Inherent limitations facilitate design and verification of concurrent programs Hagit Attiya...

Post on 28-Mar-2015

214 views 0 download

Transcript of Inherent limitations facilitate design and verification of concurrent programs Hagit Attiya...

Inherent limitations facilitate design and

verificationof concurrent programs

Hagit Attiya Technion

Concurrent Programs

• Core challenge is synchronization

• Correct synchronization is hard to get right

• Efficient synchronization is even harder

Ad-hoc VS PrincipledManual VS Automatic

Work with Ramalingam and Rinetzky (POPL 2010)

EXAMPLE I:VERIFYING LOCKING PROTOCOLS

The Goal: Sequential Reductions

Verify concurrent data structures• Pre-execution static analysis

E.g., linked list with hand-over-hand locking

• no memory leaks, shape (it’s a list), serializability

Find sequential reductions Consider only sequential executions But conclude that properties hold in all

executions

Back-of-envelop estimate of gain

Static analysis of a linked-list algorithm

[Amit, Rinetzky, Reps, Sagiv, Yahav, CAV 2007]

– Verifies e.g., memory safety, sortedness, pointed-to by a variable, heap sharing One thread (sequential) 10s 3.6MB

Two threads (interleaved) ~4h 886MB

Three threads (interleaved) > 8h ----

Serializability

operation

interleaved execution

complete non-interleaved execution

~~~~~~ ~~~ to the thread locally

[Papadimitriou ‘79]

If M is serializableThen Π ⊨ φ cni-Π ⊨ φ

Serializability assists verification

Concurrent code M Π = all executions of Mφ = a property local to the threadscni-Π: complete non-interleaved

executions of M (small subset of Π)

Easily derived from [Papadimitriou ‘79]

How do we know that M is serializable, w/o considering all executions? E.g., from only complete non interleaved executionsIf M is serializableThen Π ⊨ φ cni-Π ⊨ φ

Special (and common) case: Disciplined programming with

locksGuard access to data with locks

– Lock() acquire the lock– Unlock() release the lock

Only one process holds the lock at each time

Follow a locking protocol that guarantees conflict serializability

E.g., two-phase locking (2PL) or tree locking (TL)

Two-phase locking

[Papadimitriou `79]

• Locks acquire (grow) phase followed by locks release (shrink) phaseNo lock is acquired after some lock is

released

t1

H

t1t1

t2

t1

Tree (hand-over-hand) locking

[Kedem & Sliberschatz ‘76] [Smadi ‘76] [Bayer & Scholnick ‘77]

• Except for the first lock, acquire a lock only when holding the lock on its parent

• No lock is acquired after being released

t1

H

t1t1

t2

Tree (hand-over-hand) locking

[Kedem & Sliberschatz ‘76] [Smadi ‘76] [Bayer & Scholnick ‘77]

• Except for the first lock, acquire a lock only when holding the lock on its parent

• No lock is acquired after being released

t1

t2

t2

H

t1

void p() { acquire(B)B = 0release(B)int b = Bif (b)

acquire(A)}

void q() {acquire(B)B = 1release(B)

}

Yes!– for databases– concurrency control monitor

ensures that M follows the locking policy at run-time M is serializable

No!– for static analysis– no central monitor

Not two-phase lockedBut only in interleaved executions

Our Goal

Statically verify that M follows a locking policy

Applies to local conflict-serializable locking protocols – Depending only on thread’s local variables &

global variables locked by it

E.g., two phase locking, tree locking, (dynamic) DAG locking…

But not protocols that rely on a concurrency control monitor!

Our contribution: Easy step

ni-Π: complete non-interleaved executions of M

For any local conf serializable locking policy LPΠ ⊨ LP ni-Π ⊨ LP

non-interleaved execution

For any thread-local property φΠ ⊨ φ ni-Π ⊨ φ

Two phase locking

Tree lockingDynamic tree

lockingDynamic DAG

locking

Reduction to non-interleaved executions: Proof idea

σ is the shortest execution that does not follow LP

σ’ follows LP, guarantees conflict-serializability

non interleaved execution “equivalent” to σ’

σ (t,e)

σ’

Reduction to non-interleaved executions: Proof idea

σ is the shortest execution that does not follow LP

σ’ follows LP, guarantees conflict-serializability

non interleaved execution “equivalent” to σ’

σ (t,e)

σ’

σ’ni

Reduction to non-interleaved executions: Proof idea

σ is the shortest execution that does not follow LP

σ’ follows LP, guarantees conflict-serializability

non interleaved execution “similar” to σ’

non interleaved execution “similar” to σ’ where LP is violated

σ (t,e)

σ’σni

(t,e)

Further reduction

acni-Π: almost-complete non-interleaved executions of M

For any LCS locking policy LPΠ ⊨ LP acni-Π ⊨ LP

almost complete non-interleaved execution

Reduction to non-interleaved executions: A complication

Need to argue about termination

int X=0, Y=0

void p() {acquire(Y)y = Yrelease(Y); if (y ≠ 0)

acquire(X)X = 3release(X)

}

void q() {if (random(5) == 3){

acquire(Y)Y = 1release(Y)while (true) nop

}}

Y is set to 1 & the method

enters an infinite

loop

Observe Y == 1 & violates

2PL

Reduction to non-interleaved executions: Termination

Can use sequential reduction to verify termination

For any “terminating” local conflict serializable locking policy LPΠ ⊨ LP acni-Π ⊨ LP

Initial analysis results

Shape analysis of hand-over-hand lists

* Does not verify sortedness of list and fails to verify linearizability in some cases

Shape analysis of hand-over-hand trees (for the first time)

Our method 3.5s 4.0MB

TVLA prior 596.1s 90.3MB

Separation logic*

0.4s 0.2MB

Our method 124.6s 90.6MB

What’s next?

• Extend to shared (read) locks• Extend to software transactional

memory– aborted transactions– non-locking non-conflict based

serializability (e.g., using timestamps)

• Combine with other reductions [Guerraoui, Henzinger, Jobstmann, Singh]

EXAMPLE II:REQUIRED MEMORY ORDERINGS

Work with Guerraoui, Hendler, Kuznetsov, Michael and Vechev (POPL 2011)

Relaxed memory models

Out of order execution of memory accesses, to compensate for slow writes

Optimize to issue reads before following writes, if they access different locations

Reordering may lead to inconsistency

Read-after-write (RAW) Reordering

Process P:

Write(X,1)

Read(Y)

Process Q:

Write(Y,1)

Read(X)

P

QW(Y,1)

R(Y)W(X,1)

R(X)

W(X,1)

Avoiding out-of-order:Read-after-write (RAW) Fence

Process P:

Write(X,1)FENCERead(Y)

Process Q:

Write(Y,1)FENCERead(X)

P

QW(Y,1)

R(Y)W(X,1)

R(X)

Avoiding out-of-order:Atomic Operations

Atomic operations: atomic-write-after-read (AWAR)

E.g., CAS, TAS, Fetch&Add,…

RAW fences / AWAR are ~60 slower than (remote) memory accesses

atomic{read(Y) …write(X,1)

}

Our result

34

Any concurrent program in a certain class must use RAW/AWARs

Which programs?

• Concurrent data types:– queues, counters, hash tables, trees,…– Non-commutative operations– Linearizable solo-terminating

implementations

• Mutual exclusion

Non-commutative operations

Operation A is non-commutative if there is operation B where:

A influences Band

B influences A

Example: Queue

enq(v) add v to the end of the queuedeq() dequeues item at the head of the queue

Q.deq():1;Q.deq():2 Q.deq():2;Q.deq():1deq() influence each other

Q.enq(3):ok;Q.deq():1 Q.deq():1;Q.enq(3):okenq() is not non-commutative

1 2Q

1 2Q 3

1 2Q 3

38

Proof Intuition: Writing

If an operation does not write, it does not influence anyoneIt would be commutative

no shared write

1

deq do not influence each other

1deq deq

39

Proof Intuition: Read

If an operation does not read, it is not influenced by anyoneIt would be commutative

1

deq do not influence each other

1deq deq

no shared read

40

Proof Intuition: RAWdeq

1deq

1

W

no RAW

deq 1 1deq

Linearization

Mutual exclusion

(Mutex) Two processes do not hold lock at the same time

(Deadlock-freedom) If a process calls Lock() then some process acquires the lock

Two Lock() operations influence each other!

Every successful lock acquire incurs a RAW/AWAR fence

42

Who should care?

• Concurrent programmers: when is it futile to avoid expensive synchronization

• Hardware designers: motivation to lower cost of specific synchronization constructs

• API designers: choice of API affects synchronization

• Verification engineers: declare incorrect when synchronization is missing

“…although I hope that these shortcomings will be addressed, I hasten to add that they are insignificant compared to the huge step forward that this paper represents….”

-- Linux Weekly News, Jan 26, 2011

What else?

• Weaker operations? E.g., idempotent Work Stealing

• Tight lower bounds? • Other patterns

– Read-after-read, write-after-write, barriers

And beyond…

• The cost of verifying adherence to a locking policy

• (Semi-) Automatic insertion of lock acquire / release commands or fences

Thank you!