I am doing explicit unstructured Finite Volume CFD on a 24 node message passing parallel computer. I am getting efficiencies of more than 100% (around 115%) when using 2 & 4 nodes.
The problem size is relatively small (< 7 MB per node) comapared to the available memory (30 MB per node). Can anyone attribute reasons towards this?. Also, the efficiency is always maximum when i'm using 4 nodes. I have run codes for 4 different levels of grid and 'am consistently getting the same numbers.
Note: eff = time on 1 proc / (n * time on n proc) time : time per iteration.
Re: Parallelization efficiency..
Efficiencies greater than 100% are quite common for codes which are well suited to parallelisation. (This is termed "super-linear speed-up"). The reason is supposedly to do with increases in the efficiency of the cache use (at least on distributed memory machines).
As the number of processors increases the efficiency will then decrease because of the extra message passing so you should expect a maximum at some point.
Hope this helps a little.
|All times are GMT -4. The time now is 18:27.|