our institute wants to upgrade its hardware and looks for the best bang per buck. We have different students working on different projects on the system. Most complains are about the RAM size (they claim Palabos needs about 45GB RAM for 3D, we have 8GB RAM). I know a ultra-high end server or cluster is the way to go, but the budget is a bit limited.
The current idea is a Tyan S8212 (AMD SR5670 Chipset) with 2 AMD Opteron 2382 (4-way Shanghai) and 64GB RAM (16 x 4 GB DDR2-800). But i’m very skeptic about the CPU and RAM-speed.
What would the forum recommend in the same price-range? Can I double the general LBM performance by doubling the budget? I don’t know the exact simulation setups used by the students and the setups will change anyway from semester to semester.
Or do you have any benchmarks online for more recent workstation class systems in large 3D simulations? I know these sims will take a long time …
I went through this dilema in february. For £3,000 we got a workstation with 2x intel xeon quad cores, 12gb ram and enough PCI-E slots to fill with next years budget. The idea is that we run LB simulations using the CPUs, then when the budget allows we will run them using nvidia tesla/fermi cards. Now I see you want to use palabos, so i think GPU cores is out of the question unless this functionality is supported in the future.
Have you considered using a good SSD (solid state disk) rather than lots of ram? This was suggested to me a while ago as a good way to get fast data transfer for large amounts of data.
It’s true we haven’t published any benchmark on recent hardware; I guess we should do the effort at some point. The current hardware landscape is pretty confusing and unfortunately we can’t make any clear recommendation.
A quick comment for your students who run out of ram: you can cut down the memory needs quite dramatically by switching from D3Q19 to D3Q15, and from double to single precision arithmetics. In many applications we tried, single precision was sufficient to get to the expected result.
I will run my own benchmarks on the available systems and pick a CPU based on the results, hopping the scale-up from about 4GB RAM to 16 - 32 - … GB does not change the performance too much.
We have added recent benchmarks to the Palabos web page: http://www.lbmethod.org/palabos/benchmarks.html . They might be of some use. If anybody is making benchmarks on other platforms, please don’t hesitate to send them to us, we’ll be glad to add them. It would be useful if the same code was used for all benchmarks, I think (the code we used is provided on the benchmark page).
I will run the tests on our new system (may take a month or 2) and post the results. The BlueGene results make me a a bit jealous… but i geuss it will not fit into our IT-room
