Plotting a large number of simulations

OK…so we have our huge number of simulation cases for our educational project.

I think I want to organize the simulations as follows:
1. Do a mkillum calculation for the glazing that includes the site model.
2. Use BDSF for the interior blinds in their several positions
3. Other than that, straight Radiance simulation.

Does this seem realistic? Or am I off the deep end somewhere?

···

--
Randolph M. Fritz

This seems reasonable. I assume you have access to sufficient computing
resources. If not, read on....

If you need cheap resources recommend using Amazon's EC2 SPOT instances.
The SPOT instances are not guaranteed. However, they are cheap: you can get
the equivalent of a dual-core Intel Core 2 Duo for about 6 cents an hour
(high-CPU medium instance). You can check the prices half way down this
page: http://aws.amazon.com/ec2/spot-instances/. You can buy up to 100
virtual machines in a region. The list of regions is constantly expanding,
but I used Virginia, N. California, Ireland and Tokyo simultaneously at one
point to run 800 parallel simulations for HMG's PIER work. The cost was
reasonable and the machines are easy to start up and shut down. I would
recommend using the "Simple Queue Service" to keep track of which model
need to be run -- it is easier to have each machine poll the queue than to
try and schedule jobs before hand and deal with the (potentially) disparate
simulation times.

Out of curiosity, how many runs are you thinking of? Have you timed each
step?

Good luck and let me know if this was helpful.

Sincerely,
Tim

···

On Mon, Feb 27, 2012 at 12:54 PM, Randolph M. Fritz <[email protected]> wrote:

OK…so we have our huge number of simulation cases for our educational
project.

I think I want to organize the simulations as follows:
1. Do a mkillum calculation for the glazing that includes the site model.
2. Use BDSF for the interior blinds in their several positions
3. Other than that, straight Radiance simulation.

Does this seem realistic? Or am I off the deep end somewhere?
--
Randolph M. Fritz

______________________________**_________________
Radiance-general mailing list
Radiance-general@radiance-**online.org<[email protected]>
http://www.radiance-online.**org/mailman/listinfo/radiance-**general<http://www.radiance-online.org/mailman/listinfo/radiance-general>

This seems reasonable. I assume you have access to sufficient computing resources. If not, read on....

Thanks. We do--the Labs has a cluster, and we are one of its small users--, but this is interesting even so.

If you need cheap resources recommend using Amazon's EC2 SPOT instances. The SPOT instances are not guaranteed. […]

Out of curiosity, how many runs are you thinking of? Have you timed each step?

In the thousands. No, not timing them yet. I am guessing in the hour or so for the final simulations; I have no idea yet for the mkillum and genbsdf runs.

Good luck and let me know if this was helpful.

Thanks!

Randolph

···

On 2012-02-27 22:46:33 +0000, Tim Perry said:

Hi Randolph,

Is there any reason in particular for doing it this way? Having a BSDF surface in front of an illum has a number of disadvantages, the main one being that the room cannot sample the illum directly. You would be better off doing a 3-phase calculation where the exterior-to-sky matrix were computed once using rtcontrib, and the interior-to-image matrix in another precalculation. Then, your thousands of runs are just genskyvec and dctimestep calculations that go very quickly by substituting different blinds BSDFs.

If there is some reason you don't want to do this with a 3-phase method, you should include the blinds as part of your mkillum computation, even if that means running it multiple times for different blinds positions. That way, the interior can "see" the illum source rather than having every source ray blocked by the blinds, which completely defeats the purpose of mkillum.

Does this make sense?
-Greg

···

From: "Randolph M. Fritz" <[email protected]>
Date: February 27, 2012 12:54:38 PM PST

OK…so we have our huge number of simulation cases for our educational project.

I think I want to organize the simulations as follows:
1. Do a mkillum calculation for the glazing that includes the site model.
2. Use BDSF for the interior blinds in their several positions
3. Other than that, straight Radiance simulation.

Does this seem realistic? Or am I off the deep end somewhere?
--
Randolph M. Fritz

Couldn't he put a BSDF surface behind the mkillum? Then mkillum could sample the BSDF instead of the blind geometry, saving mkillum from having to sample the blinds in all 1000+ simulations. You could also include blind geometry with direct pass through option for the BSDF allowing the blinds to be seen in the renderings too. No?

···

On Feb 27, 2012, at 4:43 PM, Greg Ward wrote:

If there is some reason you don't want to do this with a 3-phase method, you should include the blinds as part of your mkillum computation, even if that means running it multiple times for different blinds positions. That way, the interior can "see" the illum source rather than having every source ray blocked by the blinds, which completely defeats the purpose of mkillum.

Hi Andy,

Mkillum produces an illum surface that is too close to the BSDF surface for the blinds to be sampled properly as a light source, and at that point everything becomes indirect. If the BSDF is serving as a proxy for actual blinds geometry, then the illum will be bypassed as well during source sampling.

The only thing that would "sort of" work is to run mkillum in "l+" mode so the window gets replaced by a light source, then use the BSDF as a proxy. You still wouldn't end up seeing what's outside through the blinds or get striped patterns during solar penetration, but at least the direct and indirect calculations would be reasonable.

I don't see the advantage over a three-phase calculation if you are using BSDFs, which would be a thousand times faster. The shortcomings are similar, but the costs are very different.

The highest quality results are still going to come from handing the blinds geometry to mkillum and letting it figure out the light distribution on its own. You can refer to my "first class citizen" talk from the last workshop, where I tried to explain this reasoning. Unfortunately, I don't know of a good shortcut in the case where you want to look at many blinds positions for the same time step. Do you have any ideas?

-Greg

···

From: Andy McNeil <[email protected]>
Date: February 27, 2012 5:25:27 PM PST

On Feb 27, 2012, at 4:43 PM, Greg Ward wrote:

If there is some reason you don't want to do this with a 3-phase method, you should include the blinds as part of your mkillum computation, even if that means running it multiple times for different blinds positions. That way, the interior can "see" the illum source rather than having every source ray blocked by the blinds, which completely defeats the purpose of mkillum.

Couldn't he put a BSDF surface behind the mkillum? Then mkillum could sample the BSDF instead of the blind geometry, saving mkillum from having to sample the blinds in all 1000+ simulations. You could also include blind geometry with direct pass through option for the BSDF allowing the blinds to be seen in the renderings too. No?

Hi Randolph,

Is there any reason in particular for doing it this way?

An error on my part. Ooof!

Having a BSDF surface in front of an illum has a number of disadvantages, the main one being that the room cannot sample the illum directly. […]

If there is some reason you don't want to do this with a 3-phase method, you should include the blinds as part of your mkillum computation, even if that means running it multiple times for different blinds positions. That way, the interior can "see" the illum source rather than having every source ray blocked by the blinds, which completely defeats the purpose of mkillum.

There's sometimes direct sun, and I thought the three-phase method wouldn't work in that situation. Fortunately, there are not too many times involved (those come in the next phase of the project), so I will do as you suggest and roll the blinds into the mkillum calculation

Does this make sense?

Yes. Thank you. You've saved me from a huge mistake and a lot of lost time.

Randolph

···

On 2012-02-28 00:43:09 +0000, Greg Ward said: