Help with Parallel Computing on multiple remote computers

[anmorgan]

Hello, I have used the Parallel Computing Toolbox very minimally and am having troubles finding information on how to use it with other computers besides local.

It is easy enough to use on a local machine, but I cannot find any straight forward generic guides on how to do it remotely with other computers. Can anyone help with more information?

If someone can layout a simple example of starting with a local machine and using other computers where you know their IP addresses, etc. That would be a good start, unless I am over simplifying it and you can't do it this easily.

Thanks for the help.

 

[IRstuff]

I think there are fundamental issues with that approach, namely, the data interface and coordination of the computations. That's why SETI has a problem set that is configured so that each chunk can run standalone on a single machine. So while there's parallel processing, each processing node is essentially independent and isolated from another.

TTFN
faq731-376

 

[LiteYear]

IRstuff - that's what the PCT does. You program as normal but use the PCT version of for-loops and arrays etc, then PCT takes care of the data interface and coordination.

anmorgan - first things first, are you aware that PCT by itself only does local (multicore) computation? You need MATLAB Distributed Computing Server to spread your PCT execution over a cluster (multi-computer).

 

[IRstuff]

Given that the OP was talking about "remote" computers, neither may realistically work. From the PCT's description:

http://www.mathworks.com/cmsimages/51122_wl_pct5_fig1_wl.jpg

it's intended to run the same operations can be on different parts of the a unitary chunk of data. Note the links between processors.

Since remote computers aren't necessarily in lock-step with your local computer, the system will get bogged down waiting for data, particularly if every processor needs to touch every part of the data.

The SETI mode partitions the data along boundaries where a single computer does not need data from any other computer. This allows each processing node to run at its own speed without affecting other nodes, but requires particular problems and datasets that are amenable to that type of partitioning.

Classical parallel programming does not ncessarily fit into that paradigm.

TTFN
faq731-376