[clearsign]

Hey Community, I am creating a small cluster and wanted to bounce these specs off anyone with knowledge in this field.

Software: StarCCM+ Unlimited Cores
Poly Cell Size: 100-300million cells (130million on average)
Objective: Take 1.5-3 day jobs down to 1 day or faster.

3 Nodes
2x AMD EPYC 7371 16-Core 3.1-3.6GHz
16GB DDR4 ECC SDRAM 2666 each slot (8 channel setup)
256GB per Node
768GB cluster total
96 cores cluster total
8GB Ram Per Core

56GB InfiniBand
2x 10GB is also available if InfiniBand is unnecessary

Head Node will use an 1TB 850EVO and a couple 2TB WD Black 7200RPM

Current workstation before the new hardware.
Dell Precision 7920T
2x Intel Xeon Platinum 8168 24 cores 2.7-3.7GHz
64GB x 8 DDR4 ECC 2666 (2 channel setup)
512GB Total
48 Cores Total

Comments?

Will I be ok running head node as a compute node?

Thank you for any input!

[flotus1]

Sounds quite reasonable overall.

Comments?

You might get away with only 10Gig Ethernet as an interconnect with only 3 nodes. Depending on how much Infiniband would add to the overall cost and the hassle of installing it as an afterthought...I think I personally would not skip it. The cluster our group ordered has 8 nodes with 48 cores each so the decision was much clearer.

Head Node will use an 1TB 850EVO and a couple 2TB WD Black 7200RPM

Samsung is now at 860 for their SATA SSDs. However, I would recommend something NVMe for the huge files you will be handling. SSDs are pretty cheap these days. But avoid QLC models like the Intel 660p. I don't know which form factor could fit in your boxes.
And I personally consider 3.5" drives below 6TB a waste of real estate

Will I be ok running head node as a compute node?

Sure, no problem.

I keep scratching my head why Dell and other large players keep selling their workstations with these odd memory configurations. We got a quote from a HP reseller with the same 4 DIMMs per CPU on Skylake-SP when we specifically ordered workstations for CFD.