Allen D. Malony Department of Computer and Information Science Performance Research Laboratory University of Oregon Performance Technology for Productive, High-End Parallel Computing
LLNL, Oct Outline of Talk Research motivation Scalability, productivity, and performance technology Application-specific and autonomic performance tools TAU parallel performance system developments Application performance case studies New project directions Performance data mining and knowledge discovery Concluding discussion
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Research Motivation Tools for performance problem solving Empirical-based performance optimization process Performance technology concerns characterization Performance Tuning Performance Diagnosis Performance Experimentation Performance Observation hypotheses properties Instrumentation Measurement Analysis Visualization Performance Technology Experiment management Performance database Performance Technology
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Large Scale Performance Problem Solving How does our view of this process change when we consider very large-scale parallel systems? What are the significant issues that will affect the technology used to support the process? Parallel performance observation is clearly needed In general, there is the concern for intrusion Seen as a tradeoff with performance diagnosis accuracy Scaling complicates observation and analysis Nature of application development may change Paradigm shift in performance process and technology? What will enhance productive application development?
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Scaling and Performance Observation Consider “traditional” measurement methods Profiling: summary statistics calculated during execution Tracing: time-stamped sequence of execution events More parallelism more performance data overall Performance specific to each thread of execution Possible increase in number interactions between threads Harder to manage the data (memory, transfer, storage) How does per thread profile size grow? Instrumentation more difficult with greater parallelism? More parallelism / performance data harder analysis More time consuming to analyze and difficult to visualize
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Concern for Performance Measurement Intrusion Performance measurement can affect the execution Perturbation of “actual” performance behavior Minor intrusion can lead to major execution effects Problems exist even with small degree of parallelism Intrusion is accepted consequence of standard practice Consider intrusion (perturbation) of trace buffer overflow Scale exacerbates the problem … or does it? Traditional measurement techniques tend to be localized Suggests scale may not compound local intrusion globally Measuring parallel interactions likely will be affected Use accepted measurement techniques intelligently
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Role of Intelligence and Specificity How to make the process more effective (productive)? Scale forces performance observation to be intelligent Standard approaches deliver a lot of data with little value What are the important performance events and data? Tied to application structure and computational mode Tools have poor support for application-specific aspects Process and tools can be more application-aware Will allow scalability issues to be addressed in context More control and precision of performance observation More guided performance experimentation / exploration Better integration with application development
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Role of Automation and Knowledge Discovery Even with intelligent and application-specific tools, the decisions of what to analyze may become intractable Scale forces the process to become more automated Performance extrapolation must be part of the process Build autonomic capabilities into the tools Support broader experimentation methods and refinement Access and correlate data from several sources Automate performance data analysis / mining / learning Include predictive features and experiment refinement Knowledge-driven adaptation and optimization guidance Address scale issues through increased expertise
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU Parallel Performance System Goals Multi-level performance instrumentation Multi-language automatic source instrumentation Flexible and configurable performance measurement Widely-ported parallel performance profiling system Computer system architectures and operating systems Different programming languages and compilers Support for multiple parallel programming paradigms Multi-threading, message passing, mixed-mode, hybrid Support for performance mapping Support for object-oriented and generic programming Integration in complex software, systems, applications
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU Parallel Performance System Architecture
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU Parallel Performance System Architecture
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Advances in TAU Instrumentation Source instrumentation Program Database Toolkit (PDT) automated Fortran 90/95 support (Flint parser, very robust) statement level support in C/C++ (Fortran soon) TAU_COMPILER to automate instrumentation process Automatic proxy generation for component applications automatic CCA component instrumentation Python instrumentation and automatic instrumentation Continued integration with dynamic instrumentation Update of OpenMP instrumentation (POMP2) Selective instrumentation and overhead reduction Improvements in performance mapping instrumentation
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Advances in TAU Measurement Profiling Memory profiling global heap memory tracking (several options) Callpath profiling user-controllable calling depth Improved support for multiple counter profiling Online profile access and sampling Tracing Generation of VTF3 traces files (fully portable) Inclusion of hardware performance counts in trace files Hierarchical trace merging Online performance overhead compensation Component software proxy generation and monitoring
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Advances in TAU Performance Analysis Enhanced parallel profile analysis (ParaProf) Callpath analysis integration in ParaProf Embedded Lisp interpreter Performance Data Management Framework (PerfDMF) First release of prototype In use by several groups S. Moore (UTK), P. Teller (UTEP), P. Hovland (ANL), … Integration with Vampir Next Generation (VNG) Online trace analysis Performance visualization (ParaVis) prototype Component performance modeling and QoS
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU Performance System Status Computing platforms (selected) IBM SP / pSeries, SGI Origin 2K/3K, Cray T3E / SV-1 / X1, HP (Compaq) SC (Tru64), Sun, Hitachi SR8000, NEC SX-5/6, Linux clusters (IA-32/64, Alpha, PPC, PA- RISC, Power, Opteron), Apple (G4/5, OS X), Windows Programming languages C, C++, Fortran 77/90/95, HPF, Java, OpenMP, Python Thread libraries pthreads, SGI sproc, Java,Windows, OpenMP Compilers (selected) Intel KAI (KCC, KAP/Pro), PGI, GNU, Fujitsu, Sun, Microsoft, SGI, Cray, IBM (xlc, xlf), HP, NEC, Absoft
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Component-Based Scientific Applications How to support performance analysis and tuning process consistent with application development methodology? Common Component Architecture (CCA) applications Performance tools should integrate with software Design performance observation component Measurement port and measurement interfaces Build support for application component instrumentation Interpose a proxy component for each port Inside the proxy, track caller/callee invocations, timings Automate the process of proxy component creation using PDT for static analysis of components include support for selective instrumentation
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Flame Reaction-Diffusion (Sandia, J. Ray) CCAFFEINE
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Component Modeling and Optimization Given a set of components, where each component has multiple implementations, what is the optimal subset of implementations that solve a given problem? How to model a single component? How to model a composition of components? How to select optimal subset of implementations? A component only has performance meaning in context Applications are dynamically composed at runtime Application developers use components from others Instrumentation may only be at component interfaces Performance measurements need to be non-intrusive Users interested in a coarse-grained performance
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct MasterMind Component (Trebon, IPDPS 2004)
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Proxy Generator for other Applications TAU (PDT) proxy component for: QoS tracking [Boyana, ANL] Debugging Port Monitor for CCA (tracks arguments) SCIRun2 Perfume components [Venkat, U. Utah] Exploring Babel for auto-generation of proxies: Direct SIDL-to-proxy code generation Generating client component interface in C++ Using PDT for generating proxies
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Earth Systems Modeling Framework Coupled modeling with modular software framework Instrumentation for ESMF framework and applications PDT automatic instrumentation Fortran 95 code modules C / C++ code modules MPI wrapper library for MPI calls ESMF Component instrumentation (using CCA) CCA measurement port manual instrumentation Proxy generation using PDT and runtime interposition Significant callpath profiling used by ESMF team
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Using TAU Component in ESMF/CCA
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU’s Paraprof Profile Browser (ESMF Data) Callpath profile
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct CUBE Browser (UTK, FZJ) (ESMF Data) metriccalltree location TAU profile data converted to CUBE form
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU Traces with Counters (ESMF)
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Visualizing TAU Traces with Counters/Samples
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Uintah Computational Framework (UCF) University of Utah, Center for Simulation of Accidental Fires and Explosions (C-SAFE), DOE ASCI Center UCF analysis Scheduling MPI library Components Performance mapping Use for online and offline visualization ParaVis tools F 500 processes
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Scatterplot Displays (UCF, 500 processes) Each point coordinate determined by three values: MPI_Reduce MPI_Recv MPI_Waitsome Min/Max value range Effective for cluster analysis Relation between MPI_Recv and MPI_Waitsome
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Online Unitah Performance Profiling Demonstration of online profiling capability Multiple profile samples Each profile taken at major iteration (~ 60 seconds) Colliding elastic disks Test material point method (MPM) code Executed on 512 processors ASCI Blue Pacific at LLNL Example 3D bargraph visualization MPI execution time Performance mapping Multiple time steps
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Online Unitah Performance Profiling
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Miranda Performance Analysis (Miller, LLNL) Miranda is a research hydrodynamics code Fortran 95, MPI Mostly synchronous MPI_ALLTOALL on Np x,y communicators Some MPI reductions and broadcasts for statistics Good communications scaling ACL and MCR Linux cluster Up to 1728 CPUs Fixed workload per CPU Ported to BlueGene/L Breaking News! (see next slide)
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Profiling of Miranda on BG/L (Miller, LLNL) 128 Nodes512 Nodes1024 Nodes Profile code performance (automatic instrumentation) Scaling studies (problem size, number of processors) Run on 8K and 16K processors this week!
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Fine Grained Profiling via Tracing on Miranda Use TAU to generate VTF3 traces for Vampir analysis Combines MPI calls with HW counter information Detailed code behavior to focus optimization efforts
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Max Heap Memory (KB) used for problem on 16 processors of ASC Frost at LLNL Memory Usage Analysis BG/L will have limited memory per node (512 MB) Miranda uses TAU to profile memory usage Streamlines code Squeeze larger problems on the machine TAU’s footprint is small Approximately 100 bytes per event per thread
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Kull Performance Optimization (Miller, LLNL) Kull is a Lagrange hydrodynamics code Physics packages written in C++ and Fortran Parallel Python interpreter run-time environment! Scalar test problem analysis Serial execution to identify performance factors Original code profile indicated expensive functions CCSubzonalEffects member functions Examination revealed optimization opportunities Loop merging Amortizing geometric lookup over more calculations Apply to CSSubzonalEffects member functions
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Kull Optimization Optimized Exclusive Profile Original Exclusive Profile CSSubzonalEffects member functions total time Reduced from 5.80 seconds to 0.82 seconds Overall run time reduce from 28.1 to seconds
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Important Questions for Application Developers How does performance vary with different compilers? Is poor performance correlated with certain OS features? Has a recent change caused unanticipated performance? How does performance vary with MPI variants? Why is one application version faster than another? What is the reason for the observed scaling behavior? Did two runs exhibit similar performance? How are performance data related to application events? Which machines will run my code the fastest and why? Which benchmarks predict my code performance best?
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Multi-Level Performance Data Mining New (just forming) research project PSU: Karen L. Karavanic Cornell: Sally A. McKee UO: Allen D. Malony and Sameer Shende LLNL: John M. May and Bronis R. de Supinski Develop performance data mining technology Scientific applications, benchmarks, other measurements Systematic analysis for understanding and prediction Better foundation for evaluation of leadership-class computer systems “Scalable, Interoperable Tools to Support Autonomic Optimization of High-End Applications,” S. McKee, G. Tyson, A. Malony, begin Nov. 1, 2004.
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct General Goals Answer questions at multiple levels of interest Data from low-level measurements and simulations use to predict application performance data mining applied to optimize data gathering process High-level performance data spanning dimensions Machine, applications, code revisions Examine broad performance trends Need technology Performance instrumentation and measurement Performance data management Performance analysis and results presentation Automated performance experimentation and exploration
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Specific Goals Design, develop, and populate a performance database Discover general correlations application performance and features of their external environment Develop methods to predict application performance on lower-level metrics Discover performance correlations between a small set of benchmarks and a collection of applications that represent a typical workload for a give system Performance data mining infrastructure is important for all of these goals Establish a more rational basis for evaluating the performance of leadership-class computers
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct PerfTrack: Performance DB and Analysis Tool PSU: Kathryn Mohror, Karen Karavanic UO: Kevin Huck LLNL: John May, Brian Miller (CASC)
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU Performance Data Management Framework
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct TAU Performance Regression (PerfRegress) Prototype developed by Alan Morris for Uintah Re-implement using PerfDMF
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Background – Ahn & Vetter, 2002 “Scalable Analysis Techniques for Microprocessor Performance Counter Metrics,” SC2002 Applied multivariate statistical analysis techniques to large datasets of performance data (PAPI events) Cluster Analysis and F-Ratio Agglomerative Hierarchical Method - dendogram identified groupings of master, slave threads in sPPM K-means clustering and F-ratio - differences between master, slave related to communication and management Factor Analysis shows highly correlated metrics fall into peer groups Combined techniques (recursively) leads to observations of application behavior hard to identify otherwise
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Similarity Analysis Can we recreate Ahn and Vetter’s results? Apply techniques from the phase analysis (Sherwood) Threads of execution can be compared for similarity Threads with abnormal behavior show up as less similar Each thread is represented as a vector (V) of dimension n n is the number of functions in the application V = [f 1, f 2, …, f n ] (represent event mix) Each value is the percentage of time spent in that function normalized from 0.0 to 1.0 Distance calculated between the vectors U and V: ManhattanDistance(U, V) = ∑ |u i - v i | i=0 n
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct sPPM on Blue Horizon (64x4, OpenMP+MPI) TAU profiles 10 events PerfDMF threads 32-47
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct sPPM on MCR (total instructions, 16x2) TAU/PerfDMF 120 events master (even) worker (odd)
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct sPPM on MCR (PAPI_FP_INS, 16x2) TAU profiles PerfDMF master/worker higher/lower Same result as Ahn/Vetter
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct sPPM on Frost (PAPI_FP_INS, 256 threads) View of fewer than half of the threads of execution is possible on the screen at one time Three groups are obvious: Lower ranking threads One unique thread Higher ranking threads 3% more FP Finding subtle differences is difficult with this view
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Dendrogram shows 5 natural clusters: Unique thread High ranking master threads Low ranking master threads High ranking worker threads Low ranking worker threads sPPM on Frost (PAPI_FP_INS, 256 threads) TAU profiles PerfDMF R direct access to DM R routine threads
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct sPPM on MCR (PAPI_FP_INS, 16x2 threads) masters slaves
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct sPPM on Frost (PAPI_FP_INS, 256 threads) After K-means clustering into 5 clusters Similar clusters are formed (seed with group means) Each cluster’s performance characteristics analyzed Dimensionality reduction (256 threads to 5 clusters!) SPPMINTERFDIFUZEDINTERFBarrier [OpenMP:runhyd3.F ]
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct PerfExplorer Design (K. Huck, UO) Performance knowledge discovery framework Use the existing TAU infrastructure TAU instrumentation data, PerfDMF Client-server based system architecture Data mining analysis applied to parallel performance data Technology integration Relational DatabaseManagement Systems (RDBMS) Java API and toolkit R-project / Omegahat statistical analysis Web-based client Jakarta web server and Struts (for a thin web-client)
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct PerfExplorer Architecture Client is a traditional Java application with GUI (Swing) Server accepts multiple client requests and returns results PerfDMF Java API used to access DBMS via JDBC Server supports R data mining operations built using RSJava Analyses can be scripted, parameterized, and monitored Browsing of analysis results via automatic web page creation and thumbnails
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct ZeptoOS: Extreme Performance Scalable OS’s DOE, Office of Science OS / RTS for Extreme Scale Scientific Computation Argonne National Lab and University of Oregon Investigate operating system and run-time (OS/R) functionality required for scalable components used in petascale architectures Flexible OS/R functionality Scalable OS/R system calls Performance tools, monitoring, and metrics Fault tolerance and resiliency Approach Specify OS/R requirements across scalable components Explore flexible functionality (Linux) Hierarchical designs optimized with collective OS/R interfaces Integrated (horizontal, vertical) performance measurement / analysis Fault scenarios and injection to observe behavior
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct ZeptoOS Plans Explore Linux functionality for BG/L Explore efficiency for ultra-small kernels Scheduler, memory, IO Construct kernel-level collective operations Support for dynamic library loading, … Build Faulty Towers Linux kernel and system for replaying fault scenarios Extend TAU Profiling OS suites Benchmarking collective OS calls Observing effects of faults
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Concluding Discussion As high-end systems scale, it will be increasingly important that performance tools be used effectively Performance observation methods do not necessarily need to change in a fundamental sense Just need to be controlled and used efficiently More intelligent performance systems for productive use Evolve to application-specific performance technology Deal with scale by “full range” performance exploration Autonomic and integrated tools Knowledge-based and knowledge-driven process Deliver to community next-generation tools
Performance Technology for Productive, High-End Parallel ComputingLLNL, Oct Support Acknowledgements Department of Energy (DOE) Office of Science contracts University of Utah ASCI Level 1 sub-contract ASC/NNSA Level 3 contract NSF High-End Computing Grant Research Centre Juelich John von Neumann Institute Dr. Bernd Mohr Los Alamos National Laboratory