Indeed, some of these operations do allocate additional data structures.
Other problems were (1) that our memory estimates do not account for the
explicit copy into commons math data structures (e.g.,
Array2DRowRealMatrix), and (2) unnecessarily raised exceptions due to
unknowns. However, both issues can be addressed and since we're talking
about warnings, false positives/negatives are probably ok.
Regards,
Matthias
On 10/24/2016 9:35 PM, Berthold Reinwald wrote:
if I remember correctly then it is not trivial to accurately estimate the
memory foot print for these commons math functions at compile time
depending on what intermediates they produce ... Meaning you may still end
up with java heap space OOM at runtime.
Regards,
Berthold Reinwald
IBM Almaden Research Center
office: (408) 927 2208; T/L: 457 2208
e-mail: reinw...@us.ibm.com
From: Matthias Boehm <mboe...@googlemail.com>
To: dev@systemml.incubator.apache.org
Date: 10/24/2016 11:54 AM
Subject: Re: Local versions of Linear Algebra Operators in DML
well, we still compute memory estimates for these operations. So I
guess, a good compromise would be to raise a warning whenever the memory
estimate is known to exceed the local memory budget.
Regards,
Matthias
On 10/24/2016 8:29 PM, Deron Eriksson wrote:
Would it be acceptable for a user to receive a log warning if the user
uses
an operation that is currently only implemented for single node? My
concern
is that there is an expectation for operations to be distributed with
SystemML, and if an operation is not currently distributed, the user
needs
to made aware of this.
Thoughts?
Deron
On Mon, Oct 24, 2016 at 10:38 AM, Nakul Jindal <naku...@gmail.com>
wrote:
Hi,
There is an initial implementation and PR.
https://github.com/apache/incubator-systemml/pull/273
-Nakul
On Oct 24, 2016, at 12:59 AM, Berthold Reinwald <reinw...@us.ibm.com>
wrote:
Thanks, Imran. I think it is a good idea to start off with the
DML-bodied
function implementation. This will hold until we can have a built in
implementation.
We prototyped an implementation of distributed Cholesky as a DML
bodied
function as well. For performance optimization, as the matrix becomes
"small" enough, we switched over and exploit a single node
implementation.
Adding a new svd() built in function that initially routes to a local
library is fine. I don't know whether Apache commons math has an
implementation that can be re-used.
I object renaming the functions or changing the externals. Eventually
distributed instructions need to be added to these implementations,
and
there are open jiras for it.
Regards,
Berthold Reinwald
IBM Almaden Research Center
office: (408) 927 2208; T/L: 457 2208
e-mail: reinw...@us.ibm.com
From: Niketan Pansare/Almaden/IBM@IBMUS
To: dev@systemml.incubator.apache.org
Date: 10/21/2016 01:14 PM
Subject: Re: Local versions of Linear Algebra Operators in DML
I am also comfortable with option (2) ... "with a plan to implement
its
distributed version"
Thanks,
Niketan Pansare
IBM Almaden Research Center
E-mail: npansar At us.ibm.com
http://researcher.watson.ibm.com/researcher/view.php?person=us-npansar
Matthias Boehm ---10/21/2016 01:00:51 PM---thanks Nakul for reaching
out
before starting work on this. Actually, the introduction of these CP-
From: Matthias Boehm <mboe...@googlemail.com>
To: dev@systemml.incubator.apache.org
Date: 10/21/2016 01:00 PM
Subject: Re: Local versions of Linear Algebra Operators in DML
thanks Nakul for reaching out before starting work on this. Actually,
the introduction of these CP-only builtin functions was a big mistake
because (as you already mentioned) they mistakenly suggest that we
provide distributed operations for them too. The intend was to support
them in later versions with our own local and distributed
implementations. So far, this had low priority though because these
O(n^3) operations are seldom used over large data. However, a while
back, we lost potential users who were specifically interested in
distributed eigen - so there are still use cases.
Despite the good intentions behind the renaming, I would strongly
argue
against it. First, it would unnecessarily lose compatibility with R
syntax. Second, it would defeat our clean abstraction by exposing
explicit local operations.
This leaves us with two options here: (1) you could use an external
(java-implemented) function, which gives you virtually the same
runtime
behavior but a clear separation via an explicit registration, or (2)
add
it to the list of CP-only operations (with a plan to implement its
distributed version) but name it 'svd' as in R.
Regards,
Matthias
On 10/21/2016 9:34 PM, Nakul Jindal wrote:
Hi,
Imran was planning on implementing a distributed SVD as a DML bodied
function.
The algorithm is described in the paper titled "A Distributed and
Incremental SVD Algorithm for Agglomerative Data Analysis on Large
Networks" available at https://arxiv.org/abs/1601.07010.
This algorithm requires the availability of a local SVD function,
which
we
currently do not have in SystemML.
Seeing as how there are other linear algebra functions (eigen, lu,
qr,
cholesky) in DML that reroute to Apache Common Math and only operate
in
standalone/CP mode, would it be ok to add "svd" to this set?
Also, since these operations are local and not distributed and the
documentation doesn't make it clear that these operations wont
operate
in
distributed mode, would it make sense to rename them to
"local_eigen",
"local_qr", "local_cholesky", etc?
Obviously, this change would go into the version after 0.11.
I understand that the ideal solution to this problem is to have a
distributed version of the aforementioned linear algebra routines,
but
for
the time being, would it be ok to go ahead do the rename, while also
introducing a "local_svd" ?
Niketan, Berthold, Matthias, Sasha - Any thoughts?
Thanks,
Nakul Jindal
