[Paraview] Rendering in parallel

Thu Jul 3 06:05:16 EDT 2014

Actually, I don't use mpirun to launch pvserver but mpiexec in
ParaView-4.1.0-RC1-Linux-64bit/lib/paraview-4.1/ and it is why the option
is not recognized.

Jérémy

2014-07-03 10:20 GMT+02:00 Jérémy Santina <jeremy.santina at gmail.com>:

> Thank you very much for your help.
>
> As you said, my DISPLAY environment variable is pointing back to my
> desktop.
> The problem now is that I have tried using the option -display for the
> command pvserver but apparently it is unknown. Is it normal ? Is there
> another way to make sure that the display is set correctly ? I didn't
> mention it but my version of Paraview is 4.1.0-RC1-Linux-64bits. I don't
> know if it might help.
>
> Jérémy.
>
>
> 2014-07-02 22:34 GMT+02:00 Moreland, Kenneth <kmorel at sandia.gov>:
>
>  OK, I can see where things are going majorly wrong here. Let's start
>> with the worst of the problems.
>>
>>  I notice on the bottom of your screenshot that your desktop has 4
>> windows named ParaView Server #0, ParaView Server #1, etc. Those are X
>> windows that the server is opening up on your desktop. You really don't
>> want the server to do that. Those windows are used for OpenGL rendering. If
>> they are opened on your desktop, that means that all four of those
>> processes on your server are sending *all* the geometry to your desktop,
>> your desktop renders *all* the geometry, and then the images get shipped
>> to the server. The server then composites those images together and sends
>> the result *back* to your desktop.
>>
>>  I'm sure that when you are running the server, your DISPLAY environment
>> variable is pointing back to your desktop, which is causing the problem.
>> You need to make sure the server is run with display set to localhost:0.
>> More information is on the ParaView wiki at:
>>
>>  http://www.paraview.org/Wiki/Setting_up_a_ParaView_Server#X_Connections
>>
>>  That said, I'm not sure using your server is going to give you a big
>> rendering performance boost over your desktop. The parallel rendering is
>> really designed for large clusters with many GPUs. The rendering should
>> work OK on your desktop as long as you're not thrashing your virtual memory
>> (which is possible).
>>
>>  -Ken
>>
>>   From: Jérémy Santina <jeremy.santina at gmail.com>
>> Date: Wednesday, July 2, 2014 4:17 AM
>>
>> To: Kenneth Moreland <kmorel at sandia.gov>
>> Cc: "paraview at paraview.org" <paraview at paraview.org>
>> Subject: [EXTERNAL] Re: [Paraview] Rendering in parallel
>>
>>   Sorry for my poor description. I will try to give more information.
>>
>> I am loading a Multi-block Dataset without applying any filters and the
>> rendering is surface rendering. In order to understand how it works, I am
>> just running a pvserver in parallel on another computer (with a better GPU)
>> connected via SSH. The graphics card is an NVIDIA Quadro FX 4600 and you
>> have to know that I am not alone using this machine.  Server and client
>> both work on Linux. So would the problem be because there is only one GPU ?
>>
>>  I join a picture with this message.
>>
>>  I would have another question. When I launch the rendering in parallel,
>> a variable called vtkProcessId is generated. What is it ? Does it do the
>> same thing if I apply Process Id Scalars filter ? Or are they two different
>> things ?
>>
>>  Jérémy
>>
>>
>> 2014-07-01 18:08 GMT+02:00 Moreland, Kenneth <kmorel at sandia.gov>:
>>
>>>  To check the distribution of the data, use the Process Id Scalars
>>> filters. That should color the data based on which processor it is located.
>>>
>>>  It might help if you described your system more completely. What kind
>>> of data are you loading? Is it image data? Polygon data? AMR? An
>>> unstructured grid? Are you applying any filters? How are you rendering it?
>>> Is it surface or volume rendering? Is there any transparency? Can you send
>>> a picture? What kind of parallel computer are you using? Are you running
>>> ParaView on your desktop in multi-core mode (I think rendering actually
>>> serializes in that case because you still have only one GPU.), or are you
>>> connecting to a cluster? How many nodes on your cluster and how are they
>>> configured?
>>>
>>>  -Ken
>>>
>>>   From: Jérémy Santina <jeremy.santina at gmail.com>
>>> Date: Tuesday, July 1, 2014 2:31 AM
>>> To: Kenneth Moreland <kmorel at sandia.gov>
>>> Cc: "paraview at paraview.org" <paraview at paraview.org>
>>> Subject: [EXTERNAL] Re: [Paraview] Rendering in parallel
>>>
>>>   Actually, I did try the D3 filter but I didn't really see any better
>>> results. Maybe it is because I don't know how to configure it. How does D3
>>> filter work ?
>>>
>>>
>>>
>>> 2014-06-30 16:21 GMT+02:00 Moreland, Kenneth <kmorel at sandia.gov>:
>>>
>>>>   Jeremy,
>>>>
>>>>  Like the other parallel processing in ParaView, the efficiency is
>>>> dictated by the distribution of the data. If your data distribution is
>>>> highly imbalanced such as when all the data is on one process as in your
>>>> case, then all the processing will happen where the data is and the rest of
>>>> the processors will remain idle.
>>>>
>>>>  You could try running the D3 filter. That should redistribute the
>>>> point data more evenly.
>>>>
>>>>  -Ken
>>>>
>>>>   From: Jérémy Santina <jeremy.santina at gmail.com>
>>>> Date: Monday, June 30, 2014 2:55 AM
>>>> To: "paraview at paraview.org" <paraview at paraview.org>
>>>> Subject: [EXTERNAL] [Paraview] Rendering in parallel
>>>>
>>>>    Good morning,
>>>>
>>>>  I am a novice user of Paraview and there are some aspects which I am
>>>> not familiar with. Here is one of the issues I am having :
>>>>
>>>> I run Paraview in Client-Server mode, performing the data processing
>>>> and the rendering on the remote server, and I read a Tecplot Binary File
>>>> (.plt) composed of more than 30 millions of points. This take a lot of
>>>> time. An idea to speed up the calculation is to launch the server in
>>>> parallel. I know that many readers can not read in parallel (it is the case
>>>> of TecplotBinaryFileReader I think) so I don't expect any improvment in
>>>> this way.
>>>>
>>>> But, examining the Timer Log, I noticed that it doesn't speed up the
>>>> rendering either. I tested many times displaying the points and both
>>>> experiment with parallelism and without gave the same results (about 40-50
>>>> sec). I don't understand why.
>>>>
>>>>  Do I misinterpret the Timer Log ? Is the time of rendering long
>>>> enough to conclude ? Do I have to set specific parameters to make it works ?
>>>>
>>>>  I thank you in advance for your help.
>>>>
>>>>  Jérémy
>>>>
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/paraview/attachments/20140703/d2ac3b8e/attachment-0001.html>