[petsc-dev] Handling ESSL + Lapack and PESSL

[Aron Ahmadia]

Just to confirm that we were seeing illegal instruction crashes linking
against:

-lesslbgsmp

but

-lesslbg

appears to be fine.  There may be an environment variable or compile option
to set for the SMP-friendly ESSL to work properly.  I don't have any clue.

A

On Tue, Sep 14, 2010 at 11:56 PM, Aron Ahmadia <aron.ahmadia at kaust.edu.sa>wrote:

> Jed, I went back and recompiled a PETSc that didn't make use of the KTH
> LAPACK/CBLAS build, so the problem definitely lies within the daxpy calls to
> the essl library, and not PETSc's interpretation of the interface.  At this
> point the potential culprits are misaligned memory or some bad
> initialization to the multithreaded ESSL.  In the past I've used the
> smp-friendly ESSL without any trouble, so I am surprised that this code is
> throwing illegal instruction signals.  I am rebuilding PETSc now to point to
> the single-threaded BG ESSL.  If this fails as well I will get some help
> from the IBM guys to shed some light on what's going on.
>
> --Aron
>
>
> On Mon, Sep 13, 2010 at 2:59 PM, Aron Ahmadia <aron.ahmadia at kaust.edu.sa>wrote:
>
>> ESSL contains a BLAS, but you can get a nice standard LAPACK/CBLAS on
>> BlueGene by following these instructions:
>> http://www.pdc.kth.se/resources/computers/bluegene/LAPACK-CBLAS/LAPACK-CBLAS-build
>>
>> A
>>
>>
>> On Mon, Sep 13, 2010 at 2:55 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>>
>>>
>>> On Sep 13, 2010, at 1:56 PM, Jed Brown wrote:
>>>
>>> > On Mon, 13 Sep 2010 13:42:05 -0500, Barry Smith <bsmith at mcs.anl.gov>
>>> wrote:
>>> >> My point is not only do they get a slower GEEV but they get a slower
>>> >> everything in LAPACK. [...] Why not just use LAPACK?
>>> >
>>> > Yup, I get your point.  But Lapack usually needs a BLAS, and the
>>> > original (flawed) idea was that a fast BLAS would produce a competitive
>>> > Lapack.
>>> >
>>>    What, Jack made a mistake? Never.
>>>
>>>    Does ESSL itself contain "the BLAS" or is there also a libblas type
>>> thingy on the IBM you could link against?
>>>
>>>    None of this really matters because you are still stuck writing the
>>> configure test for geev.
>>>
>>>    Barry
>>>
>>>
>>> > Jed
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20100915/8e1d5d09/attachment.html>