Parallelizing Data Race Detection

Benjamin WesterFacebook

David Devecsery, Peter Chen, Jason Flinn, Satish NarayanasamyUniversity of Michigan

Data Races

Benjamin Wester - ASPLOS '13 2

TIM

E

t1 t2

lock(l)x = 1unlock(l)

x = 3lock(l)x = 2unlock(l)

Race Detection

Benjamin Wester - ASPLOS '13

Normal run time

TIM

E With race detector

3

Instrumentation+

Analysis

Detection is slow!8x–200x•Managed vs. binary•Granularity•Completeness•Debug info gatheredMatches app concurrency

Race Detection

Benjamin Wester - ASPLOS '13

Normal run time

TIM

E With race detector

4

Parallel race detector

Detection is slow!8x–200x•Managed vs. binary•Granularity•Completeness•Debug info gatheredMatches app concurrency

How to use parallel hardware?

• Scale program?– Performance tied to app

• Parallel algorithm?– Lots of fine-grained dependencies

• Uniparallelism– Converts execution into parallel pipeline– Works for instrumentation

Benjamin Wester - ASPLOS '13 5

[Wallace CGO’07,Nightingale ASPLOS’08]

Uniparallelism

Benjamin Wester - ASPLOS '13

TIM

E

6

Uniparallelism

Benjamin Wester - ASPLOS '13

TIM

E

7

Epoch-parallelExecution

Epoch-sequentialExecution

Uniparallelism

• Scales well:Epochs are independent execution units

• More efficient:Synchronization

Benjamin Wester - ASPLOS '13 8

TIM

E

Add Detector…

• Scales well:Epochs are independent execution units

• More efficient:Synchronization& Lock elision

Race detection depends on state!

Benjamin Wester - ASPLOS '13 9

TIM

E Here!Here!

Architecture

Benjamin Wester - ASPLOS '13 10

Epoch-sequentialEpoch-sequential Epoch-parallelEpoch-parallel CommitCommit

Moving Work

Identify meaningful fast subset of analysis state– Low frequency, lightweight, self-contained– Run in Epoch-Sequential phase– Predicted and usable in parallel phase

Symbolic execution– Replace missing state with symbols– Defer some computation to commit phase– Commit: replace symbols with concrete values

Benjamin Wester - ASPLOS '13 11

Architecture

Benjamin Wester - ASPLOS '13 12

Epoch-sequentialEpoch-sequential Epoch-parallelEpoch-parallel CommitCommit

Instrument and predictfast subset of analysis state Perform deferred work

on concrete values

Full instrumentationSymbolic execution

Fast

Slow

State: Vector clocks – e.g. ⟨0, 1, 0⟩ – for each:– Thread– Lock– Variable x 2: last read(s), last write

Example computation:Check(read_x ≤ thread_i ); write_x = thread_i

Transitive reductionUse knowledge of happens-before relationship

Parallel FastTrack

Benjamin Wester - ASPLOS '13 13

Fast

Slow

Parallel Eraser

State:– Set of locks held by thread (lockset)– Lockset for each variable– Position in state machine for each variable

Example computation:If state_x == SHARED then ls_x = ls_x ∩ {L, M}

Lockset factorizationFactor out common behavior

Benjamin Wester - ASPLOS '13 14

Parallel Performance

EfficiencyScalability

2 worker threads as baseline8-CPU PlatformBenchmarks:

SPLASH-2 (water, lu, ocean, fft, radix) Parallel app: pbzip2

Benjamin Wester - ASPLOS '13 15

Efficiency

Benjamin Wester - ASPLOS '13 16

Exactly 2 cores

Efficiency

Benjamin Wester - ASPLOS '13 17

Median 13% faster

Median 13% faster

Median 8% faster

Median 8% faster

Exactly 2 cores

Scaling Parallel FastTrack

Benjamin Wester - ASPLOS '13 18

2 Worker Threads

Scaling Parallel FastTrack

Benjamin Wester - ASPLOS '13 19

Median 4.4x with 8 coresMedian 4.4x with 8 cores

2 Worker Threads

Scaling Parallel Eraser

Benjamin Wester - ASPLOS '13 20

Median 3.3x with 8 coresMedian 3.3x with 8 cores

2 Worker Threads

Conclusion

3-Phase Uniparallel architecture– Parallel FastTrack– Parallel Eraser

Parallel algorithms:– Are more efficient– Scale better

Benjamin Wester - ASPLOS '13 21

Questions?

Benjamin Wester - ASPLOS '13 22