SherLock is a synchronization inference tool described in the paper "SherLock: Unsupervised Synchronization-Operation Inference" in ASPLOS 2021.
We define the synchronization operation as the program points that introducing happens-before relation, for example, the lock read/write, thread start/join and while-loop on flag. The format of synchronization could be very flexible while they are very important to analyze the correctness of program. That’s why we design SherLock to infer them with an unsupervised approach. In general, SherLock classifies the synchronization as two types: release and acquire. The release always happens before the corresponding acquire.
Python3, numpy in Python, flipy in Python and Visual Studio 2019 or newer.
SherLock is a dynamic analyze tool that instruments the program to collect runtime trace and analyzes the trace to infer synchronizations.
instrument-tool is the directory contain the instrumentation tool. It is modified from ‘https://github.com/microsoft/TSVD’. By compiling the project in this directory generates a TorchLite.exe
.\TorchList.exe [path of binaries]; instrumenting the binaries in the path.
log-analysis has two parts: a preprocessing and linear solver. After running the instrumented binaries, a ‘Runtime.log’ file is generated in the running directory. Compiling the project in preprocessing generates Analyzer.exe
.\Analyze.exe [path contains Runtime.log]; preprocessing the log.
Analyze.exe produces the result in the same directory with Runtime.log. Eventually, in the linear-solver directory:
Python log_analyzer.log --batch [path to preprocessed result] -refine ; inferring the synchronizations.
example\Example contains a toy example. TestApp is the application we want to analyze. inst_run.ps1 is the script example to instrument and run the app TestApp. log_analysis.ps1 is the script for both preprocessing and linear-solving. By running inst_run.ps1 then log_analysis.ps1, we can see the inferred synchronizations:
Release: thread.start()
Acquire: update-begin() ; this is the beginning of a thread.
More details will be found here about going through the basic workflow multiple times for better results. You can check our artifact for more examples.