perftools

perftools Overview

Perftools 5.3.2
===============
Release Date: April 19, 2012

On Cray XK systems:
Perftools 5.3.2 and Papi 4.3.0.1 must be used as a set with the CLE 4.0.UP03
operating system and requires Cudatoolkit 4.1.28. Perftools 5.3.2 is NOT
supported on CLE 4.0 UP02 on XK systems. Swapping Perftools
or Papi with an older version is not supported and will cause problems.
Using Cuda 4.0.17a and Cuda tools 4.0.17a is not supported and will cause
problems. Cuda 4.0.17a and Cuda tools 4.0.17a are replaced by
Cudatoolkit 4.1.28 which must be installed as default and used with
Perftools 5.3.2.

Purpose:
--------
Support for Cray XK systems running CLE 4.0 UP03 and with CudaToolkit 4.1.28
installed. Perftools 5.3.2 is not compatible with Cuda 4.0.17a and Cuda Tools
4.0.17a or earlier versions. This version contains no other bug fixes or
features.

CrayPat
-------
Support for CUDA 4.0 is deprecated.  To collect accelerator performance
statistics the cudatoolkit module must be loaded.

New accelerator hardware counter groups starting with Nvidia device driver
version 285.05.xx are available.  See the accpc man page for more information on
the new counter groups.  To determine the device driver version run the
following on a GPU compute node.  This command will not work on a non-GPU
compute node.

  aprun -n 1 cat /proc/driver/nvidia/version

  Bugs Fixed:
  -----------
  783476 - Nvidia device driver 285.05.36 accelerator hardware counter name
           change breaks perftools ACCPC groups

  Known Problems:
  ---------------
  783852 -  GPU codes using multiple OpenMP threads hang intermittently when
                ACCPC's are collected
	    For OpenMP codes that use the GPU with OpenAcc, CUDA, or PGI, the
            number of threads must be set to one when collecting accelerator
            hardware counters. If the number of threads is not set to one the
            application may hang or raise an error similar to the following:
                "Error 14 for CUPTI API function 'cuptiEventGroupEnable'.
                cuptiQuery failed"

  778771 - Data transfers with host time, but zero data transferred
  774412 - app2/5.2.2:  HWPC overview tab comes up blank
  777133 - (GCC_Bug 50497) pat_build produces no user function
               traces for gnu-4.6 compiled code
           As of GCC 4.6.x the .debug_pubnames section is no longer produced.
           CrayPat  relies on this section to gather global entry points that
           are eligible for tracing. To work around this issue until an
           alternative is provide by the performance tools, users can use GCC
           4.5.X or recompile those source files whose entry points are desired
           to be traced with the -finstrument-functions option, then instrument
            with pat_build -w and any other -g options.

Product and OS Dependencies:
----------------------------
Perftools/5.3.2 is supported on Cray XE systems running the
CLE 3.1UP00 CNL or later operating systems and on Cray XK systems
running CLE 4.0UP03 and later and the following product(s).
Update UP03 with Cuda Driver version 4.1 is required.

  xt-asyncpe 5.09 or later
  One or more compilers:
      CCE 8.0.0 or later
      GCC 4.5.X
      Intel 12.0 or later
      PGI 10.9 or later
  Windows 7 (for the desktop version of Cray Apprentice2)
  PAPI 4.3.0.1
  CUDATOOLKIT 4.1.28 (for Cray XK systems)

  Notes: - Not supported with Cuda 4.0.17a and Cuda Tools 4.0.17a or earlier
         - Seastar is not supported.
         - GCC 4.6.X is not supported.

Documentation:
--------------
See the following documents at http://docs.cray.com/
Cray Performance Analysis Tools 5.3 Release Overview and
    Installation Guide S-2474-53
Using Cray Performance Analysis Tools S-2376-52

Installation of PAPI:
---------------------
rpm -ivh cray-papi-4.3.0.1-1.x86_64.rpm

Cray XK systems with GPUs also need::
rpm -ihv cray-papi-acc-4.3.0.1-1.x86_64.rpm

To make PAPI 4.3.0.1 the default version of PAPI, execute:
/opt/cray/admin-pe/set_default_files/set_default_papi_4.3.0.1

Installation of perftools:
-------------
rpm -ivh perftools-clients-5.3.2-1.x86_64.rpm
rpm -ivh perftools-5.3.2-1.x86_64.rpm

To make perftools/5.3.2 the default version of perftools, execute:
/opt/cray/admin-pe/set_default_files/set_default_perftools_5.3.2


Installation of app2 remote client on Linux desktops/laptops:
------------------------------------------------------------
tar -xvzf perftools-remote-clients-5.3.2.tar.gz
cd perftools-remote-clients-5.3.2
./Install
  Type 'yes' to agree to software license prior to rpm install.
  If you're using modules software, load the perftools module to
  access the software.
module load perftools

Installation of app2 remote client (+ server) on Windows 7 systems:
-------------------------------------------------------------------
Apprentice2Installer_5.3.2.exe

The Cray Apprentice2 installer for Windows is included in the
perftools/5.3.1 package.  Download the Cray Apprentice2 installer
onto a desktop or laptop running Windows 7.  Double click on installer
to begin installation.  The installer will walk you through the process
for your system.

PAPI 4.3.0.1
==========
Purpose:
--------
Bug fixes and support for perftools 5.3.2

Documentation:
--------------
Visit the PAPI Reference pages for more information at:
    http://icl.cs.utk.edu/projects/papi/wiki/Main_Page
And visit the PAPI website for the latest updates:
    http://icl.cs.utk.edu/papi/


To re-display perftools/5.3.2 release information,
type:    less /opt/cray/perftools/5.3.2/release_info