progress guarantee

In practice it means that A fast process doesn’t

have to wait for a slowor dead process

Can’t be designed withblocking parts

No deadlocks

Shown to scale better than blocking approaches

DefinitionFor all possible

executions, at least one concurrent

operation will succeed in a finite number of its

own steps

Libraries using Lock-free Data Objects

Java Concurrency Package Threading Building Blocks by Intel .NET Parallel Extensions NOBLE Non-Blocking Library …

OverviewCONCURRENT DATA

OBJECTS

COMPOSITION

LOCK-FREE

Composition

FC

DC

A B C

FC

DA B C

Dequeue Insert

DequeueInsert

Queue Tree

Queue Tree?

Challenge Providing efficiency and correctness

while dealing with Specialized designs Few common algorithmic components Complex proofs of correctness

We target a large class of concurrent data objects

Concurrent Data Objects

Have operations for insertion and removal

of elements

Can be composed to form move operations

Trees Skiplists

Queues

Hashtables

and other …

Contributions We provide a framework that consists of

three parts Properties used to identify compatible

objects Steps needed to adapt object Algorithmic design of operation for

performinglock-free moves between adapted objects

Characterization1. Have operations equivalent to insert

and remove2. These operations are linearizable3. …4. …

Linearizability

Operation A Operation C

Operation B Operation D

Linearizable if for any concurrent history there exists a correct sequential history where …

A happens before B, if A finished before B

started

Either C happens before D

or D happens before C,if C and D are concurrent

Linearization Points

Operation A Operation C

Operation B Operation D

Linearization point

Linearization point

Linearization point

Linearization point

Linearization Points

Operation A Operation C

Operation B Operation D

Characterization1. Have operations equivalent to insert

and remove2. These operations are linearizable3. …4. …

Insert Element into BRemove Element from A

Composition

Composition

Remove - prolog

Insert - prolog

Remove - epilog

Insert - epilog

Composition

Remove - prolog

Insert - prolog

Remove - epilog

Insert - epilog

Element to remove must be accessible here!

Characterization1. Have operations equivalent to insert

and remove2. These operations are linearizable3. The element to remove is accessible

before the linearization point4. …

Composition

Remove - prolog

Insert - prolog

Remove - epilog

Insert - epilog

Linearization point is often a successful compare-and-swap

Composition

Remove - prolog

Insert - prolog

Remove - epilog

Insert - epilog

Prolog

Epilog

if(CAS(…))Only fails if

other process succeeds

Failed! Failed!Success! Failed!Failed! Success!

Success! Success!

RemoveInsert DCASFailed!Failed!Failed!

Success!

Composition

Remove - prolog

Insert - prolog

Remove - epilog

Insert - epilog

Combined using a double-word

compare-and-swap

Characterization1. Have operations equivalent to insert

and remove2. These operations are linearizable3. The element to remove is accessible

before the linearization point4. The linearization point for a successful

operation is a successful compare-and-swap

Can be composed to move operations

Compatible Concurrent Data Objects

There are a wide variety of commonly used lock-free data structures that supports these requirements Queues [PODC ‘96] Lists [PODC ‘04] Skip-Lists [IPDPS ‘03] Priority Queues [JPDC ‘05] Hash-tables [SPAA ‘02] Dictionaries [SAC ‘04] Stacks [Treiber ‘86] …

Reverts to normal compare-and-swap if used outside

move operation

Move Operation

PrologSCAS1

PrologSCAS2

EpilogEpilog

Remove operation Insert operationMove support

Performs double-word compare-and-swap

Can only fail if other process succeeds

Case Study - Stackbool pop(value)

while(true)ltop = top;if(ltop == 0) return false;value = ltop.value;if(cas(top, ltop, ltop.next))

return true;

Not a successful operation

Successful operation connected to a

successful CAS

Accessible before linearization point

Case Study - Stackbool pop(value)

while(true)ltop = top;if(ltop == 0) return false;value = ltop.value;if(scas(top, ltop, ltop.next, value))

return true;

The scas is called with the value to move

Generic Move Operation

Performance Evaluation The evaluation was performed on a

machine with an Intel Core i7 950 3GHz processor and 6GB DDR3-1333 memory

4 Cores with Hyper-ThreadingEnqueue Dequeue

Move

Enqueue Dequeue

Queue – Insert and Remove

1 2 4 6 8 10 12 14 160100200300400500600700

Blocking

Threads

Tim

e (m

s)

Queue – Move Operations

1 2 4 6 8 10 12 14 160

500

1000

1500

2000

2500Blocking

Threads

Tim

e (m

s)

Queue – Insert/Remove/Move

1 2 4 6 8 10 12 14 160200400600800

1000120014001600

Blocking

Threads

Tim

e (m

s)

1 2 4 6 8 10 12 14 160

50

100

150

200

250

Before After

Threads

Tim

e (m

s)Queue – Before/After Adaptation

Summary We provide a framework that consists of

three parts Properties used to identify compatible objects Steps needed to adapt object Algorithmic design of operation for performing

lock-free moves between adapted objects

Adaptation does not affect standard operations

For more information:www.cse.chalmers.se/research/group/dcs

Thank you!