Light64: Lightweight Hardware Support for Data Race Detection during Systematic Testing of Parallel Programs A. Nistor, D. Marinov and J. Torellas to appear.

Light64: Lightweight Hardware Support for Data Race Detection during Systematic Testing of Parallel Programs

A. Nistor, D. Marinov and J. Torellas

to appear MICRO’09LBA reading group – 09/29/09

(by Evangelos)

Introduction – Context Debugging of parallel applications

Even for 1 input too many interleavings Systematic Testing

Execute many times - explore all interleavings

Assumptions: Input provided Thread Interleaving only cause of non-determinism

Goal: Hardware support for data race detection under Systematic Testing

Background of Systematic Testing

• Serializing of threads (multiplexing)

• New scheduler implementation

• Happens-before definition

• Segment-based interleaving

Background of Systematic Testing

State: represented by a Serial Log; ordered list of segments

Background of Systematic Testing

Light64 – The Idea

“Two different thread interleavings that have the same happens-before graph but a flipped data race, will very likely have at least a small deviation in the execution history”

Corner cases?

No false positives; few false negatives Systematic tester environment highly

deterministic Extremely improbable for two different

streams of values to generate the same hash

Cannot identify benign races; races on data that will never be consumed

By construction…

Design

Small hardware modifications CRC logic at the head of ROB ISA extensions; start/stop – save/load hash

history Two modes of execution

Passive Mode Active Mode Tradeoff between accuracy and

performance

Passive Mode

During step 4 Augment each state with the Execution

History Hash. Check if executions with same happens-before have the same hash value (e.g., S2 & S11)

No guarantees on coverage Dependable on systematic tester’s exploration

strategy and pruning heuristics No practical overhead

Active Mode

During step 2; While re-executing to reach the selected state ‘S’,

flip as many segments as possible. Compare Execution History Hash against original execution

Heuristic 1 – efficient segment reordering Smallest-ID Thread first during first run Biggest-ID Thread first during re-execution

Heuristic 2 – additional re-executions to increase coverage

ActiveFIN – re-execute all final states ActiveFULL – re-execute all states

Experimental Setup

Used Pin to model a system running a systematic tester

Instruction count as a performance metric

SPLASH-2 benchmarks (modified & unmodified)

6 versions of a system: Plain, Plain+RD, ActiveNO, ActiveFIN,

ActiveFULL, Passive

State Space Characterization

Race Detection Capability

Runtime Overhead

Runtime Overhead – Software-based

Conclusions

Lightweight support for data race detection in a Systematic Tester world

Relatively low overhead for S.T. Not a conventional MICRO paper