S3D on Jaguar

From PERI

Contents 1 From Sameer Shende, 8/27/08 2 TAU Performance Data for S3D on Jaguar 3 From Heike Jagode, 9/23/08: S3D results 4 From Heike Jagode, 8/13/08

From Sameer Shende, 8/27/08

Please see the TAU S3D webpage for further information. Slides (PPT) about running S3D on Jaguar.

TAU was configured using: ./configure -useropt=-O2 -arch=craycnl -pdt=/ccs/home/amorris/apps/s3d/pdtoolkit/ -mpi -pdt_c++=g++ -LINUXTIMERS -papi=/opt/xt-tools/papi/default/v22/xt-cnl -MULTIPLECOUNTERS

TAU Performance Data for S3D on Jaguar

• 1p: .ppk, Java Webstart
• 4p: .ppk, Java Webstart
• 64p: .ppk, Java Webstart
• 512p: .ppk, Java Webstart
• 1728p: .ppk, Java Webstart
• 4096p: .ppk, Java Webstart
• 8000p: .ppk, Java Webstart
• 12000p: .ppk, Java Webstart

From Heike Jagode, 9/23/08: S3D results

Informal Report: Performance Measurement, Analysis, and Modeling of the S3D Application for Petaflops-scale Systems

From Heike Jagode, 8/13/08

• UTK/ORNL is collecting performance data for S3D on Jaguar (so far).
• focus on using hardware counters to measure occurrences of

performance related events:

• here a few examples:
  • to get information about application's efficiency (e.g. how many instructions per cycle)
  • L1 and L2 cache misses compared to hits
  • How is the memory access on quad-core for different core cases on Jaguar and BGP?
  • Control transfer-related measurements: how is the flow of instructions into processing pipeline (depending on branch predictions/mispredictions)
• run application on different node- as well as core cases starting from 64 cores (nodes)
• use TAU and SCALASCA to collect data