[Paraview] client/server display problem
Wylie, Brian
bnwylie at sandia.gov
Tue May 16 16:05:06 EDT 2006
Okay here's my general response, with more specific responses below.
Sandia runs ParaView in client/server probably more than any other
institution in the world. ParaView client/server is run about 500 times
a month on 16 different servers(clusters), by about 40 different users
(not including the vis people).
Running ParaView server with Mesa runs absolutely fine! In fact for
small data you can actually see small performance ^increases^ over
OpenGL because grabbing the frame buffer is instantaneous.
So my recommendation for everyone is if you don't have graphics cards,
simply compile ParaView server with mesa. Period. Double stamp. No take
backs.
...
That mode is useful if you do not want to move data from one
cluster to another. If you are running on around 10 nodes, there is a
huge difference in rendering performance between software and hardware
accelerated OpenGL. So it makes sense to render on a cluster with
hardware support.
I agree, I wasn't arguing that you don't want graphics cards, I
was asserting that splitting up the server into data and render is not
useful in my opinion. If you have a cluster that has graphics cards on
it, have that cluster mount the disk where the data is and your done.
If mounting the disk isn't possible, then you have three options
1) The first and best option is to run paraview server with
Mesa.
2) FTP, SCP the data over to the render cluster (I don't like
this option just have it for completeness)
3) Set up data server on the computation cluster, start up the
render server on the graphics cluster, and set up all the sockets
between the two.
Here's my problem with option 3. First it's complex, second you
are moving ALL the geometry over the sockets from data server to the
render server. Moving data (even polygonal) is a major no, no. Again
take my use case of 400 timesteps and the user hits 'play'. Even worse,
the sockets connections from the data server to the render server are
most likely a shared 100T hubbed connection.
Also, there is a sweet spot when you have, let's say, 128 nodes
data server and 16 nodes hardware accelerated render server where you
should get better performance because Mesa is an order of magnitude
slower than hardware accelerated rendering. To boot, compositing with
128 nodes degrades performance further. On the other hand, you obviously
get a lot of performance improvement if you process data with 128 nodes
as opposed to 16.
I bet if you actually ran this test and you compared having the
128 nodes just render their small part of the data locally vs. having
the 128 nodes send their data down to the 16 to have them render it.
That for many, many operations the first would be faster...
1) I'm cutting, I'm clipping, I'm changing time steps, I'm
animating iso surface values, I'm basically trying to get my job done...
:)
2) I'm just timing rendering speed because I'm writing a paper
on why data server / render server is a good idea
ALL the operations in category 1 will be faster with 128 nodes
and mesa.
I don't care about category 2.
Another use case is when you want to render data from a compute
server on a tiled display. The only fast way to do that is to use data
server/render server.
If you have a separate cluster to drive your tiled display and your
can't mount the disk and you can't copy the data and the data is too big
to load on to 16 nodes, then I concede, the data server / render server
would be your option of last resort. :)
Brian Wylie - Org 9227
Sandia National Laboratories
MS 0822 - Building 880/A1-J
(505)844-2238 FAX(505)845-0833
____ _ __
/ __ \____ _________ | | / (_)__ _ __
/ /_/ / __ `/ ___/ __ `/ | / / / _ \ | /| / /
/ ____/ /_/ / / / /_/ /| |/ / / __/ |/ |/ /
/_/ \__,_/_/ \__,_/ |___/_/\___/|__/|__/
Unleash the Beast
________________________________
From: Berk Geveci [mailto:berk.geveci at gmail.com]
Sent: Tuesday, May 16, 2006 1:25 PM
To: Wylie, Brian
Cc: Renato N. Elias; paraview at paraview.org
Subject: Re: [Paraview] client/server display problem
Sandia runs on LOTS of clusters that don't have graphics
cards. We NEVER run in data server mode (I personally believe that mode
is not useful...ever... (waiting for the flames on that statement... :)
)
It has been a while since I had the opportunity to flame Brian
:-)
That mode is useful if you do not want to move data from one
cluster to another. If you are running on around 10 nodes, there is a
huge difference in rendering performance between software and hardware
accelerated OpenGL. So it makes sense to render on a cluster with
hardware support.
Also, there is a sweet spot when you have, let's say, 128 nodes
data server and 16 nodes hardware accelerated render server where you
should get better performance because Mesa is an order of magnitude
slower than hardware accelerated rendering. To boot, compositing with
128 nodes degrades performance further. On the other hand, you obviously
get a lot of performance improvement if you process data with 128 nodes
as opposed to 16.
Another use case is when you want to render data from a compute
server on a tiled display. The only fast way to do that is to use data
server/render server.
Brian Wylie - Org 9227
Sandia National Laboratories
MS 0822 - Building 880/A1-J
(505)844-2238 FAX(505)845-0833
____ _ __
/ __ \____ _________ | | / (_)__ _ __
/ /_/ / __ `/ ___/ __ `/ | / / / _ \ | /| / /
/ ____/ /_/ / / / /_/ /| |/ / / __/ |/ |/ /
/_/ \__,_/_/ \__,_/ |___/_/\___/|__/|__/
Unleash the Beast
________________________________
From: Berk Geveci [mailto:berk.geveci at gmail.com]
Sent: Tuesday, May 16, 2006 1:25 PM
To: Wylie, Brian
Cc: Renato N. Elias; paraview at paraview.org
Subject: Re: [Paraview] client/server display problem
Sandia runs on LOTS of clusters that don't have graphics
cards. We NEVER run in data server mode (I personally believe that mode
is not useful...ever... (waiting for the flames on that statement... :)
)
It has been a while since I had the opportunity to flame Brian
:-)
That mode is useful if you do not want to move data from one
cluster to another. If you are running on around 10 nodes, there is a
huge difference in rendering performance between software and hardware
accelerated OpenGL. So it makes sense to render on a cluster with
hardware support.
Also, there is a sweet spot when you have, let's say, 128 nodes
data server and 16 nodes hardware accelerated render server where you
should get better performance because Mesa is an order of magnitude
slower than hardware accelerated rendering. To boot, compositing with
128 nodes degrades performance further. On the other hand, you obviously
get a lot of performance improvement if you process data with 128 nodes
as opposed to 16.
Another use case is when you want to render data from a compute
server on a tiled display. The only fast way to do that is to use data
server/render server.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://public.kitware.com/pipermail/paraview/attachments/20060516/70dec357/attachment-0001.htm
More information about the ParaView
mailing list