Features Description

Memory tiering is a technology that transparently expands the available memory capacity of virtual machines by using high-performance NVMe SSDs as an extended cold tier for physical memory (DRAM). It is completely transparent to upper-layer applications, operating systems and virtual machines, requiring no code modifications or adaptation. This technology is designed to effectively increase virtual machine density when host physical memory resources are limited, and it is especially suitable for business scenarios with large memory usage but low access frequency.

Configuration Prerequisites

1. Software Version: It is mandatory to upgrade SCP (Cloud Management Platform) and HCI (Hyper-Converged Infrastructure) to version 6.11.3 or above.
2. Hardware Requirements: The host must have free NVMe disk slots and be installed with qualified NVMe SSDs.
3. Disk Status: The NVMe disks used for memory tiering must be dedicated and independent, and shall not be shared with the cache disks of hyper-converged distributed storage (aSAN). In addition, the disks must be free of residual data.

Key Constraints

1. Environmental Limitations: It does not support domestic IT innovation (ARM) environments, security versions (SM versions), and low-latency versions. Currently, the C86 domestic IT innovation environment requires X86 packages for testing and cannot be used in production yet.
2. Virtual Machine Function Limitations: Virtual machines with memory tiering enabled do not support the following features: huge page memory, any hardware passthrough (such as SRIOV, vGPU, PCI passthrough), and ultra-large specification configurations (such as 1TB memory and 256-core CPU).
3. Upgrade Limitations: In scenarios with severely over-provisioned memory, in-place hot upgrades cannot be performed, as the pre-check for hot upgrades will fail due to insufficient memory.

Notes

1. Failure Impact: An NVMe disk failure is a device-level fault, which will cause virtual machines with tiered memory enabled on the corresponding host to crash and restart. However, no data loss will occur if business data has been persistently written to disk.
2. Hybrid Cluster Deployment: It supports the hybrid deployment of hosts with and without memory tiering enabled within the same cluster.
3. Scheduling Mechanism: In the current version, HA or DRS scheduling does not adjust the scoring strategy based on whether memory tiering is configured on hosts. Virtual machines with memory tiering enabled may be migrated to hosts without this feature (where they will fully use physical memory), and vice versa.
4. Other Constraints: The host operating system itself does not consume tiered memory. If the memory of transparent huge page-enabled virtual machines enters tiered storage, the performance optimization brought by huge pages may be invalidated. After disabling the memory tiering function of a virtual machine, memory reclamation is an asynchronous process, leading to a short-term inconsistency between configuration take effect and actual resource release.