I agree : the ratio 3.2/3.5 is anything but significative.
But the ratio 4/8 of the number of cores nodes should be :
given that all the cores recieve the same amount of runs to execute (resp 2500 on the 4 cores PC and 1250 on the 8 cores one) and that all the cores are active (I do not use Grid:-Launch plus a Send-Recieve protocol), the expected execution time should (?) divided by 2 on the 8 cores machine ... all other things being equal, and more specifically with the same OS.
Now, I agree with your suggestion "so you should run your test using the same number of nodes on each machine"
But two difficulties arise :
- Considering the performances I announced beforehand, I would have like to proceed to some extended comparisons. But (company policy oblige) the migration of operating systems is generally an opportunity to upgrade the working station if not to change it. This is what was done for me and I am no longer capable to test my code on my previous machine
Accordingly, my comparisons are probably biased
- I have observed the following behaviour while using Grid-Run as described in my initial post :
Let us suppose I'm working on a 2x2 cores machine and that (1) I distribute 10000 runs over 4 cores and next (2) I distribute these 10000 runs over 2 cores (same proc or not ???)
Let T(4) and T(2) the corresponding execution times. I can expect that T(2) is twice T(4) .. but, for a reason I don't know, it is not the case (I'm not a specialist of parallel computing or processors architecture).
A quick look to the performance tab of the task monitor shows in case (1) that the 4 cores are loaded the same way (saying 95% during the whole computation sequence) ... whereas in case (2) two cores are loaded up to a level of 75% (with large deviations) while the 2 others remain between 10% and 30%.
Furthermore, the performance history is very chaotic in case (2) whereas quite flat in case (1) ... something I (mis)interpreted as a better task control by the operating system in case (1)
On "my" new Windows-7 machine (4 dual core processors) I have obtained the following results
- Distribution over 8 cores (nodes) : 343 s
- Distribution over 4 cores 409 s (?!?!?!)
These results seem to corroborate your claim "using more nodes will incur a higher percentage of administrative costs" (???)
On "my" old Windows-XP machine (2 dual core processors) I had obtained these results
- Distribution over 4 cores (nodes) : 504 s
- Distribution over 2 cores 983 s
The expected 1:2 ratio is realized here, suggesting a higher efficiency in task control (???)
So I keep thinking that something "is not going well" (more likely with Windows-7)
Other point : now that I have 8 nodes avaliable to me, it is perhaps better to use Grid-Launch with a "master" node and to distribute the compuation over the remaining 7 ???
There are a lot of questions and posts here I need to look at : even the distribution of similar computations is not as simple as we think it is
Even if your answer is far from the luminous solution to which I was expecting to, it leads me to ask myself a lot of questions.
I thank you for that