TurbSim on a 64-bit computer

Hi,

The NWTC acquired a 64-bit server this year. The server has two 3.6-GHz, 64-bit, Xeon processors and 6 GB of RAM. It’s running Windows 2003 Standard x64 Edition. I think we paid about $6K for the server.

One reason for getting a 64-bit server was to be able to run denser turbulence grids in TurbSim. On a 32-bit computer, Turbsim has a maximim grid resolution of 19x20 points. I can run a 25x25 grid within the 6 GB of physical RAM of the new server. I can run a larger grid if I’m willing to take a major performance hit by using virtual memory on disk.

I used Intel Visual Fortran (IVF) v9 for EM64T (64-bit) to compile TurbSim on the server. IVF’s predecessor, Compaq Visual Fortran cannot generate 64-bit code.

Here are some benchmarks on the new server:

[code]Grid Size CPU Time (sec) Wall Clock (hh:mm)


15x15 0.6 GB 45.20 same
17x17 1.0 GB 79.38 same
19x19 1.6 GB 132.38 same
21x21 2.4 GB 213.58 same
23x23 3.4 GB 330.92 same
25x25 4.7 GB 499.17 same
27x27 6.4 GB 788.31 2:05[/code]

Here are some from my 3.2 GHz, 32-bit, Pentium 4 PC with just 1 GB of RAM:

[code]Grid Size CPU Time (sec) Wall Clock (hh:mm)


15x15 0.6 GB 149.20 same
17x17 1.0 GB 326.22 1:48
19x19 1.6 GB 553.31 3:15[/code]

Here are some from Alan’s brand new, 3.0 GHz, 64-bit (running 32-bit WinXP), dual-core, Pentium D PC with 2 GB of RAM:

[code]Grid Size CPU Time (sec) Wall Clock (hh:mm)


15x15 0.6 GB 93.51 same
17x17 1.0 GB 178.52 same
19x19 1.6 GB 321.89 same[/code]

I didn’t measure the wall clock time for the server or Alan’s PC when TurbSim ran within memory, so wall-clock and the CPU time were about the same. When having to use on-disk virtual memory instead of physical memory, the CPU time doesn’t increase, but the wall time increases dramatically.

Most of us cannot afford a $6000 server, but for $2100 you can buy a Dell PC with a dual-core, 3 GHz, 64-bit, Pentium D computer with 4 GB of RAM, Windows XP x64 and a decent configuration including a high-end 19" LCD monitor.

Not all software works with 64-bit Windows, so you probably shouldn’t use such a system for your main PC. But it would be great for running simulations.

Alan’s computer cost us about $1400 without a monitor. It is similar to the suggested computer but with only 2 GB of RAM. Even though the clock speed of Alan’s Pentium D PC is less than my 2.5-year-old Pentium 4, it is 60-70% faster.

It’s interesting that the server is 2-2.5 times faster than Alan’s new PC. I don’t know if that is because it is running TurbSim in 64-bit mode, or if a Xeon is significantly faster than a Pentium D. The Xeon has a 2 MB secondary cache for each processor, while the Pentium D shares a 1 MB cache between the two cores. That may explain the difference. I’ve ordered myself a system similar to Alan’s. I’m going to test it temporarily with the 64-bit version of XP to see if that could be the difference. I’ll let know know after it arrives.

Marshall

Marshall,

Would you please let me know if there is an associated increase in the AeroDyn/FAST/ADAMS run time with an increase in the wind grid matrix dimensions.

Because people still refer to this post, I will note that Marshall’s results were obtained prior to TurbSim v1.50. Memory requirements in v1.50 were significantly reduced: TurbSim can now generate grids of 75 x 75 points (630 seconds at 20 Hz) using about 1.8 GB of RAM, which will run on a 32-bit computer. A grid that size will take a long time to generate, though!

I’ve posted some other TurbSim timing results here: [url]http://forums.nrel.gov/t/turbsim-cpu-times/321/1]