Shortcut: NVMe Memory Tiering an optimization boost for your TCO
The future of memory management in VMware vSphere is here, and it’s set to revolutionize how we handle workloads and optimize our infrastructure. The “Advanced NVMe Memory Tiering” feature is now generally available with the GA release of VCF 9, a feature that promises to significantly boost memory capacity and reduce your total cost of ownership (TCO).
What is NVMe Memory Tiering?
NVMe Memory Tiering allows you to use NVMe devices as a secondary memory tier alongside traditional DRAM. This innovative approach dramatically increases the available memory capacity of your ESXi hosts. Imagine expanding from 64GB of DRAM to several hundred GB with NVMe – the possibilities are immense!
Key Benefits of NVMe Memory Tiering
- Massive Memory Expansion: Significantly increase your host’s memory capacity, enabling you to run more and larger virtual machines (VMs).
- Cost Efficiency: NVMe is significantly cheaper per GB than DRAM, allowing for larger memory footprints at a fraction of the cost.
- Improved Workload Consolidation: Run more VMs on a single host, enhancing utilization and consolidation ratios.
- Reduced TCO: Lower overall hardware costs by supplementing DRAM with NVMe, reducing the need for expensive DRAM upgrades.
- Sizing for Average Usage: Optimize memory sizing for average workloads instead of peak demand, saving on resources and costs.
Real-World Scenarios and Cost Savings
The following diagram shows a cost comparison between 32GB, 48GB, 64GB, 96GB and 128GB modules, for a host with 32 modules (net prices).
The result: 128GB, 96GB and 48GB modules are currently not economic compared to 64GB and 32GB modules.
Consider an environment with low CPU utilization but high memory usage. Without NVMe tiering, you’d likely need to purchase additional servers, increasing both capital expenditure (Capex) and operational expenditure (Opex). With NVMe tiering, you can avoid this, enhancing workload density on existing hardware.
A comparison shows significant cost savings:
| Scenario | Component | Details | Costs |
|---|---|---|---|
| Scenario 1 (RAM modules only) | HPE GreenLake 380 Gen12, 2 socket, 24 core each | 32x64GB memory (2TB), moderate discount, 5 years hardware refresh | 2x €20.000 |
| Scenario 1 (RAM modules only) | Opex (Facility, annually) | Power, cooling, rack | 2x €6.000 |
| Scenario 1 (RAM modules only) | Opex (Labor, annually) | 1h/month per server, fully burdened salary = €150k | 2x €1.020 |
| Scenario 1 (RAM modules only) | Sum (over 5 years) | €110.200 | |
| Scenario 2 (NVMe Memory Tiering for suitable workloads) | HPE GreenLake 380 Gen12, 2 socket, 24 core each | 32x64GB memory (2TB), moderate discount, 5 years hardware refresh | €20.000 |
| Scenario 2 (NVMe Memory Tiering for suitable workloads) | Additional NVMe for Memory Tiering | 2x 3.84TB, moderate discount, 2x because of RAID1, 3.84TB recommended due to performance and lifetime | €2.835 |
| Scenario 2 (NVMe Memory Tiering for suitable workloads) | Opex (Facility, annually) | Power, cooling, rack | €6.000 |
| Scenario 2 (NVMe Memory Tiering for suitable workloads) | Opex (Labor, annually) | 1h/month per server, fully burdened salary = €150k | €1.020 |
| Scenario 2 (NVMe Memory Tiering for suitable workloads) | Sum (over 5 years) | €57.935 |
Yes, you are reading it right, with NVMe Memory Tiering we can talk about a DRAM:NVMe ratio of 1:1!
Of course, not all workloads are suitable for NVMe tiering. It’s necessary to categorize the existing and planned workloads due to their characteristics. Below you will find some more information that needs to be considered before using the technology. Important to note, that NVMe RAM will be available as unmappable RAM in ESXi. That means, NVMe tiering cannot be addressed directly like DRAM and therefore has different characteristics in terms of latency and access speed:
- Higher Latency than DRAM: NVMe is fast, but still much slower than traditional memory. Latency-sensitive workloads (e.g., databases) may experience performance degradation.
- Dedicated NVMe Devices Required: The NVMe device cannot be used for other purposes (e.g., as a datastore) at the same time.
- Compatibility and Support: Chosen hardware must follow the VMware compatibility guidelines.
- Lifespan and Heat Generation: NVMe SSDs have a limited lifespan and can generate more heat under heavy load compared to DRAM, potentially requiring additional cooling.
Suitable Workloads
NVMe Memory Tiering is ideal for:
- Virtual Desktop Infrastructure (VDI).
- Modern apps workloads.
- Workloads with high memory requirements but low latency sensitivity.
- Workloads with infrequent high memory activity, such as end-of-week transactions or backups.
- SQL Server and Oracle deployments (specific configurations recommended).
However, it’s currently not recommended for:
- Latency- and performance-critical applications like real-time analytics.
- Real-time environments.
- In-memory databases.
Conclusion
NVMe Memory Tiering is set to be an optimization boost for environments where DRAM is limited, cost is a major concern, and workloads have large but not extremely latency-sensitive memory demands. It can help organizations increase capacity, reduce costs, and improve consolidation — but only for the right workloads.