commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From S├ębastien Brisard <sebastien.bris...@m4x.org>
Subject Re: [Math] Moving on or not?
Date Fri, 08 Feb 2013 19:15:15 GMT
Large does not necessarily mean sparse. And please note that iterative
linear solvers was I think a big step towards large scale problems.

Of course, you are very welcome to contribute a robust implementation for
sparse matrices. The reason why we dropped it (momentarily, hopefully) was
that the current implementation was broken, in the sense that the same
computation carried out with sparse and non sparse implementations would
not lead to the same result. While this might seem perfectly acceptable (it
probably is in most applications), people who took part to the
corresponding discussion (you were not among those!) decided that it was
*not* acceptable.

I've spent a lot of time on this issue, and could not come up with a
satisfactory solution. I think the best we could do is
1. Restore the broken implementation, clean it up, and put a lot of
warnings in the javadoc.
2. Create a sub-project for sparse implementations, with lower expectations
regarding boundary cases (1.0 / +0.0, 1.0 / -0.0, ...).

Since people are reluctant with Gilles' proposal, I think that 1. is
probably the best solution, but the javadoc must be very clear about all
possible issues.

As a side note, if I remember correctly, people complained about the
implementation of sparse matrices not being very efficient.
S


2013/2/8 Konstantin Berlin <kberlin@gmail.com>

> Sorry, but not of this is making sense to me. We had a long discussion
> about how the library doesn't test for large scale problem
> performance. A lot of algorithms probably do not scale well as the
> result. There was talk of dropping sparse support in linear algebra.
> So instead of fixing that, you jump to parallelization, which is
> needed only for large scale problems, which this library does not
> handle well even in single thread right now.
>
> The most significant impact you can have is fixing the linear algebra
> component.
>
> On Feb 7, 2013, at 5:06 PM, Gilles <gilles@harfang.homelinux.org> wrote:
>
> > On Thu, 07 Feb 2013 08:32:46 -0800, Phil Steitz wrote:
> >> On 2/7/13 8:04 AM, Gilles wrote:
> >>> On Thu, 07 Feb 2013 07:01:42 -0800, Phil Steitz wrote:
> >>>> On 2/7/13 4:58 AM, Gilles wrote:
> >>>>> On Wed, 06 Feb 2013 09:46:55 -0800, Phil Steitz wrote:
> >>>>>> On 2/6/13 9:03 AM, Gilles wrote:
> >>>>>>> On Wed, 06 Feb 2013 07:19:47 -0800, Phil Steitz wrote:
> >>>>>>>> On 2/5/13 6:08 AM, Gilles wrote:
> >>>>>>>>> Hi.
> >>>>>>>>>
> >>>>>>>>> In the thread about "static import", Stephen noted
that
> >>>>>>>>> decisions
> >>>>>>>>> on a
> >>>>>>>>> component's evolution are dependent on whether the
future of
> >>>>>>>>> the
> >>>>>>>>> Java
> >>>>>>>>> language is taken into account, or not.
> >>>>>>>>> A question on the same theme also arose after the
> >>>>>>>>> presentation of
> >>>>>>>>> Commons
> >>>>>>>>> Math in FOSDEM 2013.
> >>>>>>>>>
> >>>>>>>>> If we assume that efficiency is among the important
> >>>>>>>>> qualities for
> >>>>>>>>> Commons
> >>>>>>>>> Math, the future is to allow usage of the tools
provided by the
> >>>>>>>>> standard
> >>>>>>>>> Java library in order to ease the development of
multi-threaded
> >>>>>>>>> algorithms.
> >>>>>>>>>
> >>>>>>>>> Maintaining Java 1.5 source compatibility for the
reason
> >>>>>>>>> that we
> >>>>>>>>> may need
> >>>>>>>>> to support legacy applications will turn out to
be
> >>>>>>>>> self-defeating:
> >>>>>>>>> 1. New users will not consider Commons Math's features
that are
> >>>>>>>>> notably
> >>>>>>>>>   apt to parallel processing.
> >>>>>>>>> 2. Current users might at some point simply switch
to another
> >>>>>>>>> library if
> >>>>>>>>>   it proves more efficient (because it actually
uses
> >>>>>>>>> multi-threading).
> >>>>>>>>> 3. New Java developers will be turned away because
they will
> >>>>>>>>> want
> >>>>>>>>> to use
> >>>>>>>>>   the more convenient features of the language in
order to
> >>>>>>>>> provide
> >>>>>>>>>   potential contributions.
> >>>>>>>>>
> >>>>>>>>> If maintaining 1.5 source compatibility is kept
as a
> >>>>>>>>> requirement, the
> >>>>>>>>> consequence is that Commons Math will _become_ a
legacy
> >>>>>>>>> library.
> >>>>>>>>> In that perspective, implementing/improving algorithms
for
> >>>>>>>>> which a
> >>>>>>>>> parallel version is known to be more efficient is
plainly a
> >>>>>>>>> waste of
> >>>>>>>>> development and maintenance time.
> >>>>>>>>>
> >>>>>>>>> In order to mitigate the risks (both of upgrading
and of not
> >>>>>>>>> upgrading
> >>>>>>>>> the source compatibility requirement), I would propose
to
> >>>>>>>>> create a
> >>>>>>>>> new
> >>>>>>>>> project (say, "Commons Math MT") where we could
implement new
> >>>>>>>>> features[1]
> >>>>>>>>> without being encumbered with the 1.5 requirement.[2]
> >>>>>>>>> The "Commons Math MT" would depend on "Commons Math"
where we
> >>>>>>>>> would
> >>>>>>>>> continue developing single-thread (and thread-safe)
"tasks",
> >>>>>>>>> i.e.
> >>>>>>>>> independent units of processing that could be used
in
> >>>>>>>>> algorithms
> >>>>>>>>> located in "Commons Math MT".
> >>>>>>>>>
> >>>>>>>>> In summary:
> >>>>>>>>> - Commons Math (as usual):
> >>>>>>>>>  * single-thread (sequential) algorithms,
> >>>>>>>>>  * (pure) Java 5,
> >>>>>>>>>  * no dependencies.
> >>>>>>>>> - Commons Math MT:
> >>>>>>>>>  * multi-thread (parallel) algorithms,
> >>>>>>>>>  * Java 7 and beyond,
> >>>>>>>>>  * JNI allowed,
> >>>>>>>>>  * dependencies allowed (jCuda).
> >>>>>>>>>
> >>>>>>>>> What do you think?
> >>>>>>>>
> >>>>>>>> There are several other possibilities to consider:
> >>>>>>>>
> >>>>>>>> 0) Implement multithreading using JDK 1.5 primitives
> >>>>>>>> 1) Set things up within [math] to support parallel execution
in
> >>>>>>>> JDK
> >>>>>>>> 1.7, Hadoop or other frameworks
> >>>>>>>> 2) Instead of a new project, start a 4.x branch targeting
JDK
> >>>>>>>> 1.7
> >>>>>>>>
> >>>>>>>> I think we should maintain a version that has no dependencies
> >>>>>>>> and no
> >>>>>>>> JNI in any case.
> >>>>>>>>
> >>>>>>>> Starting a branch and getting concrete about how to
parallelize
> >>>>>>>> some
> >>>>>>>> algorithms would be a good way to start.  One thing
I have not
> >>>>>>>> really investigated and would be interested in details
on is
> >>>>>>>> what
> >>>>>>>> you actually get in efficiency gain (or loss?) using
fork /
> >>>>>>>> join vs
> >>>>>>>> just using 1.5+ concurrency for the kinds of problems
we
> >>>>>>>> would end
> >>>>>>>> up using this stuff for.
> >>>>>>>>
> >>>>>>>> Thinking about specific parallelization problem instances
would
> >>>>>>>> also
> >>>>>>>> help decide whether 1) makes sense (i.e., whether it
makes
> >>>>>>>> sense as
> >>>>>>>> you mention above to maintain a single-threaded library
that
> >>>>>>>> provides task execution for a multithreaded version
or
> >>>>>>>> multithreaded
> >>>>>>>> frameworks).
> >>>>>>>>
> >>>>>>>> One more thing to consider is that for at least some
users of
> >>>>>>>> [math], having the library internally spawn threads
and/or peg
> >>>>>>>> multiple processors may not be desirable.  It is a little
> >>>>>>>> misleading
> >>>>>>>> to say that multithreading is the way to get "efficiency."
> >>>>>>>> It is
> >>>>>>>> really the way to *use* more compute resources and unless
there
> >>>>>>>> are
> >>>>>>>> real algorithmic improvements, the overall efficiency
may
> >>>>>>>> actually
> >>>>>>>> be less, due to task coordination overhead.  What you
get is
> >>>>>>>> faster
> >>>>>>>> execution due to more greedy utilization of available
cores.
> >>>>>>>> Actual
> >>>>>>>> efficiency (how much overall compute resource it takes
to
> >>>>>>>> complete a
> >>>>>>>> job) partly depends on how efficiently the coordination
> >>>>>>>> itself is
> >>>>>>>> done (which JDK 1.7 claims to do very well - I have
just not
> >>>>>>>> seen
> >>>>>>>> substantiation or any benchmarks demonstrating this)
and how the
> >>>>>>>> parallelization effects overall compute requirements.
 In any
> >>>>>>>> case,
> >>>>>>>> for environments where library thread-spawning is not
> >>>>>>>> desirable, I
> >>>>>>>> think we should maintain a single-threaded version.
> >>>>>>>
> >>>>>>> Unless I missed the point, those reasons are exactly why
I
> >>>>>>> propose to
> >>>>>>> have 2 projects/components. One, "Commons-Math", does not
fiddle
> >>>>>>> with
> >>>>>>> resources, while the other would provide a "parallelizationLevel"
> >>>>>>> setting for the algorithms written to possibly take advantage
of
> >>>>>>> the
> >>>>>>> Java 5+ "task framework".
> >>>>>>
> >>>>>> OK, what about the 4.x option?
> >>>>>>>
> >>>>>>> Yes, we could still be good by using only Java 5's concurrency
> >>>>>>> features
> >>>>>>> but the issue I raise is not only about concurrency but
about
> >>>>>>> evolution/progress/maintenance, all things that require
raising
> >>>>>>> interest
> >>>>>>> from new contributors (unless it's fine that Commons Math
be
> >>>>>>> tagged as a
> >>>>>>> "library of the past"...).
> >>>>>>
> >>>>>> +1 for experimenting with parallelization.  I would just like
to
> >>>>>> understand if the JDK 7 stuff really adds much - in particular,
> >>>>>> does
> >>>>>> it handle coordination / cpu allocation better than you could
> >>>>>> easily
> >>>>>> do it with 1.5.  More supported JDKs == more potential users,
so I
> >>>>>> like to see a real reason to bump the JDK level.
> >>>>>>>
> >>>>>>> But using concurrency features in "Commons Math" would also
> >>>>>>> contradict
> >>>>>>> your own point ("we should maintain a single-threaded
> >>>>>>> version"): I
> >>>>>>> agree,
> >>>>>>> and that's why I proposed this other project...
> >>>>>>>
> >>>>>>> As for efficiency (or faster execution, if you want), I
don't
> >>>>>>> see the
> >>>>>>> point in doubting that tasks like global search (e.g. in
a
> >>>>>>> genetic
> >>>>>>> algorithm) will complete in less time when run in parallel...
> >>>>>>>
> >>>>>>> As I summarized previously, having a "Commons Math MT" would
> >>>>>>> bring no
> >>>>>>> inconvenience, contrary to either your points 0, 1, or 2.
[No
> >>>>>>> inconvenience to me, that is, but to people with requirements
> >>>>>>> like
> >>>>>>> "Java 5 compatible" or "no multi-threading").
> >>>>>>> As I indicated, the basic "task" could be defined in "Commons
> >>>>>>> Math" and
> >>>>>>> "Commons Math MT" would provide the parallelization "glue"
(e.g.
> >>>>>>> to divide
> >>>>>>> the search space of the GA).
> >>>>>>
> >>>>>> I think it is best at this point to cut a branch and actually
> >>>>>> start
> >>>>>> working on specific algorithms.  Having a set of candidate
> >>>>>> algorithms for parallelization will help us decide what we
> >>>>>> actually
> >>>>>> need and how it might work.  I would personally favor the 4.x
> >>>>>> approach, with thread-spawning behavior configurable.
> >>>>>
> >>>>> It seems fair to wait until parallel algorithms are actually
> >>>>> implemented.
> >>>>>
> >>>>> However it is not clear what you mean with "the 4.x approach": if
> >>>>> it is
> >>>>> actually allowing Java 7, that would mean that, starting from 4.0,
> >>>>> we'll
> >>>>> indeed drop support of earlier JVMs!
> >>>>> Why would this be preferred to having 2 projects? Of course, if
> >>>>> everyone
> >>>>> agrees to that move to Java 7, that's fine. :-)
> >>>>
> >>>> What I meant was that instead of creating a new component, we would
> >>>> just create a new release line.  Like what tomcat does for servlet
> >>>> spec versions.  I guess this does mean that we end up having to
> >>>> stabilize the 3.x APIs because no additional "major" release would
> >>>> be allowed in that line.  That would be a *good thing* IMO as long
> >>>> as we can do it cleanly.  If not, maybe we end up having to use 5.x
> >>>> for the JDK 1.7+ version, using 4.0 to get to a stable API for the
> >>>> current trunk code.
> >>>
> >>> There's a still the human resource problem: we don't have it to
> >>> maintain
> >>> a single branch; having two will only make it worse.
> >>
> >> Yes, but the "new project" approach has the same problem.
> >
> > Yes.
> > However, I meant it as a way to separate concerns, as shown
> > by diverging opinions, even in the few people who take part
> > in this discussion or in previous ones about the same subject.
> >
> > A sibling (not separate!) project could allow interested
> > people to experiment while not adding yet another "distraction"
> > to the main project, where people more focused on the
> > mathematical (for lack of a better word) side can continue
> > their own improvements.
> > A healthy interaction could even come out of having a "public"
> > use-case in the form of a project that needs certain facilities
> > (algorithms as tasks) in order to provide multi-thread
> > utilities to users (who might prefer not to have to implement
> > them themselves at a higher level).
> >
> >>>>> On the other hand, if we keep Java 5, at least until we get use
> >>>>> cases or
> >>>>> contributions that would benefit from features in JDKs newer than
> >>>>> 1.5,
> >>>>> there is no need to create a branch; we can just go on with adding
> >>>>> multi-thread codes to the trunk (to become part[1] of the upcoming
> >>>>> 3.x
> >>>>> releases).
> >>>>
> >>>> That is why I wanted to get a feel for what the JDK 1.7 stuff really
> >>>> buys you.   Has anyone seen benchmarks showing better performance
> >>>> using 1.7 than can be obtained just using 1.5 concurrency
> >>>> primitives?
> >>>
> >>> Again, there are separate issues:
> >>> 1. Coding in Java 7
> >>> 2. Running with the JVM shipped with JDK 1.7
> >>>
> >>> The newer JVMs are faster, independently of whether new features
> >>> of the
> >>> language are used.
> >>> But it could well be that some of the new features allow even better
> >>> performance (as is foreseen for Java 8).
> >>
> >> Agreed.  I am interested in understanding better both how much
> >> easier it actually is to code and whether the 1.7 framework
> >> materially improves scheduling / allocation over what you could do
> >> just using 1.5 primitives.
> >
> > I cannot provide proof, but nor is anyone on this list
> > eager to prove the contrary; hence the proposal to set
> > up a "playground".
> >
> >>>> Has anyone used 1.7 to parallelize numerical algorithms
> >>>> and found it really easier / more performant?
> >>>
> >>> Where are those people who could answer?
> >>
> >> This is a public list :)
> >>> That is one of the points I raised. If we maintain source
> >>> compatibility
> >>> with a language version that is 9 years old, not many contributors
> >>> are
> >>> going to be interested. Thus reducing the chance to get answers...
> >>>
> >>>> Any opinions /
> >>>> responses to Konstantin's comment on where parallelization should be
> >>>> implemented - i.e. in the library vs somewhere up the stack?
> >>>
> >>> What was the _question_?  ...
> >>
> >> The question he implicitly raised was whether or not it makes sense
> >> for a low-level library to parallelize tasks / run across cores.
> >
> > In several areas, CM is not a low-level library (GA, multi-start
> > optimizers for example). In other areas like FFT, a user can
> > legitimately expect top performance without having to handle
> > parallelization by himself.
> >
> >> This is a legitimate question.  It may be better actually to set
> >> things up so that higher-level frameworks or applications can
> >> arrange parallel execution rather than embedding it in the low-level
> >> library itself.  This is also what I was referring to when I said
> >> that in some contexts, thread-spawning / cpu hogging may not be
> >> desirable.
> >
> > For several cases (GA, FFT, multi-start optimizers), I have the
> > opposite viewpoint: multi-threading is a implementation detail,
> > that could be handled at a _lower_ level. Of course, the user can
> > decide whether to enable more than one thread.
> >
> >>>> Any
> >>>> ideas how to set things up so that [math] code can play nicely with
> >>>> concurrency frameworks?
> >>>
> >>> That's a strange question in the context of a project that tries hard
> >>> not to have any dependency.
> >>
> >> I did not mean necessarily to bring in dependencies; but rather to
> >> make it easy for computational tasks executed by [math] code to be
> >> managed by external concurrency frameworks, e.g. Hadoop.
> >
> > In the context of Commons Math, we often heard that "no dependency"
> > is good. Then, it is also good to not impose _implicit_ dependencies
> > (like: "If you use Hadoop, you could have better performance"). In a
> > way, the CM development "model" is: "We provide a toolkit of efficient
> > procedures, and you, the user, get top performance (on a best effort
> > basis of course)."
> > If we can provide better performance through multi-threading, why not?
> > Nobody will be forced to use it: they will use the "basic" (sequential)
> > tasks, or set the "parallelizationLevel" setting to 1.
> >
> > Gilles
> >
> >> Phil
> >>> If the requirement is to only depend on the standard JDK: the
> >>> framework
> >>> is in
> >>> java.util.concurrent
> >>> and all we need to do is to define "tasks" that can be "submitted to
> >>> an executor:
> >>>
> >>>
> http://docs.oracle.com/javase/1.5.0/docs/api/java/util/concurrent/AbstractExecutorService.html#submit(java.util.concurrent.Callable)
> >>>
> >>>
> >>> Regards,
> >>> Gilles
> >
> >
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
> > For additional commands, e-mail: dev-help@commons.apache.org
> >
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
> For additional commands, e-mail: dev-help@commons.apache.org
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message