CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   CD-adapco (http://www.cfd-online.com/Forums/cd-adapco/)
-   -   cluster on Linux (http://www.cfd-online.com/Forums/cd-adapco/54230-cluster-linux.html)

Ivan December 3, 2004 08:03

cluster on Linux
 
1) What is the upper limit on number of cells if I use 2 Xeon and 4 Pentium IV machines?

2) If I use 6 Xeon machines?

3) If I use 1 Itanium and 5 Xeon machines?

Brian December 3, 2004 10:54

Re: cluster on Linux
 
The limit has little to do with the processor, but rather the amount of RAM available to each processor.

I typically use a rule of thumb of 1 million cells per megabyte of RAM. This is a conservative rule - you can often get away with more, but you run the risk of memory paging which slows things down to glacier speed.

That means if you have four processors each with 1MB of RAM, you should be able to handle a 4 million cell case.

But keep in mind that pre and post processing is done on a single processor, so the practical limit to a 32 bit machine, which can't address more than 3.5 - 4MB of RAM (depending on architecture), is about 4 million cells.

Tom December 3, 2004 12:03

Re: cluster on Linux
 
The other problems are building the model or writing the geometry file. Using a 32 bit machine you can only build a model of 3 to 4 million cells no matter how much memory you have. To build a larger model you need to go to a 64 bit machine.

If you try writing the geometry file of a model in excess of 4 million cells using a 32 bit machine you get at least one error for every cell you have. It takes awhile for >4 million errors to scroll by. Tom


All times are GMT -4. The time now is 16:32.