Design and Optimize your code for high-performance with Intel® Advisor and Intel® VTune™ Profiler

Editor's Notes

• #4: What’s New in 2019 Enhanced Roofline Analysis Hierarchical – visualize the call chain Use Customizable Roofline Analysis - tailor roofs for the number of threads Share Roofline chart with others, save in HTML format Technical Preview: Try Roofline analysis with integers - great for machine learning MacOS* User Interface - Analyze data collected from Linux* or Windows* targets Flow Graph Analyzer - Visualize Parallelism - Interactively build, validate, and analyze algorithms Use new rapid visual prototyping and analysis tool Interactively build, validate, and visualize algorithms Visually generate code stubs Generate parallel C++ programs Click and zoom through your algorithm’s nodes and edges to understand parallel data and program flow Analyze load balancing, concurrency, and other parallel attributes to fine tune your program Use Intel® TBB or OpenMP* 5 (draft) OMPT APIs --------------------------------- Roofline analysis helps you optimize effectively Find high impact, but under optimized loops Does it need cache or vectorization optimization? Is a more numerically intensive algorithm a better choice?  Faster data collection Filter by module - Calculate only what is needed. Track refinement analysis – Stop when every site has executed Make better decisions with more data, more recommendations Intel MKL friendly – Is the code optimized? Is the best variant used? Function call counts in addition to trip counts Top 5 recommendations added to summary Dynamic instruction mix – Expert feature shows exact count of each instruction Easier MPI launching -- MPI support in the command line dialog Flow Graph Analyzer: Design, Validate and Model for Heterogeneous Systems FGA provides a rapid visual prototyping environment for Threading Building Blocks flow graph API, which has built-in support for designing, validating, and modeling the design before generating TBB source code. Using this tool, you can build algorithms for heterogeneous systems. FGA also enables you to collect traces from an TBB flow graph application and analyze the application for performance issues.
• #5: Offload Advisor Identify which kernels to offload Predict kernel performance on current or future GPUs Identify bottlenecks and potential issues (for example data transfer to GPU) The output generated from offload advisor is self contained on an HTML page. Everything is neatly integrated, including code snippets of identified kernels.
• #6: Some key observations The workload was accelerated 4.4x You can see in program metrics that the original workload ran in 25.07s and the accelerated workload ran in 5.85s
• #8: Your performance will ultimately have an upper bound based on your hardware’s limitations. There are several limitations that Offload Advisor can indicate but they generally come down to compute, memory and data transfer. Knowing what your application is bounded by is critical to developing an optimization strategy
• #9: Gen9 – No efficient to offload Gen11 – 1 offload 98% of code accelerated Accelerated 1.6x … 98% bound by compute (not on slide)
• #11: As your port your application to a discrete GPU, it is important to consider how much of your data will be transferred from your CPU to your GPU and also back to your CPU. This data transfer cost can often dictate whether GPU offload is worthwhile for your application. Offload Advisor gives the data transferred and uses this in addition to other metrics in determining whether you should offload based upon your GPUs characteristics.
• #12: Backup Vectorization Advisor allows you to identify high-impact, under-optimized loops, what is blocking vectorization, and where it is safe to force vectorization. Threading Advisor allows you to analyze, design, tune, and check threading design options without disrupting your normal development. Offload Advisor allows you to collect performance predictor data in addition to the profiling capabilities of Intel Advisor. View output files containing metrics and performance data such as total speedup, fraction of code accelerated, number of loops and functions offloaded, and a call tree showing offloadable and accelerated regions. Flow Graph Analyzer (FGA) is a rapid visual prototyping environment. It assists developers with analyzing and designing parallel applications that use the Intel® Threading Building Blocks (Intel® TBB) flow graph interface.
• #13: Offload Advisor is currently run from the command-line
• #17: GPU Roofline Performance Insights Highlights poor performing loops Shows performance ‘headroom’ for each loop Which can be improved Which are worth improving Shows likely causes of bottlenecks Memory bound vs. compute bound Suggests next optimization steps As an example you can see from the roofline chart, our L3 dot is very close to the L3 maximum bandwidth, to get more FLOPS we need to optimize our caches further. A cache blocking optimization strategy can make better use of memory and should increase our performance. The GTI (traffic between our GPU, GPU uncore (LLC) and main memory)is far from the GTI roofline, transfer costs between out CPU to GPU does not seem to be an issue.
• #23: See more info in product help articles: https://software.intel.com/en-us/vtune-help-gpu-application-analysis
• #34: Your performance will ultimately have an upper bound based on your hardware’s limitations. There are several limitations that Offload Advisor can indicate but they generally come down to compute, memory and data transfer. Knowing what your application is bounded by is critical to developing an optimization strategy