1. Allen D. Malony malony@cs.uoregon.edu Department of Computer and Information Science TAU Performance Research Laboratory University of Oregon Discussion: How to Address Tools Scalability

2. IBM Petascale Tools2 Scale and Scaling  What is meant by scale?  Processors  execution concurrency/parallelism   Memory  memory behavior, problem size  Network  concurrent communications  File system  parallel file operations / data size  Scaling in the physical size / concurrency of the system  What else?  Program  code size / interacting modules  Power  electrical power consumption  Performance  potential computational power  Dimension  Terascale … Petascale … and beyond

3. Discussion: How to Address Tools ScalabilityIBM Petascale Tools3 Tools Scalibility  Types of tools  Performance  analytical / simulation / empirical   Debugging  detect / correct concurrency errors  Programming  parallel languages / computation  Compiling  parallel code / libraries  Scheduling  systems allocation and launching  What does it mean for a tool to be scalable?  Tool dependent (different problems and scaling aspects)  What changes about the tool?  Naturally scalable vs. change in function / operation  Is a paradigm shift required?  To what extent is portability important?  What tools would you say are scalable? How? Why?

4. Discussion: How to Address Tools ScalabilityIBM Petascale Tools4 Focus – Parallel Performance Tools/Technology   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 data storage Performance Technology

5. Discussion: How to Address Tools ScalabilityIBM Petascale Tools5 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 required  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  What will enhance productive application development?  Paradigm shift in performance process and technology?

6. Discussion: How to Address Tools ScalabilityIBM Petascale Tools6 Instrumentation and Scaling  Make events visible to the measurement system  Direct instrumentation (code instrumentation)  Static instrumentation modifies code prior to execution  does not get removed (always will get executed)  source instrumentation may alter optimization  Dynamic instrumentation modifies code at runtime  can be inserted and deleted at runtime  incurs runtime cost  Indirect instrumentation generates events outside of code  Does scale affect the number of events?  Runtime instrumentation is more difficult with scale  Affected by increased parallelism

7. Discussion: How to Address Tools ScalabilityIBM Petascale Tools7 Measurement and Scaling  What makes performance measurement not scalable?  More parallelism  more performance data overall  performance data specific to each thread of execution  possible increase in number interactions between threads  Harder to manage the data (memory, transfer, storage)  Issues of performance intrusion  Performance data size  Number of event generated X metrics per event  Are there really more events? Which are important?  Control number of events generated  Control what is measured (to a point)  Need for performance data versus cost of obtaining it  Portability!

8. Discussion: How to Address Tools ScalabilityIBM Petascale Tools8 Measurement and Scaling (continued)  Consider “traditional” measurement methods  Profiling: summary statistics calculated during execution  Tracing: time-stamped sequence of execution events  Statistical sampling: indirect triggers, PC + metrics  Monitoring: access to performance data at runtime  How does the performance data grow?  How does per thread profile / trace size grow?  Consider communication  Strategies for scaling  Control performance data production and volume  Change in measurement type or approach  Event and/or measurement control  Filtering, throttling, and sampling

9. Discussion: How to Address Tools ScalabilityIBM Petascale Tools9 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

10. Discussion: How to Address Tools ScalabilityIBM Petascale Tools10 Analysis and Visualization Scalability  How to understand all the performance data collected?  Objectives  Meaningful performance results in meaningful forms  Want tools to be reasonably fast and responsive  Integrated, interoperable, portable, …  What does “scalability” mean here?  Performance data size  Large data size should not impact analysis tool use  Data complexity should not overwhelm interpretation  Results presentation should understandable  Tool integration and usability

11. Discussion: How to Address Tools ScalabilityIBM Petascale Tools11 Analysis and Visualization Scalability (continued)  Online analysis and visualization  Potential interference with execution  Single experiment analysis versus multiple experiments  Strategies  Statistical analysis  data dimension reduction, clustering, correlation, …  Scalable and semantic presentation methods  statistical, 3D  relate metrics to physical domain  Parallelization of analysis algorithms (e.g., trace analysis)  Increase system resources for analysis / visualization tools  Integration with performance modeling  Integration with parallel programming environment

12. Discussion: How to Address Tools ScalabilityIBM Petascale Tools12 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

13. Discussion: How to Address Tools ScalabilityIBM Petascale Tools13 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

14. Discussion: How to Address Tools ScalabilityIBM Petascale Tools14 ParaProf – Histogram View (Miranda) 8k processors 16k processors

15. Discussion: How to Address Tools ScalabilityIBM Petascale Tools15 ParaProf – 3D Full Profile (Miranda) 16k processors

16. Discussion: How to Address Tools ScalabilityIBM Petascale Tools16 ParaProf – 3D Scatterplot (Miranda)  Each point is a “thread” of execution  A total of four metrics shown in relation  ParaVis 3D profile visualization library  JOGL

17. Discussion: How to Address Tools ScalabilityIBM Petascale Tools17 Hierarchical and K-means Clustering (sPPM)

18. Discussion: How to Address Tools ScalabilityIBM Petascale Tools18 Vampir Next Generation (VNG) Architecture Merged Traces Analysis Server Classic Analysis:  monolithic  sequential Worker 1 Worker 2 Worker m Master Trace 1 Trace 2 Trace 3 Trace N File System Internet Parallel Program Monitor System Event Streams Visualization Client Segment Indicator 768 Processes Thumbnail Timeline with 16 visible Traces Process Parallel I/O Message Passing