I conducted a test of the varying processor abilities of my array of computers for the purpose of determining the best setup to run the Elite version of Profit. The results were quite interesting to me, and certainly confirm the value of using Profit Elite, if one is going to crank out a lot of models.

The computers that I had available to link using my gigabit network were as follows:

#1 - Intel Dual Xeon CPU Workstation, 3.6 GHz, 4.0 GB Ram, Windows XP Pro

#2 - Intel Dual Xeon CPU Workstation, 3.6 GHz, 2.0 GB Ram, Windows XP Pro

#3 - Intel P3 CPU Desktop, 886 MHz, 1.0 GB Ram, Windows 2000

#4 - Intel P4 CPU Laptop, 3.0 GHz, 512 MB Ram, Windows XP Home

#5 - Intel P4M CPU Laptop, 2.0 GHz, 512 MB Ram, Windows XP Pro

Using Windows XP, one is able to either disable or enable the Hyper-Threading (HT) function of the Xeon and P4 CPUs. If one disables the HT feature (NonHT), then Elite recognizes one processor for each installed CPU within a given computer. If HT is enabled, then Elite recognizes two CPUs for each installed CPU. Consequently, a dual processor computer with XP that is HT enabled will be recognized as having four CPUs. A Windows 2000 computer will only recognize a CPU as one processor. Elite can access up to eight recognized processors (I think this is a Microsoft limitation).

Elite has the ability to either use Multi-Processing or Distributive Multi-Processing in building Models. If I am correct in my assessment, Multi-Processing is only done within the designated Home Computer, while Distributive Multi-Processing is conducted amongst the Home Computer and up to seven other computers … based on your computer array … with eight recognized CPUs always being the maximum.

To make my tests meaningful, I configured the Profit modeling parameters in such a manner as to guarantee that all tests were equal. Here is how I proceeded with my setup:

Six transformations, four years of data, Profit Metric, 30 min. and 30 max. Passes of Degree of Learning, Construct 100 Models, Save 10, Construct Process 1 Time, Free Form only, Seek Best Model, Simple Complexity, Use All Inputs (6), Pool Size 20, Explore Widely, No Tendency To Reduce, Explore Widely. All other settings were default settings.

My first tests were of the Home Computer using the Multi-Processing feature. Here are the results to generate 100 models:

One can determine from these results that most of the Floating Point capabilities of the two CPUs are being reached whether one is using 2 NonHT or 2 HT. There is a marginal improvement using 3 HT, but it can be seen that the Floating Point capabilities are almost being used to their full abilities with two threads. The third and fourth threads don’t do very much. Nonetheless, HT is the winner.

A more beautiful picture emerges, however, when one brings into play additional processors … using the Distributive Multi-Processing feature of Elite.

In testing the Distributive Multi-Processing feature of Elite, I tried almost all of my possible CPU combinations, and will present most of the data in this report. Here are the data:

Case 1 - 168 seconds = #1 – 1 HT (1 CPU)

Case 2 - 145 seconds = #1 – 1 NonHT (1 CPU)

Case 3 - 110 seconds = #1 – 2 HT (2 CPUs)

Case 4 - 105 seconds = #1 – 2 NonHT (2 CPUs)

Case 5 - 100 seconds = #1 – 3 HT (3 CPUs)

Case 6 - 97 seconds = #1 – 4 HT (4 CPUs)

Case 7 - 79 seconds = #1 – 4 HT, plus #3 – 1 NonHT (5 CPUs)

Case 8 - 73 seconds = #1 – 4 HT, plus #5 – 1 NonHT (5 CPUs)

Case 9 - 71 seconds = #1 – 4 HT, plus #4 – 1 NonHT (5 CPUs)

Case 10 - 68 seconds = #1 – 4 HT, plus #4 – 1 HT (5 CPUs)

Case 11 - 62 seconds = #1 – 4 HT, plus #4 – 2 HT (5 CPUs)

Case 12 - 58 seconds = #1 – 4 HT, plus #2 – 1 HT (5 CPUs)

Case 13 - 57 seconds = #1 – 4 HT, plus #3 – 1 NonHT, plus #4 – 1 HT (6 CPUs)

Case 14 - 54 seconds = #1 – 4 HT, plus #3 – 1 NonHT, plus #4 – 2 HT (7 CPUs)

Case 15 - 49 seconds = #1 – 4 HT, plus #2 – 2 HT (6 CPUs)

Case 16 - 47 seconds = #1 – 4 HT, plus #2 – 4 HT (8 CPUs)

Case 17 - 45 seconds = #1 – 4 HT, plus #2 – 3 HT (7 CPUs)

Case 18 - 45 seconds = #1 – 4 HT, plus #2 – 2 HT, plus #4 – 2 HT (8 CPUs)

Case 19 - 43 seconds = #1 – 4 HT, plus #2 – 3 HT, plus #3 – 1 NonHT (8 CPUs)

Case 20 - 40 seconds = #1 – 4 HT, plus #2 – 1 HT, plus #3 – 1 NonHT, plus #4 – 2 HT (8 CPUs)

Case 21 - 38 seconds = #1 – 4 HT, plus #2 – 1 HT, plus #3 – 1 NonHT, plus #4 – 1 HT, plus #5 – 1 NonHT (8 CPUs)

Case 22 - 37 seconds = #1 – 4 HT, plus #2 – 2 HT, plus #3 – 1 NonHT, plus #4 – 1 HT (8 CPUs)

Case 23 - 37 seconds = #1 – 4 HT, plus #2 – 2 HT, plus #4 – 1 HT, plus #5 – 1 NonHT (8 CPUs)

The conclusion that I draw from these data is that more CPUs is better than more processor speed. The best of all worlds, though, would probably be to network three of the fastest dual CPU computers that you can buy, and link them 4HT, plus 2HT, plus 2HT. But, don’t throw away your old computers. The P3s are still amazingly effective. Put them to work using Profit Elite!

Now, if any of you guys have a better mousetrap, I would be interested in learning about it … now that you have a basis for comparison. I assume that the 2-Processor 64-Bit AMD Opteron would win out over the Xeon, since the Opteron is supposed to be superior in performing Floating Point calculations. Can anyone confirm this? And if this is the case, do I need to throw everything away and start over? I think I’ll keep my P3, though.

Ted Richardson