#CUDALAUNCH VISUAL PROFILER SERIAL#
#CUDALAUNCH VISUAL PROFILER MAC#

#CUDALAUNCH VISUAL PROFILER SERIAL#

nsight, nvvp, nvprof)ġ0 Why Profile? GPU Application Code CPU 100 s of cores 10,000 s of threads Great memory bandwidth Best at parallel execution A few cores 10 s of threads Good memory bandwidth Best at serial executionġ1 NVIDIA Profilers NVIDIA Visual Profiler Standalone (nvvp) Integrated into NVIDIA Nsight Eclipse Edition (nsight) NVIDIA Nsight Visual Studio Edition nvprof Command-line Driver-based profiler still available Command-line, controlled by environment variablesġ8 Measuring Time Measure time with horizontal rulers. MPI directly from GPU memoryĨ What Limits Kernel Performance? Memory bandwidth Low number of operations per byte Instruction bandwidth High number of operations per byte Latency Unable to fill memory or arithmetic pipelines How do we determine what is limiting a given kernel? Profiling tools (e.g.

#CUDALAUNCH VISUAL PROFILER MAC#

1 GPU Performance Analysis and Optimisation Thomas Bradley, NVIDIA CorporationĢ Outline What limits performance? Analysing performance: GPU profiling Exposing sufficient parallelism Optimising for Keplerģ Additional Resources More information on topics we cannot cover today Kepler architecture: GTC on Demand: Session S0642 Inside Kepler Kepler whitepapers: More details: GTC on Demand - S0514 GPU Performance Analysis and Optimization - S0419 Optimizing Application Performance with CUDA ProfilingTools - S0420 Nsight IDE for Linux and Mac CUPTI documentation (describes all the profiler counters) - Included in every CUDA toolkit (/cuda/extras/cupti/doc/cupti_users_guide.pdf GPU computing webinars in general: In particular: register spillingĤ Example Workflow Getting Started on GPU Overall goal is application performance, combination of factors inter-node/inter-process communication CPU-GPU communication CPU/GPU performance Start by analysing realistic data at realistic scale Various tools such as Vampir, TAU, Scalasca help identify hotspots Extract data for development Select timesteps for shorter runtime Select subdomain for smaller scaleĥ Example Workflow Optimising Communication (node/process) Computation In this talk we re focussing on the GPU CPU, DMA, GPU overlap Kernel optimisationħ What Limits Communication with the GPU? PCIe bus connects GPU to CPU/network Gen 2 (Fermi): 8 GB/s in each direction Gen 3 (Kepler): 16 GB/s in each direction Tesla GPUs have dual DMA engines Two memcpys (in different streams, different directions) Overlap with kernel and CPU execution GPUDirect RDMA e.g.

Cudalaunch visual profiler