1、Correcting Threading Errors with Intel Parallel Inspector2Intel Parallel InspectorObjectivesAfter successful completion of this module you will be able toUse Parallel Inspector to detect and identify a variety of threading correctness issues in threaded applications Determine if library functions ar

2、e thread-safe3Intel Parallel InspectorAgendaWhat is Intel Parallel Inspector?Detecting race conditions Detecting potential for deadlockChecking library thread-safety4Intel Parallel InspectorMotivationDeveloping threaded applications can be a complex taskNew class of problems are caused by the intera

3、ction between concurrent threadsData races or storage conflictsMore than one thread accesses memory without synchronizationDeadlocks Thread waits for an event that will never happen5Intel Parallel InspectorIntel Parallel InspectorDebugging tool for threaded softwarePlug-in to Microsoft* Visual Studi

4、o*Finds threading bugs in OpenMP*, Intel Threading Building Blocks, and Win32* threaded softwareLocates bugs quickly that can take days to find using traditional methods and toolsIsolates problems, not the symptomsBug does not have to occur to find it!Intel Parallel Inspector FeaturesIntegrated into

5、 Microsoft Visual Studio .NET* IDE2005 & 2008 EditionsSupports different compilersMicrosoft* Visual* C+ .NET*Intel Parallel ComposerView (drill-down to) source code for DiagnosticsOne-click help for diagnosticsPossible causes and solution suggestions6Intel Parallel Inspector7Intel Parallel Inspector

6、Parallel Inspector: Analysis Dynamic as software runsData (workload) -driven executionIncludes monitoring of:Thread and Sync APIs usedThread execution order Scheduler impacts resultsMemory accesses between threadsCode path must be executed to be analyzed8Intel Parallel InspectorParallel Inspector: B

7、efore You StartInstrumentation: backgroundAdds calls to library to record informationThread and Sync APIsMemory accessesIncreases execution time and sizeUse small data sets (workloads)Execution time and space is expandedMultiple runs over different paths yield best resultsWorkload selection is impor

8、tant! 9Intel Parallel InspectorWorkload GuidelinesExecute problem code once per thread to be identifiedUse smallest possible working data set Minimize data set sizeSmaller image sizesMinimize loop iterations or time stepsSimulate minutes rather than daysMinimize update ratesLower frames per secondFi

9、nds threading errors faster!10Intel Parallel InspectorBuilding for Parallel InspectorCompileUse dynamically linked thread-safe runtime libraries (/MDd)Generate symbolic information (/ZI)Disable optimization (/Od)Link Preserve symbolic information (/DEBUG)Specify relocatable code sections (/FIXED:NO)

10、11Intel Parallel InspectorBinary InstrumentationBuild with supported compilerRunning the applicationMust be run from within Parallel InspectorApplication is instrumented when executedExternal DLLs are instrumented as usedStarting Parallel InspectorBuild the Debug version of the application with appr

11、opriate flags set12Intel Parallel InspectorStarting Parallel InspectorSelect Parallel Inspector from the Tools menu13Intel Parallel InspectorYou can choose to look forMemory ErrorsThreading ErrorsStarting Parallel InspectorThe Configure Analysis window pops up14Intel Parallel InspectorSelect the lev

12、el of analysis to be carried out by Parallel InspectorThe deeper the analysis, the more thorough the results and the longer the execution timeClick Run AnalysisStarting Parallel InspectorThe initial (raw) results come up after analysis15Intel Parallel InspectorClick the Interpret Results button to f

13、ilter the raw data into more human consumable formatsStarting Parallel InspectorThe analysis results are gathered together in related categories16Intel Parallel InspectorDouble-click a line from the Problem Sets pane to see the source code that generated the diagnosticStarting Parallel InspectorThe

14、source lines involved in a data race can be shown17Intel Parallel Inspector18Intel Parallel InspectorActivity 1a - Potential Energy Build and run serial version Build threaded versionRun application in Parallel Inspector to identify threading problems19Intel Parallel InspectorRace ConditionsExecutio

15、n order is assumed but cannot be guaranteedConcurrent access of same variable by multiple threadsMost common error in multithreaded programsMay not be apparent at all times20Intel Parallel InspectorSolving Race ConditionsSolution: Scope variables to be local to threadsWhen to useValue computed is no

16、t used outside parallel regionTemporary or “work” variablesHow to implementOpenMP scoping clauses (private, shared)Declare variables within threaded functionsAllocate variables on thread stackTLS (Thread Local Storage) API21Intel Parallel InspectorSolving Race ConditionsSolution: Control shared acce

17、ss with critical regionsWhen to useValue computed is used outside parallel regionShared value is required by each threadHow to implementMutual exclusion and synchronizationLock, semaphore, event, critical section, atomicRule of thumb: Use one lock per data element22Intel Parallel InspectorActivity 1

18、b - Potential Energy Fix errors found by Parallel Inspector23Intel Parallel InspectorDeadlockCaused by thread waiting on some event that will never happenMost common cause is locking hierarchiesAlways lock and un-lock in the same orderAvoid hierarchies if possibleDWORD WINAPI threadA(LPVOID arg) Ent

19、erCriticalSection(&L1); EnterCriticalSection(&L2); processA(data1, data2); LeaveCriticalSection(&L2); LeaveCriticalSection(&L1); return(0);DWORD WINAPI threadB(LPVOID arg) EnterCriticalSection(&L2); EnterCriticalSection(&L1); processB(data2, data1) ; LeaveCriticalSection(&L1);LeaveCriticalSection(&L

20、2); return(0);ThreadA: L1, then L2ThreadB: L2, then L124Intel Parallel InspectorDeadlockAdd lock per elementLock only elements, not whole array of elementsvoid swap (shape_t A, shape_t B) lock(a.mutex); lock(b.mutex);/ Swap data between A & B unlock(b.mutex); unlock(a.mutex);typedef struct / some da

21、ta things SomeLockType mutex; shape_t;shape_t Q1024;swap(Q986, Q34);Thread 4swap(Q34, Q986);Thread 1Grabs mutex 34Grabs mutex 98625Intel Parallel InspectorWindows* Critical SectionLightweight, intra-process only mutexMost useful and most usedNew type CRITICAL_SECTION cs;Create and destroy operations

22、InitializeCriticalSection(&cs)DeleteCriticalSection(&cs);26Intel Parallel InspectorWindows* Critical SectionCRITICAL_SECTION cs ;Attempt to enter protected code EnterCriticalSection(&cs)Blocks if another thread is in critical sectionReturns when no thread is in critical sectionUpon exit of critical

23、sectionLeaveCriticalSection(&cs)Must be from obtaining thread27Intel Parallel InspectorExample: Critical Section#define NUMTHREADS 4CRITICAL_SECTION g_cs; / why does this have to be global?int g_sum = 0;DWORD WINAPI threadFunc(LPVOID arg ) int mySum = putation(); EnterCriticalSection(&g_cs); g_sum +

24、= mySum; / threads access one at a time LeaveCriticalSection(&g_cs); return 0;main() HANDLE hThreadNUMTHREADS; InitializeCriticalSection(&g_cs); for (int i = 0; i NUMTHREADS; i+) hThreadi = CreateThread(NULL,0,threadFunc,NULL,0,NULL); WaitForMultipleObjects(NUMTHREADS, hThread, TRUE, INFINITE); Dele

25、teCriticalSection(&g_cs);28Intel Parallel InspectorActivity 2 - DeadlockUse Intel Parallel Inspector to find and correct the potential deadlock problem.29Intel Parallel InspectorThread Safe RoutinesAll routines called concurrently from multiple threads must be thread safeHow to test for thread safet

26、y?Use OpenMP and Parallel Inspector for analysisUse sections to create concurrent execution30Intel Parallel InspectorThread Safety ExampleCheck for safety issues betweenMultiple instances of routine1()Instances of routine1() and routine2()Set up sections to test all permutationsStill need to provide

27、 data sets that exercise relevant portions of code#pragma omp parallel sections#pragma omp section routine1(&data1);#pragma omp section routine1(&data2);#pragma omp section routine2(&data3);31Intel Parallel InspectorIt is better to make a routine reentrant than to add synchronizationAvoids potential overhead Two Ways to Ensure