ScaLAPACK Archives

[Scalapack] ScaLAPACK: PSLACP3 suggestion

Hello, excuse me for the late answer

I've tried PDGEADD and it does work nicely.

The main problems that are still present with PDGEADD:

1) it's not clear for user that it can be used this way (think notes in 
documentation can help)
2) the performance is worse than of P?LACP3 - for now I've estimated it only 
inside local 8-core node, see the attach with preliminary P?LACP4 (off course 
things will look different for IB cluster but you got the point)

So I'm still suggesting P?LACP4 in order to have fast distribution that can be 
often critical.
Any thoughts?


Best regards,
Alexander

-----Original Message-----
From: Julien Langou [mailto:julien.langou@Domain.Removed] 
Sent: Wednesday, April 21, 2010 4:54 AM
To: Ed D'Azevedo
Cc: Langou, Julien; scalapack@Domain.Removed; Kosenkov, Alex
Subject: Re: [Scalapack] ScaLAPACK: PSLACP3 suggestion


Hello,

I have experimented with PDGEADD and it seems to work great. Thanks Ed, 
for the tip.

Alex, do you agree that PDGEADD has the functionnality of "PSLACP4" ?
Do you still want to have a PSLACP4 routine in ScaLAPACK?

--julien

On Sat, 3 Apr 2010, Ed D'Azevedo wrote:

I think PDGEADD operates in the same context and PDGEMR2D operates
across different contexts.

I think PDGEADD has the capability of the proposed PSLACP4.


Julien Langou wrote:
Hello Alex, I think, as a general rule, we are all in favor of
collaboration and contribution. In particular anything that improves the
ease of use of ScaLAPACK is welcome. This said, your proposed PDLACP4
looks close to the PDGEMR2D routine in the REDIST directory. What is the
difference? (I just went over the routines quickly.) Best wishes, Julien.

On Wed, 31 Mar 2010, Kosenkov, Alex wrote:


Hello, (please forward as appropriate)



My name is Alexander and I?m working under ScaLAPACK Intel Math Kernel 
Library component.

During the time I was supporting this component I?ve seen many questions 
related to the matrix 2d-block cyclic distribution.

For this purpose Netlib ScaLAPACK has an auxiliary routine PSLACP3:



PSLACP3 is an auxiliary routine that  copies  from  a  global  parallel

array  into  a  local replicated array or vise versa. Notice that   the

entire submatrix that is copied gets placed on one node or more.

The receiving node can be specified precisely, or all nodes

can receive, or just one row or column of nodes.



The key problem that our users have encountered is that this function does 
support only square matrices:



SUBROUTINE PSLACP3( M, I, J, A, DESCA, B, LDB, II, JJ, REV )



Because of this MKL/Netlib users do distribution by themselves and It often 
leads to arbitrary mistakes which is not desirable.



We have decided to make an extended version of PSLACP3 that supports 
non-square matrices:



SUBROUTINE PSLACP4( M, N, I, J, A, DESCA, B, LDB, II, JJ, REV )



Generally we don?t want to introduce another difference that makes us less 
compatible with the Netlib in the future.

Therefore I want to ask you if there is a possibility to include this 
routine in the subsequent Netlib?s releases if we are to provide the source 
code of the already implemented routine (PSLACP4).



I would highly appreciate any comments and suggestions on this.



Thank you,

Alexander Kosenkov



--

Intel? Math Kernel Library | Solvers team

Software Engineer

--------------------------------------------------------------------
Closed Joint Stock Company Intel A/O
Registered legal address: Krylatsky Hills Business Park,
17 Krylatskaya Str., Bldg 4, Moscow 121614,
Russian Federation

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.




_______________________________________________
Scalapack mailing list
Scalapack@Domain.Removed
http://lists.eecs.utk.edu/mailman/listinfo/scalapack


--------------------------------------------------------------------
Closed Joint Stock Company Intel A/O
Registered legal address: Krylatsky Hills Business Park, 
17 Krylatskaya Str., Bldg 4, Moscow 121614, 
Russian Federation

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: psgeadd_vs_pslacp4.jpg
Type: image/jpeg
Size: 83935 bytes
Desc: psgeadd_vs_pslacp4.jpg
Url : 
http://lists.eecs.utk.edu/mailman/private/scalapack/attachments/20100430/5224a817/psgeadd_vs_pslacp4-0001.jpg
 

<Prev in Thread] Current Thread [Next in Thread>


For additional information you may use the LAPACK/ScaLAPACK Forum.
Or one of the mailing lists, or