High Performance Computing in Aachen

(1)

in Aachen

Christian Iwainsky

[email protected]

Center for Computing and Communication

RWTH Aachen University

Produktivitätstools unter Linux

Sep 16, RWTH Aachen University

(2)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Agenda

Lecture session:

1. Introduction to productivity tools

C. Iwainsky

2. Sun Studio Thread Analyzer

C. Terboven

3. Intel Threading Tools

4. Marmot

C. Iwainsky

5. Performance Analysis Basics

6. Sun Studio Performance Analyzer

7. Intel Tracing Tools

8. Scalasca

9. Vampir

(3)

Introduction to Sun Performance Analyzer …

1. Overview of the Sun Performance Analyzer and Collector

2. How to use

3

(4)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Overview of the Sun Performance Analyzer …

Facts about Sun Performance Analyzer

• collection of tools, i.e.:

• analyzer

• collect

• works with unmodified binaries

• must be dynamically linked to gain MPI information

• supports C, C++, Fortran and Java

• supports many parallel paradigms:

• OpenMP

• MPI

• POSIX and Solaris threads

(5)

Overview of the Sun Performance Analyzer …

Facts about Sun Performance Analyzer

• provides a very sophisticated GUI

• available for

Sun Solaris

Linux

• simple to use

• uses sampling and library wrapping to obtain information

• also called Sun Performance Analyzer and Collector

Summarily called “

analyzer

” from now on.

(6)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

How to use

To use the Analyzer follow the steps:

1. Execute your application:

OpenMP:

collect -o Mdata.er yourApplication

MPI/MPI and OpenMP

mpiexec -n N collect –m on –g Mdata.erg yourApplication

2. Analyze data:

analyze MData

(7)

Thanks

The next slides have been kindly provided by:

Ruud van der Pas

Senior Staff

Technical Developer

Technologies, Sun Microsystems

(8)

(9)

The collect command

% collect

NOTE: SunOS-64-bit, 1 CPUs, i386 5.11 system amd is correctly patched and set up for use with the Performance tools.

NOTE: The J2SE[tm] version 1.6.0 found at java (picked by PATH) is supported by the Performance tools.

usage: collect <args> target <target-args>

Sun Analyzer 7.6 SunOS_i386 2007/05/03

-p <interval> specify clock-profiling

clock profiling interval range on this system is from 0.500 to 1000.000 millisec.; resolution is 0.001 millisec. -h <ctr_def>...[,<ctr_n_def>]

specify HW counter profiling for up to 4 HW counters see below for more details

-s <threshold> specify synchronization wait tracing -r <option> specify thread analyzer experiment; see man page -H {on|off} specify heap tracing

-m {on|off} specify MPI tracing

-j {on|off|path} specify Java profiling

-J <java-args> specify arguments to Java for Java profiling -P <pid> use dbx to attach and collect data from running process -t <duration> specify time over which to record data

-x specify leaving the target waiting for a debugger attach -n dry run -- don't run target or collect performance data

-y <signal>[,r] specify delayed initialization and pause/resume signal When set, the target starts in paused mode;

if the optional r is provided, it starts in resumed mode -F {on|off|all|=<regex>} specify following descendant processes -A {on|off|copy} specify archiving of load-objects; default is on -S <interval> specify periodic sampling interval (secs.)‏

-L <size> specify experiment size limit (MB.)‏ -l <signal> specify signal for samples

... <many lines of output removed> ...

Only partial output

shown

Default experiment:

expt_name = test.1.er

clock profiling enabled, 10.007 millisec.

descendant processes will not be followed

periodic sampling, 1 secs.

experiment size limit 2000 MB.

experiment archiving: on

data descriptor: "p:10007;S:1;L:2000;A:1;"

host: `amd', cpuver = 2500, ncpus = 1,

clock frequency 2393 MHz.

(10)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Some of the options on “collect”

Option

Description

-o

Experiment name

-d

Experiment directory

-C

Store a label in the experiment

-p

Clock profiling

-m

MPI tracing

-g

Experiment group

-h

Hardware Counter profiling

-j

Java profiling

-F

Follow descendant processes

-A

Archiving of load-objects

-S

Periodic sampling interval

-L

Experiment size limit

-L

Signal for samples

-v

Print expanded log of processing

-R

Show README file and exit

-s

Specify synchronization wait tracing

-H

Specify heap tracing

-x

Specify leaving the target waiting for debugger attach

-y

Specify delayed initialization and pause/resume signal

(11)

(12)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Start the Analyzer

% analyzer

Select

experiment

Information on

experiment

Note: Can add/drop

experiments from

within the analyzer

(13)

Analyzer - Main Window

Most expensive

functions

(14)

(15)

Filter Data and Show/Hide Functions

Select

samples,

threads, LPW

and CPUs

Select which

load objects to

be shown (or

hidden)‏

(16)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Set and save your Preferences

metrics

shown

timeline

source

info

Save your

settings

(17)

Main window - More information

Additional

metrics selected

(18)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Source window

Compiler

Commentary

most expensive

statements

(19)

(20)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Timeline overview

System View

Application

View

Dynamic

Call Stack

for Event

Zoom in/out

Select An

Event

(21)

Timeline - More details

Initial behaviour

appears to be

(22)

HPC unter Linux IV

16.09.2008 – Ch. Iwainsky

Timeline - Zoom in

Very useful to analyze

application behaviour in

(23)

How to use

How setup the environment here at the RWTH-Aachen?

Load the module by typing:

module load studio/12

And proceed as described!