BLIS survey on NA-Digest

[dirk....@gmail.com]

I've just completed the survey. The question that its designers implicitly ask is: What will induce you to use BLIS?

The resolution of Issue 210 will in my case go a long way. https://github.com/flame/blis/issues/210

To someone who no longer works on the bleeding edge of numerical linear algebra, if it isn't in Debian, it does not exist.

[Lee Killough]

Basic Linear Algebra Persons!!!

--
You received this message because you are subscribed to the Google Groups "blis-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blis-discuss...@googlegroups.com.
To post to this group, send email to blis-d...@googlegroups.com.
Visit this group at https://groups.google.com/group/blis-discuss.
For more options, visit https://groups.google.com/d/optout.

[Lee Killough]

Like Field? Anyone heard of him? I hope so too.

[Field G. Van Zee]

Dirk,

Thanks for your feedback.

You are right that BLIS is not (to my knowledge) officially supported in
Debian, mostly because we have not yet had the chance to develop the
right relationships to facilitate inclusion (the process is
non-trivial). However, I would like to point out that Nico Schlömer
maintains a PPA for Ubuntu/Debian packages [1].

I would also like to point out that binary distributions are actually
not ideal for packages like BLIS. Sometimes the default configuration
options work fine. However, many (if not most) users need (or want) to
configure BLIS with a specific set of non-default configuration options:
BLIS integer size, BLAS integer size, threading options, enabling CBLAS,
etc. Binary packages lock the end-user into a single combination of
options that may not suit the user's needs. That said, as long as the
user knows of this potential shortcoming, the availability of binary
packages never hurts.

I can understand your desire to stick with binary packages. Nonetheless,
I invite you to try out a github clone. BLIS is very easy to compile and
install for most users, and if you run into any trouble, we are here to
help.

Regards,
Field


[1] https://github.com/flame/blis#external-linux-packages

[Jed Brown]

"Field G. Van Zee" <fi...@cs.utexas.edu> writes:

> I can understand your desire to stick with binary packages. Nonetheless,
> I invite you to try out a github clone. BLIS is very easy to compile and
> install for most users, and if you run into any trouble, we are here to
> help.

The issue is less about convenience for application developers as it is
distribution requirements for other libraries and applications that
deliver features to less technical end users. If you deliver software
to end users who don't know what a compiler is or don't know the name of
your package (because it's only used transitively), then there is
overwhelming incentive to not use BLIS because it would drastically
increase the complexity of distributing the entire software stack.

[Field G. Van Zee]

You make very good points, Jed. Thanks for reminding me of why there is
no substitution for being officially included in a distribution.

Hopefully, someone will come along with the motivation and connections
in the Debian/Ubuntu community to help us make this happen.

Field

[Jeff Hammond]

I must ask: does Matlab exist in your universe?

Jeff

--

Jeff Hammond
jeff.s...@gmail.com
http://jeffhammond.github.io/

[Jeff Hammond]

It is terrifying to think that you are seemingly okay with a world where Linux distribution package maintainers are the gatekeepers of computational science.

I am particularly concerned about this path since many Linux distributions provide Open-MPI by default when "mpi" is a dependency and did so when Open-MPI was utterly broken for both MPI_THREAD_MULTIPLE and other major features of the MPI-2 standard (RMA).  Maintainers of software that depended on working MPI-2 software were we well aware that their peers were shipping broken software to millions of Linux users but they had no power to address it because these communities apparently lack the governance structure that would support positive change.

[Jed Brown]

Jeff Hammond <jeff.s...@gmail.com> writes:

> On Mon, Aug 27, 2018 at 3:04 PM, Jed Brown <j...@jedbrown.org> wrote:
>
>> "Field G. Van Zee" <fi...@cs.utexas.edu> writes:
>>
>> > I can understand your desire to stick with binary packages. Nonetheless,
>> > I invite you to try out a github clone. BLIS is very easy to compile and
>> > install for most users, and if you run into any trouble, we are here to
>> > help.
>>
>> The issue is less about convenience for application developers as it is
>> distribution requirements for other libraries and applications that
>> deliver features to less technical end users. If you deliver software
>> to end users who don't know what a compiler is or don't know the name of
>> your package (because it's only used transitively), then there is
>> overwhelming incentive to not use BLIS because it would drastically
>> increase the complexity of distributing the entire software stack.
>>
>
> It is terrifying to think that you are seemingly okay with a world where
> Linux distribution package maintainers are the gatekeepers of computational
> science.

I made a statement about library dependencies in any environment that
eschews bundling (the alternative is a security disaster).

The MPI Forum kneecapped themselves by neglecting to even provide
recommendations on a common ABI, but that is a much broader topic.

[Jeff Hammond]

On Mon, Aug 27, 2018 at 3:25 PM, Jed Brown <j...@jedbrown.org> wrote:

Jeff Hammond <jeff.s...@gmail.com> writes:

> On Mon, Aug 27, 2018 at 3:04 PM, Jed Brown <j...@jedbrown.org> wrote:
>
>> "Field G. Van Zee" <fi...@cs.utexas.edu> writes:
>>
>> > I can understand your desire to stick with binary packages. Nonetheless,
>> > I invite you to try out a github clone. BLIS is very easy to compile and
>> > install for most users, and if you run into any trouble, we are here to
>> > help.
>>
>> The issue is less about convenience for application developers as it is
>> distribution requirements for other libraries and applications that
>> deliver features to less technical end users.  If you deliver software
>> to end users who don't know what a compiler is or don't know the name of
>> your package (because it's only used transitively), then there is
>> overwhelming incentive to not use BLIS because it would drastically
>> increase the complexity of distributing the entire software stack.
>>
>
> It is terrifying to think that you are seemingly okay with a world where
> Linux distribution package maintainers are the gatekeepers of computational
> science.

I made a statement about library dependencies in any environment that
eschews bundling (the alternative is a security disaster).

The MPI Forum kneecapped themselves by neglecting to even provide
recommendations on a common ABI, but that is a much broader topic.

Fortran provides no mechanism for creating a common ABI even with a single compiler.  Fortran remains ~50% of the usage of MPI.  Whatever ABI solution you are imagining does not address the majority of users.

In any case, I'm not sure somebody who works on a project that has compile-time options like --with-scalar-type=<real, complex> --with-precision=<single, double, quad> --with-64-bit-indices should be making fun of MPI's ABI problems :-P

Jeff

[Jed Brown]

Jeff Hammond <jeff.s...@gmail.com> writes:

>> The MPI Forum kneecapped themselves by neglecting to even provide
>> recommendations on a common ABI, but that is a much broader topic.
>>
> Fortran provides no mechanism for creating a common ABI even with a single
> compiler. Fortran remains ~50% of the usage of MPI. Whatever ABI solution
> you are imagining does not address the majority of users.

By what definition of "usage of MPI"? People who self-selected to
complete a survey or people who use software that depends on MPI?

> In any case, I'm not sure somebody who works on a project that has
> compile-time options like --with-scalar-type=<real,
> complex> --with-precision=<single, double, quad> --with-64-bit-indices
> should be making fun of MPI's ABI problems :-P

Touché, though the comparison is imperfect for many reasons outside the
scope of this mailing list.

[dirk....@gmail.com]

Op maandag 27 augustus 2018 23:53:58 UTC+2 schreef Field Van Zee:

> You are right that BLIS is not (to my knowledge) officially supported in
> Debian, mostly because we have not yet had the chance to develop the
> right relationships to facilitate inclusion (the process is
> non-trivial). However, I would like to point out that Nico Schlömer
> maintains a PPA for Ubuntu/Debian packages [1].

That might swing it. I do use some software only available that way.

> I would also like to point out that binary distributions are actually

> not ideal for packages like BLIS. ...

> I can understand your desire to stick with binary packages.

Debian is not binary-only. It has an equally useful mechanism for building one's own binaries via its package manager. It is easy, insofar as installing from source is ever easy, to ask Debian to download a source package (they usually have "-dev" in their names) with all its dependencies and build your own binary.

> Nonetheless, I invite you to try out a github clone. BLIS is very easy
> to compile and install for most users, and if you run into any trouble,
> we are here to help.

I may still do that, but we live in a world in which the software supermarket is well-stocked and even the software soup kitchen satisfies the needs of the poor. If I just want to try buy something to decide whether I like it, I don't want to clone a repo.

[dirk....@gmail.com]

Op dinsdag 28 augustus 2018 00:13:13 UTC+2 schreef Jeff Hammond:

> On Sun, Aug 26, 2018 at 11:22 PM, <dirk....@gmail.com> wrote:
> I've just completed the survey. The question that its designers implicitly ask
>

> To someone who no longer works on the bleeding edge of numerical linear algebra, if it isn't in Debian, it does not exist.

> I must ask: does Matlab exist in your universe?

No. My universe is a subset, not a superset, of free and open source software.

[dirk....@gmail.com]

Op maandag 27 augustus 2018 08:22:00 UTC+2 schreef dirk....@gmail.com:

> I've just completed the survey. The question that its designers implicitly ask is: What will induce you to use BLIS?
>
> The resolution of Issue 210 will in my case go a long way. https://github.com/flame/blis/issues/210
>
> To someone who no longer works on the bleeding edge of numerical linear algebra, if it isn't in Debian, it does not exist.

I have now partially followed Field's advice, and at least visited the GitHub page. I must compliment the BLIS team on the excellent README.md; it (and the draft TOMS paper it cites) answered all my questions.

[Jeff Hammond]

That's great to hear.  For users who might not want to build from source, would downloadable binaries help?  This isn't trivial but I think some projects are able to automate this from Travis CI.  At the very least, we could crowd source the manual effort, although this can lead to sadness when people don't maintain their effort forever (see e.g. LLVM binary support).

[Field G. Van Zee]

On 08/28/2018 12:02 AM, dirk....@gmail.com wrote:
> I have now partially followed Field's advice, and at least visited the GitHub page. I must compliment the BLIS team on the excellent README.md; it (and the draft TOMS paper it cites) answered all my questions.

Dirk,

Thanks for the kind words on the README.md. I've spent a lot of time on
writing, editing, and maintaining BLIS's documentation. If you are
interested in further reading, we have topic-specific markdown documents
in the 'docs' directory, which, if viewed via github, can be nicely
rendered by your web browser.

Field