[pbalz]

Hi,

we are looking for a new rack-mounted workstation in our (small) company to enhance our computing power.
The main tasks will be:
- ~80% CFD calculations with OpenFOAM, mesh size 2-25 millions cells, mostly simpleFoam/pisoFoam, interDyMFoam and chtMultiRegionSimpleFoam.
- 20% FE Analysis, both implicit and explicit, using max 8 cores.

The key features for the workstation will be:
- 2 x Xeon E5 v3 cpus
- 256 GB DDR4 RAM
- SSD + HDD
- basically no graphic power, as the visualization will be done on other workstations


Regarding our area of application, various cpus come into question that are pricely viable:
- 2x Xeon E5-2667 v3: 16 * 3.2 GHz
- 2x Xeon E5-2687W v3: 20 * 3.1 GHz
- 2x Xeon E5-2690 v3: 24 * 2.6 GHz
- 2x Xeon E5-2697 v3: 28 * 2.6 GHz

Which one would be the best choice or are there any other cpus that are even better, considering both the price/performance and the viability to run small jobs (~ 5 mio cells) with a decent performance?


Thanks!

[Andrew1]

There are other people on here with a lot more knowledge who im sure can comment if they disagree with anything, however this was my reasoning when I had to make a similar decision


If you are disabling hyperthreading you can get the processors to run quicker. How much depends on the maximum frequency but a rough guide its half way between stock and max.


Also consider the memory bandwidth per core. All processors use the same 4 sticks of DDR4 memory.


If it’s one node it’s less important but otherwise you may want to consider the cost per core, or the cost per core factoring switch, rackspace costs etc.


(I know using total Ghz isn’t a good indicator as benchmarks are best, but it was a quick approximation)
Processor Speed Cores ~ total Ghz Memory bandwidth / core
E5-2667 V3 3.4Ghz 16 54.4 8.5
E5-2687W 3.45Ghz 20 69 6.8
E5-2690 V3 3.05Ghz 24 73.2 5.67
E5-2697 V3 3.1Ghz 28 86.8 4.87

The first processor overall has less processing power and isn’t actually any quicker per core. The second gets you the best per core performance if that’s most important. If overall processing power is important, the last two are the best. However it comes down to how much memory bandwidth per core you need as to whether it’s worth paying an extra $1200 for 4 more cores.

[pbalz]

Hi Andrew,

thanks for your reply.
Just before your reply we already made a decision in our company and we went with the 2 x E5-2697.
There are several reasons for this:
- 16 or 20 cores were just not enough for massive parallelization
- it turned out that most of our OpenFoam Jobs scale linearly with increasing processor cores up to around 200k cells per core (which I didn't expect!).
- With more cores there comes the possibility to run two jobs in parallel without siginficant decrease in performance.
- turbo boost is more efficient with more cores (I think).

[flotus1]

These are some interesting results. It would be great if you could post some of the benchmark results that brought you to these conclusions.