Rich Getting the Colo to sign off on the water cooling was a nightmare. Dont get me wrong, there’s a lot of massive server provides using water cooling but they have the resources to directly work with the dc and build a system in colab, we’re not that big, so we had a specification sheet, we built it, sent it, got it returned twice, and then finally hit all the goals for allowing us to have it running.
It’s currently using two super micro 2000w, one in redundancy, one in use. The dc recommended a higher psu than what was needed to handle the system.
Answering your p/e core question, from memory I’m not 100% sure will have to check back in to answer that more thoroughly, at a glance though, we have around 150 clients per server, 1:1 website to customer, all higher traffic sites with Fastly serving around 19million requests per 4-5 days, at most I’ve seen around 1000-1500 requests hitting the server. The usage was around 40%, with Fastly offloading most of the requests to cache, our throughput through Fastly is limited to around 500 concurrent origin requests per second though, so theoretically could go higher.
In terms of clock speed across cores, the minimum is around 5.4 on p cores I believe, and 4.7 on e cores. I’ve thought about disabling e cores for the sake of some performance, seeing as php is single threaded, but you lose some threads with the loss of e cores totally.
If now I was to provision a dedicated server with the same specification for a single or two to three customers I’d probably have the RAM setup 1:1 and e cores disabled.
Hope that was informative enough!
