1. Yuan Yu(MSR) Tom Rodeheffer(MSR) Wei Chen(UC Berkeley) SOSP 2005
RaceTrack: Efficient Detection of Data Race Conditions via Adaptive Tracking
Yuan Yu(MSR) Tom Rodeheffer(MSR)
Wei Chen(UC Berkeley)
SOSP 2005

2. Dynamic Data Race Detection
Lockset: verify locking discipline for shared memory
Detect race regardless of thread scheduling
False positives because other synchronization primitives (fork/join, signal/wait) not supported
Happens-before: track partial order of program events
Supports general synchronization primitives
Higher overhead compared to lockset
False negatives due to sensitivity to thread scheduling
RaceTrack = Lockset + Happens-before

3. False positive using Lockset
Tracking accesses to X
Inst
State
Lockset 1
Virgin
{ } 3
Exclusive:t 6
Shared Modified
{a} 9
Report race

4. RaceTrack Notations Lt Cx Bu Sx ti Notation Meaning
Lockset of thread t Cx
Lockset of memory x Bu
Vector clock of thread u Sx
Threadset of memory x ti
Thread t at clock time i

5. RaceTrack Algorithm Lt Cx Bt Sx t1 Notation Meaning
Lockset of thread t Cx
Lockset of memory x Bt
Vector clock of thread t Sx
Threadset of memory x t1
Thread t at clock time 1

6. Avoiding Lockset's false positive (1)
Inst Cx Sx Lt Bt Lu Bu
All
{ }
{t1} - 1
{t2}
{ t1,u1 } 2
{a} 3 4 5 6
{t2,u1} 7 8

7. Avoiding Lockset's false positive (2)
Inst Cx Sx Lt Bt Lv Bv 8
{a}
{t2,u1}
{ } - 9
{t2} 10
{t3,u1}
{t2,v1} 11 12
{t3} 13 14 15
{t3,v1} 16

8. RaceTrack's state machine

9. Additional ideas from paper
Accuracy vs performance & scalability tradeoff
Object granularity tracking
Track subset of (array) objects
Prune vector clock
Annotations to eliminate false positives
Warnings report analysis
Ranking and classification
Multiple stack traces

10. Microsoft CLR Implementation
.NET framework

11. RaceTrack Object Layout

12. RaceTrack Encodings

13. Evaluation CLR Regression tests Performance 2122 tests (0.5 MLOC)
48 warnings
Performance
5 real world programs

14. Performance 3X slowdown on memory intensive programs 1.2X memory usage
< 2X on other programs
1.2X memory usage

15. Conclusions Eliminate Lockset false positives using happens-before
Refine state machine based on common coding style
Trade off accuracy for performance/scalability
Improved warnings reports