
605-427-2649 |
I N T R O D U C I N G...
HPC Software - ezParaPerf - a Fortran Parallel Performance Toolkit
Today's High Performance Computer systems require good HPC software tools to understand and therefore improve parallel performance.
This is the tool your software team has been waiting for
Portable to allmost all HPC Software Systems
Powerful Yet Easy to Use (Very Short Learning Curve)
- Run our automatic Fortran software instrumenter
- Link to our runtime library
- Learn to Run our profiler
- Measures CPU, Network, and I/O performance
- Reports software performance variation across processors
Built-in Hardware Performance Counters Used by the HPC software
- Modern Processors have built-in hardware to measure Events
- Events can be mflops, cache misses, tlb-misses ...
- No performance penalty in counting events
- Very small penalty to retrieve counters
- Modern Processors can measure 2-8 events simultaneously
- PAPI provides portable API for many processors' event counters. (Intel, SGI, Cray, ...)
What does the Fortran Parallel software Performance Tool Provide?
- It automatically measures each routine's performance- It also makes it easy split large routines into code sections
The HPC Fortran software Instrumenter (f95inst)
- f95inst is quick and automatically instruments software subroutines for:
. Cpu and wall time ( inclusive and exclusive )
. PAPI events
. Message-passing
. Fortran I/O
- f95inst instruments sourcecode before any pre-processing or compilation
- f95inst has USER directive for blocks of code with performance issues
- f95inst has NOINST directive is for suppressing instrumentation of small routines
- f95inst allows the user to reinstrument and instrumentation is easily removed with a switch
- f95inst supports conditional instrumentation using C-preprocessor directives
- f95inst supports both free and fixed format codes
- f95inst allows the user to control and modify which PAPI events are of current interest
HPC software MPI calls
- f95inst provides cpu and wall wait-times for asynchronous operations by subroutine
- f95inst provides cpu, wall time, and transfer rates for synchronous operations by subroutine
HPC Parallel Fortran I/O
- f95inst captures wall time of standard Fortran I/O by software subroutine
The ezParaPerf Runtime Library (libparaperf.a)
- Our Runtime Library writes a runtime file for each processor in XML-format
- Each XML file contains:
- Job Stats
- Machine Stats
- User-supplied information: such as how code was built, what kind of run it was
- Performance Data
- We strongly recommend you link with both our Runtime Library and the Public Domain PAPI Library
- You may download Performance Application Programming Interface (PAPI) from http://icl.cs.utk.edu/papi/
License
The easy Fortran parallel software performance tool (ezParaPerf) is
proprietary code designed by two PhDs that are HPC experts. The Fortran95
instrumenter (f95inst) requires
two public domain programs: the Performance Database Toolkit (pdtool-kit-3.4) from
the University of Oregon Performance Research Laboratory can be downloaded
from the URL: http://www.cs.uoregon.edu/research/pdt/home.php and the Python module PLEX (Plex-v1.1.4.1) written by Greg
Ewing of the Computer Science Department at the University of Canterbury,
Christchurch, New Zealand.