Hi All, A little piece of i
A little piece of info on the performance of dual core processors with OpenFOAM. I have recently replaced my laptop with an Intel CoreDuo machine. The old one was a 2 GHz Intel Pentium 4-M with a 500 MHz bus; the new Machine is a 2.16 GHz Pentium CoreDuo on a 667 GHz bus. I wanted to know what kind of performance the second core would give me, so I ran a pretty big (1.2 million cell) LES case on both machine. The numbers are not cooked in any way, i.e. the first time-step contains all the I/O, mesh construction etc, followed by 3 time steps.
After the initial comparison (188.26 s on the old machine and 152.38 s on the new one), I have done full parallel domain decomposition runs on the CoreDup CPU. The results seem to show that my CoreDuo is behaving almost like a proper 2-CPU machine, which is very nice indeed!
1 time-step 3 time-steps
Old machine 86.63 cpu 88 clock 188.26 cpu 190 clock
New 1 process 73.57 cpu 75 clock 152.38 cpu 154 clock
New 2 process 46.03 cpu 46 clock 103.01 cpu 104 clock
New 4 process 24.69 cpu 51 clock 51.64 cpu 103 clock
Very interesting! just for
just for curiosity, what kind of laptop did you buy? (I am planning to change my laptop too =) )
It's an IBM T60p (well, Lenovo
It's an IBM T60p (well, Lenovo), model number 2007C8U. I'm running SuSe Linux 10.1.
Not too bad, some problems with the graphics card but very usable.
In case you badly need to use
In case you badly need to use proprietary ATI drivers, here is some related OpenSuse info ;)
Hrv Comparing your "New 1-p
Comparing your "New 1-proc" and "New 2-proc" times, it looks like you are only getting a x1.5 speed-up for the two processors. This is nice, but is short of a x2 speed-up that you were probably dreaming of! In fact, I have seen a similar speed-up when running a dual process Star-CD simulation on a single core hyperthreaded Pentium P5 chip (ie. 2 "virtual" machines).
Two questions for you Hrv:
1. Do you have any feel for how run-time scales with number of (true, single core) processors for OpenFOAM? My experience with Star-CD is that we get a fairly close, linear speed-up (i.e. ~4times quicker when spread across 4 nodes)
2. What was the application that you ran? Did it need a lot of communication between the nodes?
Apologies for the "inquisition", but we are looking to buy some more kit in the near future and I am interested as ever in squeezing out as much as possible from the new boxes!
Hi Ian, Well, in fact I was
Well, in fact I was pleasantly surprised even with 1.5 because this sort of application is all about moving lots of data through the computer and the data bus is probably the bottleneck. Looks like I've got a pretty balanced machine, which came as a bit of a surprise.
Regarding your questions:
1) OpenFOAM will scale very well in parallel with a decent inter-connect. There are some issues with the AMG (which is getting better) and I've got a very good feel on how to improve it further.
2) The application was very representative for a high-end CFD simulation: around 1.3 million cells, shightly loaded on the pressure solver.
Hope this was useful + good luck with the new boxes,
|All times are GMT -4. The time now is 15:17.|