We run on Cisco UCS + VMware ESX + HP 3PAR.
Host blade config : UCSB-B200-M4 , Xeon E5-2667 v3 3.1GHz , 2 sockets , 8 cores each, Hyper-Threading Active So total 16 physical cores, or 32 logical cores.
We have Software Assurance with M$, so all are Enterprise editions, and lot of additional SQL core licenses paid for, so money is not an issue.
Our primary single OLTP SQL VM is 'dedicated' to one of the hosts, i.e. no other VMs are allowed to run on it, cos it requires all 16 cores of power. Even with that, CPU regularly runs ~60-80% , so we're planning to upgrade hardware. Questions below :
SQL Server Processor NUMA config - all 16 processors are on 1 NUMA node - But I thought since the host has 2 sockets, there would be 2 NUMA nodes ?
vCenter Socket/Core config - 16 sockets with 1 core per socket - Is it better to change this to 2 sockets with 8 cores per socket since that matched underlying hardware ?
I recently heard at PASS conference, that if an entire VM is being dedicated to a single host, then leave Hyper-Threading disabled. Anybody agree/disagree ?
As a basic pen-paper calculation, we need to double our 'compute capacity' from 50 (16 cores x 3.2Ghz) to ~100 GHz.
Is it better to go for higher core-count across 2 sockets:
E5-4669 : 2 socket x 22 cores x 2.2 Ghz = 96.8 Ghz
or higher clock speed across 4 sockets:
E5-4627 : 4 socket x 10 cores x 2.6 Ghz = 104 Ghz