Server hardware - when are too many cores too much

[roardan]

Hello!

our dept. wants to buy a new server for our computational biology needs. We work in the field of metagenomics, so our biggest workloads usually include assembly of large metagenomes (spades, megahit), phylogenetic tree reconstruction (raxml, iqtree) and short read mapping/processing (bbmap, bowtie2, samtools, blast).

I've discussed some specs with our IT department and they came up some preliminary specs (4 x 16-core xeon CPUs, 3 TB RAM, large SSD scratch disks). Here, we discussed the possibility of increasing the number of cores (e.g. 4 x 24-core CPUs) since many of our programs would benefit from an increase in core count given the available RAM. In other words - our assemblies rarely use more than 500 GB of RAM, which would theoretically allow to run 6 assemblies in parallel. This would mean that each assembly gets "only" ~ 11 cores / 22 threads. We would like to bump this but our IT department mentioned that the cores might be memory bandwidth starved if we increase the cores per CPU. So, what is your opinion/experience regarding this? When is too many cores per CPU too much for the programs we use?

Main specs:

CPU: 4 x Xeon Gold 5218, 2.30 GHz, 16-Core; alternative 4 x Xeon Platinum 8268, 2.90 GHz, 24-Core

RAM: 3072GB (48x 64GB) DDR4 PC2933

[GenoMax]

Listen to your IT recommendations. It is one thing to have cores but if you don't have storage infrastructure to support the I/O needs then that will always be the bottleneck. When you run these assemblies you would not notice the 20min more it may take to complete a job especially if it is running overnight.

Highest core count xeons are hideously expensive and you can save a good bit of money by dropping that core count to second/third best.