The memory resource is of the highest importance for SQL Server workloads. Modern versions of SQL Server benefit from accessing memory resources to use different caches and reduce disk IOPS effectively.
Check with your DBA team to get memory usage metrics directly from the SQL Server instance (Distributed Management Views (DMVs), particularly sys.dm_os_sys_memory, provide a lot of useful information to accomplish this task). Avoid using OS- or vSphere-based memory metrics for the memory rightsizing task. You can find more details about rightsizing SQL Server memory resources here.
From the performance perspective, it’s recommended to stay within one physical NUMA node while allocating memory. As both physical server models used in VMware Cloud on AWS have relatively high memory density, it should not be a problem. However, if you allocate more than 256/384 GB on i3/i3en instances, respectively, ensure to align the vNUMA configuration to accommodate for the extra amount of memory.
Do not overcommit memory for any VM hosting SQL Server workload with high-performance requirements. ESXi memory overcommitment highly impacts SQL Server performance. Therefore, the following metrics on the VM and host objects should be monitored, and an alert created if the value will be above 0: