深入理解 Nsight System 与 Nsight Compute 性能分析优化工具.pdf

1、NVIDIADEEPDIVEINTONSIGHTSYSTEMS8 NSIGHTCOMPUTEBing Liiu，202012#page#Overview of ProfilersNsight SystemsNsight ComputeAGENDACase StudiesSummary#page#page#NSIGHT PRODUCT FAMILYStart hereNsight SystemsComprowokoadelperomanceRecheckoverallRecheckoverallCouAODive intographicsCUDAkonolsframosNsight Comput

2、eNsight GraphicsDetaled CuDAkenel perforFinished ifOpuimize：performanceToyaccasssatisfactory#page#page#NSIGHT SYSTEMSOverviewSystem-wide application algorithm tuningFocus on the applications algorithm- a unique perspectiveLocate optimization opportunitiess See gaps of unused CPU and GPU timeBalance

3、your workload across multiple CPUs and GPUsCPU algorithms，utilization，and thread stateGPU streams, kernels， memory transfers， etcSupport for Linux 8 Windows，X86-64 8 Tegra. Host only for Mac#page#NSIGHT SYSTEMSKey FeaturesComputeCUDA API. Kernel launch and execution correlationLibraries: cuBLAS， CUD

4、NN，TensorRTOpenACCGraphicsVulkan，OpenGL，DX11，DX12，DXR，V-syncOS Thread state and CPU utilization， pthread， file l/O，etc.User annotations API (NVTX）#page#Thread/coremigrationProcesses0andThread statethreads来科印CUDA andOpenGL API traceCuDNN andCUBLAS trace110201110101Kernel and memorytransfer activities

5、I日Multi-GPU#page#CPU THREADSThread ActivitiesGet an overview of each threads activitiesWhich core the thread is running and the utilizationCPU state and transitionOS runtime libraries usage: pthread， file l/O，etc.APl usage: CUDA，CuDNN， CuBLAS，TensorRT，口181pythonOSruntime librariesCUDAAPIcuEventSynch

6、ronizcuEventCuEverCUEVCUDNN11日CuBLAS#page#page#OS RUNTIME LIBRARIESldentify time periods where threads are blocked and the reasonLocate potentially redundant synchronizationsTPL122405pythonsemwait.jpthread.pthreadsemwaitfgetcOS runtime librariespthread cond waitBeqins:9.23185sCUDA APIlhMhEnds:9.2336