Page 1 of 1

Stalling LAPACK tests

PostPosted: Mon Oct 17, 2005 8:20 pm
by vnivargi
Hi !
I have been trying to get the LAPACK Tests to run on my machine, but am not getting consistent results. The Real Lapack Linear Equation routines tests are not completing in a predictable way ... sometimes the tests complete, other times the test appears hung [but with CPU at full throttle].

This is the Test: ---
Testing REAL LAPACK linear equation routines
./xlintsts < stest.in #> stest.out 2>&1
Tests of the REAL LAPACK routines
----
and this is where the test stalls:---
M = 5, N = 50, type 2, test 6, ratio = 0.88853E+06
M = 5, N = 50, type 3, test 2, ratio = 0.10829E+06
M = 5, N = 50, type 3, test 3, ratio = 0.71292E+06
M = 5, N = 50, type 3, test 5, ratio = 0.22219E+06
M = 5, N = 50, type 3, test 6, ratio = 0.92624E+06
M = 10, N = 50, type 1, test 3, ratio = 0.33554E+06
M = 10, N = 50, type 1, test 6, ratio = 0.33554E+06
----

Any reason for this which I may have overlooked?
I am running this on a 64-bit Dual Opteron.
Lemme know if some more info is needed.

Regards,
Vaibhav

PostPosted: Mon Oct 17, 2005 10:13 pm
by Julien Langou
Hello:
- which BLAS are you using?
- which compiler, which flags?
all this matters. Just copy paste the make.inc if you wish.
Thanks,
Julien Langou

PostPosted: Tue Oct 18, 2005 5:48 pm
by vnivargi
Hi Julien,
I am using:
1. The BLAS Lib which ships with LAPACK 3.0
2. g77
--- from make.inc
PLAT = _LINUX
FORTRAN = g77
OPTS = -funroll-all-loops -O3
DRVOPTS = $(OPTS)
NOOPT =
LOADER = g77
LOADOPTS =

ARCH = ar
ARCHFLAGS= -crv
RANLIB = ranlib
---

Regards,
Vaibhav

PostPosted: Tue Oct 18, 2005 10:01 pm
by Julien Langou
Hello,

Installation of LAPACK with g77 and reference BLAS on 64-bit Opteron should be no problem. I just tried again and everything was fine. I am using g77 version 3.4.3 in 64-bit mode with the same make.inc as yours.

so your BLAS library line in make.inc looks like:
BLASLIB = ../../blas_LINUX.a
and you have obtained blas_LINUX.a by doing
make blaslib
in the LAPACK directory.

I guess at this level, can you do a
rm -f lapack_LINUX.a blas_LINUX.a ; make clean ; make blaslib ; make

and check again the behaviour of ./xlintsts. Because I do not really know what's going on.

Julien

PostPosted: Wed Oct 19, 2005 12:54 pm
by vnivargi
Hi Julien,
Well the problem is that the test completes like 50% of the time ... and the remaining times I see the test stalled at M=10...

Is there some place where I can get documentation about the tests, and their standard runtimes, etc?

Thanks,
Vaibhav

PostPosted: Sat Oct 22, 2005 1:45 am
by Julien Langou
Hello Vaibhav,

Well the problem is that the test completes like 50% of the time ... and the remaining times I see the test stalled at M=10...


Honnestely I do not know where it comes from. I am using an AMD Opteron 246 Dual Core 2GHz using g77 3.4.4 with -m64.
What you describes look as if you have had in the compilation line a -fnof2c or strange thing like this. But a priori not.
Do you have at least something working in 32-bit mode....?

Maybe you have one dumb core so 50% of the time you stall, :)

seriously I do not know

Is there some place where I can get documentation about the tests, and their standard runtimes, etc?


Reference working note for the test suite of LAPACK is LAWN#41
http://www.netlib.org/lapack/lawnspdf/lawn41.pdf
from Susan Blackford and Jack Dongarra.

There is no document up-to-date about the standard runtime of the testing routines.

So let's start now!

Code: Select all
AMD Dual Core 246 Opteron in 64 bit mode with GNU Fortran 3.4.4 -O3 and reference BLAS (the BLAS provided with LAPACK):

Testing otal time                              : 5:26.80s

Timing Total time                              : 24:26.05s


I'll ask around to see if somebody as an idea to figure out what your problem is.

Julien