Performance Analysis of Parallel Scientific Applications In Eclipse

Transcription

1 Performance Analysis of Parallel Scientific Applications In Eclipse EclipseCon 2015 Wyatt Spear, University of Oregon

2 Supercomputing Big systems solving big problems Performance gains save time and money Development historically based in the command line Communications infrastructure to transfer data between compute nodes complicates software development Principles of parallel computing becoming principles of computing

3 Parallel Tools Platform (PTP) The Parallel Tools Platform aims to provide a highly integrated environment specifically designed for parallel application development Features include: An integrated development environment (IDE) that supports a wide range of parallel architectures and runtime systems A scalable parallel debugger Parallel programming tools (MPI, OpenMP, UPC, etc.) Support for the integration of parallel tools An environment that simplifies the end-user interaction with parallel systems

4 How Eclipse PTP is Used Editing/Compiling Local Source Code Remote Source Code

5 How Eclipse PTP is Used Launching/Monitoring Source Code Executable

6 How Eclipse PTP is Used Debugging Source Code Executable

7 How Eclipse PTP is Used Performance Tuning Source Code Executable Perf. Data

8 Synchronized Projects Projects types can be: Launch Service Run Compute Debug Service g Debu Build Service Executable Bu il d Index Service File Service Edit Search/Index Navigation Local source code Synchronize Local Source code copy Remote

9 PTP/External Tools Framework formerly Performance Tools Framework Goal: Reduce the eclipse plumbing necessary to integrate tools Provide integration for instrumentation, measurement, and analysis for a variety of performance tools Dynamic Tool Definitions: Workflows & UI Tools and tool workflows are specified in an XML file Tools are selected and configured in the launch configuration window Output is generated, managed and analyzed as specified in the workflow One-click launch functionality Support for development tools such as TAU, PPW and others. Adding new tools is much easier than developing a full Eclipse plug-in

10 SAX and JAXB Tool Definitions Prior implementations of ETFW used a simple SAX based schema to define tool workflows By default workflows now use the more powerful JAXB schema that defines PTP s resource manager Legacy workflows can still be loaded by selecting the SAX parser in PTP options Window->Preferences-> Parallel Tools->External Tools

11 TAU: Tuning and Analysis Utilities TAU is a performance evaluation tool It supports parallel profiling and tracing Profiling shows you how much (total) time was spent in each routine Tracing shows you when the events take place in each process along a timeline TAU uses a package called PDT (Performance Database Toolkit) for automatic instrumentation of the source code Profiling and tracing can measure time as well as hardware performance counters from your CPU (or GPU!) TAU can automatically instrument your source code (routines, loops, I/O, memory, phases, etc.) TAU runs on all HPC platforms and it is free (BSD style license) TAU has instrumentation, measurement and analysis tools paraprof is TAU s 3D profile browser TAU TAU-11

12 TAU Performance System Architecture

13 Direct Observation: Events Event types Interval events (begin/end events) Measures exclusive & inclusive durations between events Metrics monotonically increase Atomic events (trigger with data value) Used to capture performance data state Shows extent of variation of triggered values (min/max/mean) Code events Routines, classes, templates Statement-level blocks, loops 13

14 Inclusive and Exclusive Profiles Performance with respect to code regions Exclusive measurements for region only Inclusive measurements includes child regions int foo() { int a; a =a + 1; bar(); exclusive duration inclusive duration a =a + 1; return a; } 14

15 Hardware Counters Hardware performance counters available on most modern microprocessors can provide insight into: 1. Whole program timing 2. Cache behaviors 3. Branch behaviors 4. Memory and resource access patterns 5. Pipeline stalls 6. Floating point efficiency 7. Instructions per cycle Hardware counter information can be obtained with: 1. Subroutine or basic block resolution 2. Process or thread attribution

16 Profiling On The Command Line % export TAU_MAKEFILE=<taudir>/<arch>/lib/Makefile.tau-papi-mpipdt % export TAU_OPTIONS= -opttauselectfile=select.tau optverbose % cat select.tau BEGIN_INSTRUMENT_SECTION loops routine= # END_INSTRUMENT_SECTION % make F90=tau_f90.sh (Or edit Makefile and change F90=tau_f90.sh) % export TAU_METRICS=TIME:PAPI_FP_INS:PAPI_L1_DCM % mpirun np 8./a.out % paraprof - pack app.ppk Move the app.ppk file to your desktop. % paraprof app.ppk 16

17 PTP TAU plug-ins TAU (Tuning and Analysis Utilities) First implementation of External Tools Framework (ETFw) Eclipse plug-ins wrap TAU functions, make them available from Eclipse Full GUI support for the TAU command line interface Performance analysis integrated with development environment

18 TAU Integration with PTP TAU: Tuning and Analysis Utilities Performance data collection and analysis for HPC codes Numerous features Command line interface The TAU Workflow: Instrumentation Execution Analysis

19 Selective Instrumentation By default tau provides timing data for each subroutine of your application Selective instrumentation allows you to include/exclude code from analysis and control additional analysis features Include/exclude source files or routines Add timers and phases around routines or arbitrary code Instrument loops Note that some instrumentation features require the PDT Right click on a source file to see the Selective Instrumention context menu Results in creation of tau.selective

20 Begin Profile Configuration The ETFw uses the same run configurations and resource managers as debugging/launching Click on the Run menu or the right side of the Profile button From the dropdown menu select Profile configurations

21 Select Configuration Select an existing launch configuration or create a new one The Resource and Application configuration tabs require little or no modification from standard PTP launch Allows selection/creation of remote connection PTP provides a UI for the remote resource manager, e.g. Torque Includes options for configuring remote environment including modules Performance Analysis tab is present in the Profile Configurations dialog

22 Select Tool/Workflow Select the Performance Analysis tab and choose the TAU tool set in the Select Tool dropdown box Other tools may be available, either installed as plug-ins or loaded from workflow definition XML files Configuration sub-panes appear depending on the selected tool

23 Select TAU Configuration Choose the TAU Makefile tab: All TAU configurations in remote installation are available Check MPI and PDT checkboxes to filter listed makefiles Make your selection in the Select Makefile: dropdown box TAU provides individual stub makefiles for each configuration, tailored to the programming paradigm and data being collected.

24 Choose PAPI Hardware Counters When a PAPI-enabled TAU configuration is selected the PAPI Counter tool becomes available Select the Select PAPI Counters button to open the tool Open the PRESET subtree Select PAPI_L1_DCM (Data cache misses) Scroll down to select PAPI_FP_INS (Floating point instructions) Invalid selections are automatically excluded Select OK

25 Compiler Options TAU Compiler Options Set arguments to TAU compiler scripts Control instrumentation and compilation behavior Verbose shows activity of compiler wrapper KeepFiles retains instrumented source PreProcess handles C type ifdefs in fortran Specify use of selective instrumentation

26 Runtime Options TAU Runtime options Set environment variables used by TAU Control data collection behavior Verbose provides debugging info Callpath shows call stack placement of events Throttling reduces overhead Tracing generates execution timelines Hover help

27 Working with Profiles Profiles are uploaded to selected database A text summary may be printed to the console Profiles may be uploaded to the TAU Portal for viewing online tau.nic.uoregon.edu Profiles may be copied to your workspace and loaded in ParaProf from the command line.

28 Launch TAU Analysis Once your TAU launch is configured select Profile The project rebuilds on the remote system with TAU compiler commands The project will execute normally but TAU profiles will be generated TAU profiles will be processed as specified in the launch configuration. If you have a local profile database the run will show up in the Performance Data Management view Double click the new entry to view in ParaProf Right click on a function bar and select Show Source Code for source callback to Eclipse

29 Paraprof Use ParaProf for profile visualization to identify performance hotspots Inefficient sequential computation Communication overhead IO/Memory bottlenecks Load imbalance Suboptimal cache performance Compare multiple trials in PerfExplorer to identify performance regressions and scaling issues

30 Callpath Profile 30

31 Trace of HMPP SGEMM (CAPS Entreprise) Host Process Transfer Kernel Compute Kernel Host Process Transfer Kernel Compute Kernel

32 Communication Matrix Display Goal: What is the volume of inter-process communication? Along which calling path?

33 Performance Regression Testing

34 Evaluate Scalability

35 Roofline Analysis A Roofline chart represents peak hardware floating point performance and peak memory bandwidth The Roofline can be derived from system modeling via benchmarks or taken from system specifications Computational intensity of application kernels are plotted on the roofline to illustrate the difference between observed and achievable performance Roofline charts can be applied to system power models as well

36 Roofline Visualization Toolkit Eclipse integration for development platform support Roofline charts implemented in JavaFX. Allows for portable, standalone viewer Roofline data is stored in JSON files including recorded performance metrics and metadata for ease of searching and comparison between trials, systems and benchmarks. Preliminary remote, file based database for roofline data storage

37 Roofline UI Load roofline data from remote site or local disk Quick selection from multiple datasets

38 Eclipse Integration Select application events in Eclipse source outline Display values from TAUdb database on Roofline chart

39 Online Information Information about PTP PTP online help Main web site for downloads, documentation, etc. Wiki for designs, planning, meetings, etc. Information about Photran Main web site for downloads, documentation, etc. Information About TAU

40 PTP Mailing Lists User Mailing Lists PTP Photran Major announcements (new releases, etc.) - low volume Developer Mailing Lists Developer discussions - higher volume