Is Intel Optane PMem a Good Fit for How You Consume Memory?
In this article
Launched in 2019, Intel Optane PMem bridges the gap between storage and memory by combining the most useful features of both to provide near-instant access to more data in memory. Available in two modes, App Direct Mode and Memory Mode, Intel Optane PMem cost-effectively combines large capacity and support for data persistence, helping prevent the loss of critical data. It can result in denser, more consolidated virtualized server environments and lower costs.
Before investing in a migration to Intel Optane PMem, it is important that you determine whether the way your workloads consume available memory is within the ideal ratio that makes persistent memory technology running in Memory Mode a sensible, cost-effective choice. To do this, WWT would like to share eight steps for using one free and easy tool – Dell EMC Live Optics workload observation software – to help you assess VMware workloads for suitability on Intel Optane PMem, and how we can assist in designing your overall data center transformation.
Knowing how memory is used is key to maximizing Intel Optane PMem value
The first place to start is with an inventory of workloads that might benefit from a migration to persistent memory, such as artificial intelligence (AI), virtual desktop infrastructure (VDI), analytics and high performance computing (HPC). Specifically, it's important to assess how memory is used within those workloads.
The next step is to evaluate memory ratios to determine exactly how memory is being utilized within those workloads. In a virtualized environment, the three relevant categories are:
- Allocated memory is the total amount of memory available to a given virtual machine (VM).
- Consumed memory is the amount of that memory allocation that the VM has loaded into memory for the operating system, applications and data.
- Active memory is the portion of consumed memory that's currently in use for read/write functions versus rarely accessed or inactive datasets.
Ideally, the active-memory workloads should be handled in the DRAM cache layer, while non-active data should exist in persistent memory. As consumed memory shuffles in and out of active use, the processor employs tiering to shunt "stale" memory from the DRAM cache to persistent memory as needed. There should be sufficient DRAM to accommodate the need for active memory and enough persistent memory for all the consumed memory.
The Live Optics tool is able to assess your existing environment with traditional DRAM, measuring the ratio of memory that's actively in use versus consumed memory devoted to dormant datasets. A common ideal conservative ratio between active and total consumed memory should be approximately 1:4.
From this analysis, the Live Optics tool indicates whether your use case is within that ideal ratio. If they fall outside the ratio, they may require further exploration and a more customized solution. Customers may perform this validation step themselves, or they can bring the results to WWT for interpretation. At your direction, we can even get started on a migration solution design for optimum performance and total cost of operation (TCO).
Eight steps for assessing your memory consumption with the Live Optics tool
Live Optics is an example of a robust, full-featured OEM tool with a range of useful analytical features that can be utilized for assessing how to optimize Intel Optane PMem for your environment, workloads and memory usage. What follows is a step-by-step procedure for accessing, running and interpreting the Live Optics tool.
1. Download the latest Live Optics collector software
- Log in or sign up to the Live Optics portal account at https://www.liveoptics.com
- Download a collector
- Common process is to download the collector to a Windows VM/JumpBox that has network access to the vCenter Server Appliance
- Unzip the files
- Double-click on the Live Optics file to start the app
2. Run the Live Optics collector software
- Accept the EULA
- Go to Server & Virtualization > Optical Prime > Click Start
- Select 'Establish secure https connection' to Live Optics site > Click Next
- Name your project > Next
- Accept/update Target directory for data collection
- Click on Add Remote Server
- Select VMware vCenter and provide the IP address > Connect
- Provide the vCenter Server credentials
- Select the data center clusters to be evaluated
- Select the duration of collection
- Recommended a minimum of 24 hours
- Start Capture
3. View Live Optics project
- Log in to Live Optics portal
- Click on View Projects
- Select/click on your named project and start analysis
4. Understand current HW configuration
Use the environment tab
- Content in our use case
- 919 VMs on 28 hosts / hypervisors
- 1048 cores @2.3GHz (2388GHz) - 40% peak util%
- 20.96TB Mem - 60% Peak Util%
- 1.82PB Storage - 65% used
vSphere's Memory counters enable identification of Hot Areas
VM memory size
- The memory provisioned to the VM
- It is an upper limit defined at VM creation
- The amount of host physical memory consumed by the VM
Active memory (the Hot area)
- Is a part of the consumed memory that has been recently touched by I/O operations
*Provisioned or Consumed determine PMem needs
*Active determines DRAM needs
5. Obtain cluster memory utilization data
Provisioned, Consumed and Active Memory
- Click on the 'Virtual' tab
- Obtain Provisioned, Consumed and Active Memory(*)
- Provisioned is used to determine PMem size
- Active is used to determine DRAM size
- Consider memory rationalization when provisioned significantly higher than consumed
- Increase memory capacity when provisioned >75% of physical
Our use case information
- Provisioned (15917GB) is 30% larger than Consumed (12366GB)
- Provisioned is 75% of the total physical capacity (20.96TB) - environment tab
- Active to consumed ratio is 1:9 (1400/12366)
Provisioned = how much capacity the VMs were requesting.
Consumed = how much physical memory was allocated to the VMs.
Active = the part of the consumed memory that is frequently touched (the hot area).
6. Obtain VM level Provisioned vs. Active
- Under 'Virtual Tab'
- Scroll down to VMs section
- Click on column label to sort and identify top consumers
- Large VMs affect DRAM sizing
- DRAM capacity should be large enough to accommodate the VMs with the largest active memory
7. Identify opportunities for host consolidation/optimization
Consider CPU and memory sizes, host storage, and IO capabilities
- Identify opportunities for host consolidation when transitioning to new CPU gen
- Adjust memory requirements for new node count and identify optimization opportunities
- Adjust storage capacity and IO capabilities for new node count
Tiered memory sizing general recommendations
- Set Max DRAM & PMem utilization to 80%
- Use Provisioned to determine PMem Size
- Use Active memory to determine DRAM size
- Consider configuring a 1:4 DRAM to PMem ratio
8. The final step – evaluating your configuration options
Assess both DRAM and PMem utilization levels and evaluate cluster level costs
Once you have run the report and collected your data, WWT recommends any of these options:
- Connect directly to the Live Optics web service to view your project in the Live Optics portal.
- Save your data to an SIOKIT file and upload it to the portal for analysis.
- Reach out to WWT for assistance in analysis and recommendations for designing your most practical and beneficial solution.
Smart memory choices start with understanding your environment
Intel Optane persistent memory running in Memory Mode can deliver significant capacity or cost benefits by using DRAM more efficiently. As a general rule, a 1:4 ratio between active memory and total consumed memory is ideal. However, every use case is different and the only way to know how best to incorporate Intel Optane PMem is to run the numbers and analyze the results.
Downloading and running the Live Optics tool is a good start. WWT is here to help with advice, analysis and full assessment capabilities.