Re: [PerfSuite-bugs] [PerfSuite 0.6.1rc1] fails building with Intel compiler

[Rick Kufrin <rkufrin@nc...>]

Hello again,

I'll try to combine information about your two reports with testsuite 
output (thank you for including) into one reply.  There are several things 
going on here.

Regarding building PerfSuite with Intel compilers: you're correct that the 
inline assembly in the timing routines (and, on x86/x86-64, the processor 
detection routines) will cause it to choke.  The solution to this is to 
build PerfSuite with gcc but to cause the Intel Fortran/C calling 
convention to be used through the variable F77.  In other words, PerfSuite 
itself should always be built with gcc, but that doesn't restrict its use 
with other compilers.  There's very little computation done by PerfSuite, 
so any speed differences on IA-64 with icc/gcc don't really matter (that's 
not the case, of course with the application code!).  So all you should 
have to do is remove "CC=icc" and "CXX=icpc" from your configure line, but 
retain "F77=ifort", in order to pick up the ifort/C calling convention.

At NCSA, we typically build dual builds of PerfSuite to handle both g77 
and Intel compilers, with the Intel libraries installed in 
$PREFIX/lib/intel.

Regarding the other test case failures:  these, I believe, are due to the 
version of PAPI you're using, which appears to be 2.3.4.3.  There have 
been a number of changes in PAPI, and the most recent version (3.0.7) has 
many fixes to problems that existed before.  In fact, just yesterday a 
problem was identified and fixed that caused incorrect results (failures) 
on Altix/Itanium 2.  My experience has been that the two problem spots in
PAPI 2 have been a) multiplexing and b) pthreads.  Again, the PAPI 
development team has made numerous improvements in PAPI to address these 
spots and so the best thing to do is to upgrade to PAPI 3.0.7 from the 
UTK/ICL web site.  You might even want to wait for 3.0.8 since changes 
like the one yesterday are probably still only in the PAPI CVS tree.

I hope these suggestions help - please let me know if problems still 
persist, and thanks for the very complete problem report!

Rick

On Fri, 7 Jan 2005, Constantinos Evangelinos wrote:

> On an IA-64 system running RedHat EL AS 3, with a patched 2.4.21 kernel.org 
> kernel (with the perfmon and SCORE patches) and PAPI 2.4.3.4 build with the 
> Intel compiler (v. 8.1), configuring Perfsuite with the Intel C/C++ compiler 
> (icc -no-gcc -cxxlib-gcc) fails when building timer.c because of and tricks 
> to use gcc in libtool only for timers.c fall flat on their face later on...
> 
> Perfsuite was configured as:
> 
> MPICPPFLAGS="-I/usr/local/pkg-ia64/mpich-vmi/mpich-vmi-1.2.5/intel/include"  
> CC=icc CFLAGS=-O2 CXX=icpc CXXFLAGS=-O2 F77=ifort FFLAGS=-O2  
> PTHREAD_LIBS=-lpthread ./configure  
> --with-tdom=/usr/local/pkg-ia64/tdom/tdom-0.8.1  
> --with-papi=/usr/local/pkg-ia64/papi/papi-2.3.4.3 --enable-tests --enable-mpi  
> --prefix=/usr/local/pkg-ia64/perfsuite/perfsuite-0.6.1rc1/papi2
> 
> The error is as follows:
>  icc -DHAVE_CONFIG_H -I. -I. -I../.. -O2 -MT timers.lo -MD -MP -MF 
> .deps/timers.Tpo -c timers.c  -fPIC -DPIC -o .libs/timers.o
> timers.c(315): error: asm statements not supported in this environment
>       rdtscll(tmp);
>       ^
>  
> timers.c(315): error: expected a ")"
>       rdtscll(tmp);
>       ^
> 
> and reconfiguring with icc -no-gcc -cxxlib-gcc etc. does not solve the 
> problem.
> 
> 

Re: [PerfSuite-bugs] Incomplete installation

[Rick Kufrin <rkufrin@nc...>]

Hello and thank you for your comments and feedback, they are much
appreciated.  Here are the answers to these questions:

> a) Why do PSconfig and psprocess need to be installed by executing make 
> install in their respective directories while the global gmake install 
> ignores them? Is there a problem with their use?

This is not what should be happening with a correct "configure" use. If
it is, then the problem is likely on our end with the configure 
script generation.

PSConfig and psprocess take an extra step during the installation process
because, as Tcl scripts, there is a substitution that occurs where the
location of the tclsh/wish interpreters on your system are substituted
where they should be in the driver scripts.  It's the equivalent of 
the #!/bin/sh initial line on sh scripts.  Still, this should require no 
intervention on your part.

If tclsh and wish are not found during the configuration process on
the installer's system, then the build/install process just skips
psprocess and PSConfig entirely, and I would have thought that is
what may have happened in your case, but I see in the testsuite.log
files that you have sent that they were found in /usr/bin/tclsh8.3
and /usr/bin/wish8.3 - so I can only guess that the problem must
be on this end and I'll take a closer look.  Thanks for reporting
this, as you've possibly guessed, one can install by hand but
that shouldn't be necessary... also there are several variables
set in the drivers (which should ultimately be installed in
$PREFIX/bin) that may need to be manually tailored for your install.
If you would like any help with that, feel free to email again. 

> b) Where are PSJava, PSTcl, PerfExplore, ProfView and OptView?

These are upcoming pieces that should appear at a later time (in
addition to others, such as PSPMPI).  The goal has been to limit
things to a manageable subset, hence so far the focus has been
on the base and hardware counter-related software, which is what
comprises "0.6.x" versions.

PSJava and PSTcl are next up, and you'll probably see them in spring
2005, under an alpha 0.6.2 release.  PSJava is the most recent, and
will have a package/class that will have similar functionality
to libpshwpc with an OO design and thread support.  There also
is a JVMTI agent that requires Java 1.5 which is like psrun in
counting mode, but for Java programs.  These produce the same
XML documents that the other software does and are post-processable
with psprocess or anything else that understands the format.  If you are 
interested in trying these when available, your feedback on the design 
and use will be very helpful.

PSTcl, in some sense, is partially there as it is used by psprocess and 
psconfig. For example, there is a Tcl PAPI C extension that is used by 
psprocess to learn about performance events, names.  You'll find the 
source to that in src/tcllib/pshwpc.  The Tcl support will be organized
and documented a bit more, and that is what will comprise PSTcl.  One 
thing that will change in 0.6.2 is that there will be no need to actually
have performance counter support on the system where you run psprocess so 
that, for example, you could generate XML documents on a cluster with 
perfctr/perfmon/PAPI support, transfer the documents to your workstation 
and review them there.  Currently, that's not the case.

The other pieces you mention are a bit farther off so I can't give
you a date on those but hopefully should have more when the time
comes.

> Thanks for a very useful suite of tools.

Thank you again for your feedback, I'm glad you find PerfSuite useful!

Rick

[PerfSuite-bugs] Incomplete installation

[Constantinos Evangelinos <ce107@MIT.EDU>]

I'm installing Perfsuite for our ACES clusters at MIT and I have a two 
questions:

a) Why do PSconfig and psprocess need to be installed by executing make 
install in their respective directories while the global gmake install 
ignores them? Is there a problem with their use?
b) Where are PSJava, PSTcl, PerfExplore, ProfView and OptView?

Thanks for a very useful suite of tools.
-- 
Dr. Constantinos Evangelinos                    Room 54-1518, EAPS/MIT
Earth, Atmospheric and Planetary Sciences       77 Massachusetts Avenue
Massachusetts Institute of Technology           Cambridge, MA 02139
+1-617-253-5259/+1-617-253-4464 (fax)           USA

[PerfSuite-bugs] [PerfSuite 0.6.1rc1] fails building with Intel compiler

[Constantinos Evangelinos <ce107@MIT.EDU>]

On an IA-64 system running RedHat EL AS 3, with a patched 2.4.21 kernel.org 
kernel (with the perfmon and SCORE patches) and PAPI 2.4.3.4 build with the 
Intel compiler (v. 8.1), configuring Perfsuite with the Intel C/C++ compiler 
(icc -no-gcc -cxxlib-gcc) fails when building timer.c because of and tricks 
to use gcc in libtool only for timers.c fall flat on their face later on...

Perfsuite was configured as:

MPICPPFLAGS="-I/usr/local/pkg-ia64/mpich-vmi/mpich-vmi-1.2.5/intel/include"  
CC=icc CFLAGS=-O2 CXX=icpc CXXFLAGS=-O2 F77=ifort FFLAGS=-O2  
PTHREAD_LIBS=-lpthread ./configure  
--with-tdom=/usr/local/pkg-ia64/tdom/tdom-0.8.1  
--with-papi=/usr/local/pkg-ia64/papi/papi-2.3.4.3 --enable-tests --enable-mpi  
--prefix=/usr/local/pkg-ia64/perfsuite/perfsuite-0.6.1rc1/papi2

The error is as follows:
 icc -DHAVE_CONFIG_H -I. -I. -I../.. -O2 -MT timers.lo -MD -MP -MF 
.deps/timers.Tpo -c timers.c  -fPIC -DPIC -o .libs/timers.o
timers.c(315): error: asm statements not supported in this environment
      rdtscll(tmp);
      ^
 
timers.c(315): error: expected a ")"
      rdtscll(tmp);
      ^

and reconfiguring with icc -no-gcc -cxxlib-gcc etc. does not solve the 
problem.

-- 
Dr. Constantinos Evangelinos                    Room 54-1518, EAPS/MIT
Earth, Atmospheric and Planetary Sciences       77 Massachusetts Avenue
Massachusetts Institute of Technology           Cambridge, MA 02139
+1-617-253-5259/+1-617-253-4464 (fax)           USA

[PerfSuite-bugs] [PerfSuite 0.6.1rc1] testsuite: 5 22 23 failed

[Constantinos Evangelinos <ce107@MIT.EDU>]

On an IA-64 system running RedHat EL AS 3, with a patched 2.4.21 kernel.org 
kernel (with the perfmon and SCORE patches) and PAPI 2.4.3.4 build with the 
Intel compiler (v. 8.1).

Perfsuite was configured as:

MPICPPFLAGS="-I/usr/local/pkg-ia64/mpich-vmi/mpich-vmi-1.2.5/intel/include"  
CC=gcc CFLAGS=-O2 CXX=g++ CXXFLAGS=-O2 F77=ifort FFLAGS=-O2  
PTHREAD_LIBS=-lpthread ./configure  
--with-tdom=/usr/local/pkg-ia64/tdom/tdom-0.8.1  
--with-papi=/usr/local/pkg-ia64/papi/papi-2.3.4.3 --enable-tests --enable-mpi  
--prefix=/usr/local/pkg-ia64/perfsuite/perfsuite-0.6.1rc1/papi2

-- 
Dr. Constantinos Evangelinos                    Room 54-1518, EAPS/MIT
Earth, Atmospheric and Planetary Sciences       77 Massachusetts Avenue
Massachusetts Institute of Technology           Cambridge, MA 02139
+1-617-253-5259/+1-617-253-4464 (fax)           USA