Summertime Fun Everyone loves performance

1
Summertime FunEveryone loves performance

• Shirley Browne, George Ho, Jeff Horner, Kevin
  London, Philip Mucci, John Thurman

2
Projects

• PerfAPI
• Cache Simulator
• DOD Performance Optimization
• Rice/PET Collaboration
• Benchmarking
• HPD implementation
• Graduation

3
PerfAPI

• Research gathered on the following platforms
• Sun Ultra
• Pentium Pro/II
• IBM Power Series
• MIPS R10000
• DEC Alpha, Cray T3E
• Cray T90/SV1

4
PerfAPI

• Standard API to access hardware performance
  counters
• Standard set of definitions for performance
  metrics
• Resulting in
• Data for performance tool developers
• Data for tuning and evaluating applications
• Portable performance tools for every major
  platform

5
PerfAPI

• Research on the users needs through
• Mailing list
• Web page http//www.cs.utk.edu/mucci/pdsa
• Collaboration with existing researchers
• Vendors (Sun, Cray, Digital, SGI)
• SPDT98/SC98 poster

6
PerfAPI - Coming this fall

• Draft API by 8/31 released to mailing list
• 9/31 revisions incorporated into a tech report
• 10/31 implementations for MIPS and Ultra
• 11/31 Implementations for IBM and Intel
• 12/31 Implementations for Alpha
• 12/31 Portable hardware counter based prof

7
Cache Simulator

• Motivated by the need for information correlated
  with the source code and run-time reference
  patterns
• Redesign into object oriented structure and raw
  output format
• Statistical reduction techniques
• GUI design
• Parser design

8
Cache Simulator

• GUI written in Java
• Parsers written using Octave from Edinburg. (?)
• Tool will allow browsing and instrumentation of
  source
• Reporting will be done with perl scripts
• Money from Sandia pending
• Conflict matrix adopted by IBM/Watson

9
DOD Performance Optimization

• Optimization Tutorial and Poster at Users Group
  Meeting
• Visit to ASC to help scalability of Cobalt
• Putting together a performance team with a suite
  of in-house tools.
• Meeting at Rice to direct a possible
  collaboration with Rice on run-time data
  collection for optimization.

10
DOD Performance Optimization

• Upcoming tutorial at ARL 2 day 1 day workshop
• Speaking at annual ARL UGM
• Work with Cobalt, MAGI, GAMESS, HELIX
• Developers reluctance!
• Lack of tools!
• PET lead?

11
DOE Benchmarking

• Attended Weather/Climate modeling conference in
  June.
• Need for standardized benchmarks
• Complete lack of understanding of performance
• Opportunity for UT involvement
• Meeting in September at NCAR
• Virgin territory?

12
Benchmarking

• Modifications to 3 benchmarks plus a new one.
• MPBench
• BLASBench
• CacheBench
• ClockBench
• Standardized options and graph generation
• Integration into a Low-Level suite

13
Benchmarking MSRC Machines

• Completed benchmark runs for BlasBench,
  CacheBench and MPBench on all the MSRC platforms.
• In the Process of finishing Dedicated Runs of the
  ParkBench Suite.
• Made graphs of the results and added them into
  the BenchRib repository.

14
MPBench

• Add alltoall test
• Dynamic memory allocation
• NUMA measurement
• Cache flushing code

15
Benchmarking

• BLASbench
• Addition of solvers
• CacheBench
• Latency benchmark
• ClockBench
• Timer accuracy

16
HPD Implementation

• Elimination of Dolphin as a solution provider
• Received front end parser
• Building distributed Perl framework
• DBX/GDB as backends