Page 1 of 1

Running B3 in parallel on a cluster

PostPosted: Tue Sep 24, 2013 2:37 pm
by ManuelP
Hello B3 Team,

I would like to run B3 on multiple nodes in our cluster, and would like to ask for advice on how to do so best.

I would image the following ways to do so

1) subdivide the ROI, calculating each subarea on different nodes, using the same stack of images, and running the same number of iterations
2) use the same ROI for each node, but split the image stack

Am I mistaken about the second one? Can i use both the same time? What reduced computation time more, reducing ROI size, or image stack size?

In any case, I would have the following questions about each method

For 1)
What about the ROI border? Should the ROI's overlap, or be cleanly ( pixel by pixel ) divided? If there is a border, how thick should it be ( in pixels )?
What is the recommended minimal ROI size in pixels?

For 2)
What is the minimal number of frames I should use per node?

Thank you

PS: a very basic question, but what is the unit of measurement used in the result file, indicating the position of possible fluorescence molecules?

Re: Running B3 in parallel on a cluster

PostPosted: Fri Sep 27, 2013 1:42 pm
by edrosten
ManuelP wrote:Hello B3 Team,

I would like to run B3 on multiple nodes in our cluster, and would like to ask for advice on how to do so best.

I would image the following ways to do so

1) subdivide the ROI, calculating each subarea on different nodes, using the same stack of images, and running the same number of iterations
2) use the same ROI for each node, but split the image stack

Am I mistaken about the second one? Can i use both the same time? What reduced computation time more, reducing ROI size, or image stack size?


This depends very much on the data. 3B works best on data of between 200 and 500 frames or so. If your stacks are very much longer than that then you will almost certainly need to split the data up by frames. It is certainly possible to use both at the same time. If there is significant bleaching in the data, then there is probably one region in time which is the most suitable, and you might be able to get almost all of the information out by looking at that point.

ManuelP wrote:
In any case, I would have the following questions about each method

For 1)
What about the ROI border? Should the ROI's overlap, or be cleanly ( pixel by pixel ) divided? If there is a border, how thick should it be ( in pixels )?
What is the recommended minimal ROI size in pixels?


The smallest size ROI I would recommend using is 10x10. There's also some additional discussion about how to subdivide up an image into multiple ROIs in this thread:

viewtopic.php?f=10&t=30212#p30796

For 2)
What is the minimal number of frames I should use per node?


200 is the minimum.

PS: a very basic question, but what is the unit of measurement used in the result file, indicating the position of possible fluorescence molecules?

[/quote]

In the results file itself, all the results are in pixel coordinates. This is why you need to specify the pixel size in nm when loading an old run with the 3B plugin. Once you've done that, the result in ImageJ is a calibrated image and you can make measurements in nanometres directly.

Regards

-Ed

Re: Running B3 in parallel on a cluster

PostPosted: Fri Dec 26, 2014 7:07 am
by dieting
If your stacks are very much longer than that then you will almost certainly need to split the data up by frames. It is certainly possible to use both at the same time.??