1. University of Maryland Towards Automated Tuning of Parallel Programs Jeffrey K. Hollingsworth hollings@cs.umd.edu Department of Computer Science University of Maryland, College Park, MD 20742

2. University of Maryland Why Automated Performance Tuning? Software commonly have parameters that impact their performance. Optimal parameter values are usually variable and un-predictable. Automated Parameter tuning can be used for adaptive parameter tuning in complex software.

3. University of Maryland Automated Performance Tuning Goal: Maximize achieved performance Problems: –Large number of parameters to tune –Shape of objective function unknown –Multiple libraries and coupled applications –Analytical model may not be available Requirements: –Runtime tuning for long running programs –Don ’ t try too many configurations –Avoid gradients

4. University of Maryland Active Harmony Runtime performance optimization –Can also support training runs Automatic library selection (code) –Monitor library performance –Switch library if necessary Automatic performance tuning (parameter) –Monitor system performance –Adjust runtime parameters Hooks for Compiler Frameworks –Working to integrate USC/ISI Chill –Looking at others too

5. University of Maryland Example: Cluster Based Web Server 3-tier system Harmony Provides –Parameter updates for DB, and App Severs TPC-W Benchmark –Transactional web benchmark –Mimic operations of an e-commerce site –Uses Java implementation from Univ. of Wisconsin –Performance metrics Web Interaction Per Second (WIPS)

6. University of Maryland Cluster-Based Web Service Tuning Best configuration after 200 iterations BrowsingShoppingOrdering Improvements compared to the default configuration 15% 16% 5%

7. University of Maryland Importance of various parameters

8. University of Maryland A Bit More About Harmony Search Pre-execution –Sensitivity Discovery Phase –Used to Order not Eliminate search dimensions Online –Use Parallel Rank Order Search Different configurations on different nodes

9. University of Maryland Benefits of Searching in Parallel Performance for Two Programs –High Performance Linpack –POP Ocean Code

10. University of Maryland 10 Integrating Compiler Transformations

11. University of Maryland 11 Performance of Matrix Multiply ActiveHarmony + CHiLL vs ATLAS and MKL

12. University of Maryland Must Coordinate Auto Tuners Problem: Warring auto tuning systems Multiple components “ auto tuning ” at once Tuning based on multiple changes at once Solution: –Need some level of coordination –Possible Answer: Exposing different tuning systems –Part of PERI Auto-tuning Effort

13. University of Maryland Conclusion Active Harmony –An infrastructure for runtime tuning –Automatic tuning and environment adaptation –It works ! Auto Tuning Integration –Need to integrate multiple tools Compilers, runtime, application

14. University of Maryland Acknowledgements Coding and Experiments –I-Hsin Chung (IBM Watson) –Vahid Tabatabaee –Ananta Tiwari Chill Integration Effort –Marry Hall (Utah) –Chun Chen (USC/ISI) –Jacqueline Chame (USC/ISI) Funding –DOE – PERC/PERI –NSF –LTS (DoD)