[Trilinos-Users] [EXTERNAL] Tpetra::MatrixMarket: Write and Read MultiVector

[Hoemmen, Mark]

Oh dear, I think I get it now... The chunk approach is fine for a single
column, but breaks for multiple columns.

The quick fix is to print out MultiVector one column at a time.  That
should be fine for now, because MultiVector was never intended to be used
with a very large number of columns.

Andrey, could you create a Tpetra bug for this and CC me?  Sorry for the
trouble!
mfh

On 2/11/15, 1:28 PM, "Andrey Prokopenko" <aprokop at sandia.gov> wrote:

>It seems that Tpetra's multivector writer is fundamentally broken.
>
>It does the writing in chunks, where each processor sends its part of
>multivector to processor zero.
>The problem is that each chunk is written consequently. So if a chunk
>has, say, 5 entries and two columns, the writer outputs 5 entries in the
>first column, and then 5 entries in the second column. Globally,
>however, it means that it mixes first and second columns.
>
>I think the problem started when it was decided that the writer is not
>scalable enough, and that we should not assemble the global vector on
>processor 0. It would have worked, except that MatrixMarket format
>requires column-major order. So no cigar.
>
>Mark,
>
>Is my understanding what's happening correct?
>
>-Andrey
>On 02/09/2015 07:21 AM, Sander Schaffner wrote:
>> Hello again,
>>
>> the new version helpt in the sense that it reads now the file. But
>> still it behaves like described before:
>>
>> [1 2 3 ... 63] -> [0 1 64 65 2 3 ...]
>> [64 65 66 ... 127] -> [32 33 96 97 34 35 ...]
>>
>> I wrote a small test program: First I create a 64x2 MV on 32 cores and
>> write into the first MV  [1 2 3 ... 63] and into the second [64 65 66
>> ... 127]. Then I write out the map and the MV. After that I read both
>> files to create a new MV to finally write it out again. Since I
>> changed nothing inbetween and only read and write a valid file i
>> expect two identical files. But this is not the case.
>>
>> Can someone tell me if I'm doing something wrong in my Code, or is it
>> maybe a bug?
>>
>> Thanks
>>
>> Sander
>>
>> Output first File (test_64_2):
>>
>> %%MatrixMarket matrix array real general
>> 64 2
>> 0
>> 1
>> 64
>> 65
>> 2
>> 3
>> 66
>> 67
>> 4
>> 5
>> 68
>> 69
>> 6
>> 7
>> ...
>>
>> Second (test_64):
>>
>> %%MatrixMarket matrix array real general
>> 64 2
>> 0
>> 1
>> 32
>> 33
>> 64
>> 65
>> 96
>> 97
>> 2
>> 3
>> 34
>> 35
>> 66
>> 67
>>
>> Am 06.02.2015 um 19:04 schrieb Andrey Prokopenko:
>>> Hi Sander,
>>>
>>> I believe that was fixed (see
>>> https://software.sandia.gov/bugzilla/show_bug.cgi?id=6264).
>>> Which version of Trilinos do you use?
>>>
>>> -Andrey
>>> On 02/06/2015 02:25 AM, Sander Schaffner wrote:
>>>> Hi
>>>>
>>>> I'm trying to write a MultiVector with
>>>> Tpetra::MatrixMarket::Writer.writeDenseFile. The goal is to read the
>>>> same MultiVector afterwards with
>>>> Tpetra::MatrixMarket::Reader.readDenseFile. But there are some
>>>> reasons why this doesn't work:
>>>>
>>>> If i generate a MultiVector with two Vectors (dimension 128) where
>>>> the first holds the values 0 and the second 1 I get this output on
>>>> 32 cores:
>>>>
>>>> %%MatrixMarket matrix array real general
>>>> 128 2
>>>> 0
>>>> 0
>>>> 0
>>>> 0
>>>>
>>>> 1
>>>> 1
>>>> 1
>>>> 1
>>>>
>>>> 0
>>>> 0
>>>> 0
>>>> 0
>>>>
>>>> 1
>>>> 1
>>>> 1
>>>> 1
>>>>
>>>> and so on. So I get first the local part of processor 0 for both
>>>> Vectors and so on.
>>>>
>>>> The reader does not accept this file. It complains about the spaces.
>>>> If I manually get ridd of them it reads the file. But then it
>>>> handles the first half of the values as Vector 1 and the second part
>>>> as Vector 2. So in the End we have the following:
>>>>
>>>> [0 0 0 0 0 ... 0] -> [0 0 0 0 1 1 1 1 ...]
>>>> [1 1 1 1 ... 1] -> [0 0 0 0 1 1 1 1 ...]
>>>>
>>>> Is there a way of directly read in a MatrixMarket file created by
>>>> the Writer? What other options are there to write/read MultiVectors?
>>>>
>>>> Any help is appreciated.
>>>>
>>>> Sander
>>>> _______________________________________________
>>>> Trilinos-Users mailing list
>>>> Trilinos-Users at software.sandia.gov
>>>> https://software.sandia.gov/mailman/listinfo/trilinos-users
>>>
>>
>>
>>
>> Am 06.02.2015 um 19:04 schrieb Andrey Prokopenko:
>>> Hi Sander,
>>>
>>> I believe that was fixed (see
>>> https://software.sandia.gov/bugzilla/show_bug.cgi?id=6264).
>>> Which version of Trilinos do you use?
>>>
>>> -Andrey
>>> On 02/06/2015 02:25 AM, Sander Schaffner wrote:
>>>> Hi
>>>>
>>>> I'm trying to write a MultiVector with
>>>> Tpetra::MatrixMarket::Writer.writeDenseFile. The goal is to read the
>>>> same MultiVector afterwards with
>>>> Tpetra::MatrixMarket::Reader.readDenseFile. But there are some
>>>> reasons why this doesn't work:
>>>>
>>>> If i generate a MultiVector with two Vectors (dimension 128) where
>>>> the first holds the values 0 and the second 1 I get this output on
>>>> 32 cores:
>>>>
>>>> %%MatrixMarket matrix array real general
>>>> 128 2
>>>> 0
>>>> 0
>>>> 0
>>>> 0
>>>>
>>>> 1
>>>> 1
>>>> 1
>>>> 1
>>>>
>>>> 0
>>>> 0
>>>> 0
>>>> 0
>>>>
>>>> 1
>>>> 1
>>>> 1
>>>> 1
>>>>
>>>> and so on. So I get first the local part of processor 0 for both
>>>> Vectors and so on.
>>>>
>>>> The reader does not accept this file. It complains about the spaces.
>>>> If I manually get ridd of them it reads the file. But then it
>>>> handles the first half of the values as Vector 1 and the second part
>>>> as Vector 2. So in the End we have the following:
>>>>
>>>> [0 0 0 0 0 ... 0] -> [0 0 0 0 1 1 1 1 ...]
>>>> [1 1 1 1 ... 1] -> [0 0 0 0 1 1 1 1 ...]
>>>>
>>>> Is there a way of directly read in a MatrixMarket file created by
>>>> the Writer? What other options are there to write/read MultiVectors?
>>>>
>>>> Any help is appreciated.
>>>>
>>>> Sander
>>>> _______________________________________________
>>>> Trilinos-Users mailing list
>>>> Trilinos-Users at software.sandia.gov
>>>> https://software.sandia.gov/mailman/listinfo/trilinos-users
>>>
>>
>
>