vD vDan VMware · VCF · NSX · K8S

VCF Memory Tiering Resource

Short answer: VCF memory tiering (VMware Advanced Memory Tiering) lets an ESX host use a local NVMe device as a second memory tier behind DRAM, so the same server runs about twice the VMs at the default 1:1 ratio with a measured loss under 10 %. This resource puts everything about it on one page with six tabs: what it does and how VCF 9.0 and 9.1 differ, the host, drive and VM requirements with a readiness checklist, an interactive DRAM:NVMe sizing calculator that applies the 50 % active-memory rule and estimates the memory cost saving, the 9.0 and 9.1 configuration steps with an esxcli command generator, the thresholds to monitor with the troubleshooting table, and every official blog, document, KB, lab, Explore session and video, filterable and exportable as Markdown.

Prefer a full window? Open the resource full screen →

VCF memory tiering rules at a glance

  • One sizing rule. Keep the peak active memory of the host at or below 50 % of its DRAM. Between 50 and 75 % test first, above 75 % expect a performance loss. In practice, most workloads sit at 10 to 30 %, which is why the default 1:1 ratio doubles capacity safely; ratios up to 1:4 exist for very cold workloads such as VDI.
  • The right drive. A local PCIe NVMe from the Broadcom Compatibility Guide with Endurance Class D (7,300 TBW or more), Performance Class F or G, mixed-use at 3 DWPD, dedicated to tiering, at least as large as DRAM and partitioned up to 4 TB. However, two plain devices are not redundant in 9.0; use hardware RAID 1 there and software mirroring in 9.1.
  • 9.0 versus 9.1. 9.0 needs a manual partition, the VMkernel.Boot.memoryTiering boot option and a reboot. In contrast, 9.1 does it in maintenance mode without a reboot through vSphere Configuration Profiles (memtier > nvme) or one esxcli memtier enable command, and the old commands no longer work there.
  • What to watch. At first, Tier 1 stays empty until DRAM consumption reaches about 80 %, which is normal. Watch active memory, NVMe read latency (under 200 µs), CPU headroom on non-tiered hosts and, in 9.1, the device health report and its Available Spare.

Sources