Hey John,
Also : for dynamic load balancing, I'd like to instruct several reader to read
the same piece - since the algorithm controls (for example) the particles the
algorithm can internally communicate information about what to do amongst its
processes, but it can't talk upstream to the readers and fudge them.
I am wondering if there is any way of supporting this kind of thing using the
current information keys and my instinct says no.
I guess you can kind of do this with the current "request update" stuff
but thanks to the flexibility of the pipeline information key,values you
can also roll your own very easy.
I recently implemented dynamic load balancing in a new stream line
tracer. To get the work load balanced its crucial that each process have
to have on demand access to the entire data set. I accomplished it with
information keys and by using a "meta-reader" in place of the
traditional paraview reader. The meta reader does two things, it
populates the new keys and it gives PV a dummy dataset that is one cell
per process such that the bounds, shape, and array names are the same as
the real dataset which is not read during the meta-reader execution.
When the stream tracer executes downstream of the meta-reader he picks
the keys out of the pipeline information. The important key,value is an
out-of-core (ooc) reader. The ooc reader is a vtkDataObject so that it
can be passed through the information. Once the stream tracer has it he
can make repeated IO requests as particles move through the dataset as
needed. My interface accepts a point and returns a chunk of data. The
ooc reader internally handles caching and memory management. In this way
you can keep all processes busy all the time when tracing stream lines.
The approach worked out well and was very simple to implement, with no
modification to the executive. Also the filter has control of caching,
and can free all the memory at the end of its execution which reduces
significantly the memory footprint compared to the traditional PV
reader. And I need not worry if PV or some upstream filter uses MPI
communications in between during my IO requests. There is a little more
to our scheduling algorithm which I wont discus now but so far for
making poincare maps we scaled well up to 2E7 stream lines per frame and
96 processes and we minimize the memory footprint which is important to us.
Berk and Ken already basically gave you all the options you need but I
add this because it shows how flexible and powerful the pipeline
information really is.
Burlen
Biddiscombe, John A. wrote:
Berk,
We had a discussion back in 2008, which resides here
http://www.cmake.org/pipermail/paraview/2008-May/008170.html
Continuing from this, my question of the other day, touches on the same problem.
I'd like to manipulate the piece number read by each reader. As mentioned
before, UPDATE_PIECE is not passed into RequestInformation at first (since
nobody knows how many pieces there are yet!), so I can't (directly) generate
information in the reader which is 'piece dependent'. And I can't be sure that
someone doing streaming won't interfere with piece numbers when using the code
differently.
For the particle tracer (for example), I'd like to tell the upstream pipeline to read no
pieces when certain processes are empty of particles (currently they update and
generate{=read} data when they don't need to). I may be able to suppress the forward
upstream somehow, but I don't know of an easy way for the algorithm to say
"Stop" to the executive to prevent it updating if the timestep changes, but the
algorithm has determined that no processing is required (ForwardUpstream of Requests
continues unabated). I'd like to set the UPdatePiece to -1 to tell the executive to stop
operating.
Also : for dynamic load balancing, I'd like to instruct several reader to read
the same piece - since the algorithm controls (for example) the particles the
algorithm can internally communicate information about what to do amongst its
processes, but it can't talk upstream to the readers and fudge them.
I am wondering if there is any way of supporting this kind of thing using the
current information keys and my instinct says no. It seems like the update pice
and numpieces were really intended for streaming and we need two kinds of
'pieces', one for streaming, another for splitting in _parallel_ because they
aren't quite the same. (Please note that I haven't actually tried changing
piece requests in the algorithms yet, so I'm only guessing that it won't work
properly)
<cough>
UPDATE_STREAM_PIECE
UPDATE_PARALLEL_PIECE
<\cough>
Comments?
JB
I would have the reader (most parallel readers do this) generate empty
data on all processes of id >= N. Then your filter can redistribute
from those N processes to all M processes. I am pretty sure
RedistributePolyData can do this for polydata as long as you set the
weight to 1 on all processes. Ditto for D3.
-berk
On Fri, Dec 11, 2009 at 4:13 PM, Biddiscombe, John A. <biddi...@cscs.ch>
wrote:
Berk
It sounds like M is equal to the number of processors (pipelines) and
M >> N. Is that correct?
Yes, That's the idea. N blocks, broken (in place) into M new blocks, then
fanned out to the M processes downstream where they can be processed
separately . If it were on a single node, then each block could be a
separate 'connection' to a downstream filter, but distributed, an explicit
send is needed.
JB
-berk
On Fri, Dec 11, 2009 at 10:40 AM, Biddiscombe, John A. <biddi...@cscs.ch>
wrote:
Berk
The data will be UnstructuredGrid for now. Multiblock, but actually, I
don't really care what each block is, only that I accept one block on
each
of N processes, split it into more pieces, and the next filter accepts
one
(or more if the numbers don't match up nicely) blocks and process them.
The
redistribution shouldn't care what data types, only how many blocks in
and
out.
Looking at RedistributePolyData makes me realize my initial idea is no
good. In my mind I had a pipeline where multiblock datasets are passed
down
the pipeline and simply the number of pieces is manipulated to achieve
what
I wanted - but I see now that if I have M pieces downstream mapped
upstream
to N pieces, what will happen is the readers will be effectively
duplicated
and M/N readers will read the same pieces. I don't want this to happen as
IO
will be a big problem if readers read the same blocks M/N times.
I was hoping there was a way of simply instructing the pipeline to
manage
the pieces, but I see now that this won't work, as there needs to be a
specific Send from each N to their M/N receivers (because the data is
physically in another process, so the pipeline can't see it). This is
very
annoying as there must be a class which already does this (block
redistribution, rather than polygon level redistribution), and I would
like
it to be more 'pipeline integrated' so that the user doesn't have to
explicitly send each time an algorithm needs it.
I'll go through RedistributePolyData in depth and see what I can pull
out
of it - please feel free to steer me towards another possibility :)
JB
-----Original Message-----
From: Berk Geveci [mailto:berk.gev...@kitware.com]
Sent: 11 December 2009 16:09
To: Biddiscombe, John A.
Cc: paraview@paraview.org
Subject: Re: [Paraview] Parallel Data Redistribution
What is the data type? vtkRedistributePolyData and its subclasses do
this for polydata. It can do load balancing (where you can specify a
weight for each processor) as well.
-berk
On Fri, Dec 11, 2009 at 9:59 AM, Biddiscombe, John A.
<biddi...@cscs.ch>
wrote:
I have a filter pipeline which reads N blocks from disk, this works
fine
on N processors.
I now wish to subdivide those N blocks (using a custom filter) to
produce
new data which will consist of M blocks - where M >> N.
I wish to run the algorithm on M processors and have the piece
information
transformed between the two filters (reader -> splitter), so that
blocks
are
distributed correctly. The reader will Read N blocks (leaving M-N
processes
unoccupied), but the filter which splits them up needs to output a
different
number of pieces and have the full M processes receiving data.
I have a reasonably good idea of how to implement this, but I'm
wondering
if any filters already do something similar. I will of course take
apart
the
D3 filter for ideas, but I don't need to do a parallel spatial
decomposition
since my blocks are already discrete - I just want to redistribute the
blocks around and more importantly change the numbers of them between
filters.
If anyone can suggest examples which do this already, please do
Thanks
JB
--
John Biddiscombe, email:biddisco @
cscs.ch
http://www.cscs.ch/
CSCS, Swiss National Supercomputing Centre | Tel: +41 (91)
610.82.07
Via Cantonale, 6928 Manno, Switzerland | Fax: +41 (91)
610.82.82
_______________________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
Please keep messages on-topic and check the ParaView Wiki at:
http://paraview.org/Wiki/ParaView
Follow this link to subscribe/unsubscribe:
http://www.paraview.org/mailman/listinfo/paraview
_______________________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
Please keep messages on-topic and check the ParaView Wiki at:
http://paraview.org/Wiki/ParaView
Follow this link to subscribe/unsubscribe:
http://www.paraview.org/mailman/listinfo/paraview
_______________________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
Please keep messages on-topic and check the ParaView Wiki at:
http://paraview.org/Wiki/ParaView
Follow this link to subscribe/unsubscribe:
http://www.paraview.org/mailman/listinfo/paraview
