operations icon indicating copy to clipboard operation
operations copied to clipboard

Replace karm and eddie

Open pnorman opened this issue 3 months ago • 2 comments

ref #1294

#1116 #1117 in late 2024 enlarged the arrays enough that we haven't run out of space yet, but my math shows the DB will be at 16.9TiB by early 2027 with the possibility of spikes to 18.3TiB. As the array is 18TiB, it will run out of space.

There are empty bays, but the CPU year is 11 years ago. Should we replace them? I'd prefer to have more efficient machines where we're sure they can handle production loads.

Image

pnorman avatar Oct 02 '25 01:10 pnorman

I saw you mentioned second generation Xeon Scalable as proposed baseline for new purchases.

If you need that, with high performance, possibly the Intel Xeon Gold 6248R processors are a good choice, they have considerably higher multi-threaded performance than the first iteration of the same 2nd generation processors (Xeon Gold 6248).

With a dual CPU multi-threaded Passmark score of 59000, this would be almost 3x the capacity of the current dual CPU E5-2660 v3 systems at 21931. Or Xeon Gold 6258R, with 64643 score for dual CPU.

They are still not really cheap, but second hand prices are starting to come down, with a quick search of +/- $600-700 per CPU for both Xeon types.

Just watch the 205W required cooling capacity. Maybe not all early 2nd gen Xeon Scalable server systems supported that, I saw warnings for that.

Although, if AMD is acceptable as well, then going for the equivalent of the culebre processors (AMD EPYC 7453) is even better. At an an equivalent to even lower second hand price, you push multi-threaded performance of a dual CPU system to 80327, 4x the capacity of the current systems.

mboeringa avatar Oct 22 '25 19:10 mboeringa

AMD-based DL365s are not commonly available on the used market. In fact, I don't see any 10 bay NVMe ones from our usual suppliers. We won't be picking the CPUs until time of purchase as prices are driven by available supply.

pnorman avatar Oct 23 '25 05:10 pnorman