Understand and deploy persistent memory
Applies to: Azure Stack HCI, versions 22H2 and 21H2; Windows Server 2022, Windows Server 2019, Windows Server 2016, Windows 10
Persistent memory (or PMem) is a new type of memory technology that retains its content through power cycles and can be used as top-tier storage, which is why you may hear people refer to PMem as "storage-class memory" or SCM. This article provides background on persistent memory and explains how to deploy it as the top storage tier in Azure Stack HCI and Windows Server.
What is persistent memory?
Persistent memory is a type of non-volatile media that fits in a standard DIMM (memory) slot. It's slower than DRAM, but provides higher throughput than SSD and NVMe. Compared to DRAM, persistent memory modules come in much larger capacities and are less expensive per GB, however they are still more expensive than NVMe. Memory contents remain even when system power goes down in the event of an unexpected power loss, user initiated shutdown, or system crash. This means that you can use persistent memory modules as ultra-fast, persistent storage.
Azure Stack HCI and Windows Server 2019 support using persistent memory as either a cache or a capacity drive. However, given the pricing model, persistent memory provides the most value as either a cache or as a small amount of dedicated storage for memory mapping data. In most cases, persistent memory drives will be automatically used as cache drives, and anything slower will be used as capacity drives. For more information about how to set up cache and capacity drives, see Understanding the storage pool cache and Plan volumes.
Persistent memory concepts
This section describes the basic concepts you'll need to understand in order to deploy persistent memory in Windows Server and Azure Stack HCI environments to reduce I/O bottlenecks and improve performance.
Access methods
There are two methods for accessing persistent memory. They are:
- Block access, which operates like storage for app compatibility. In this configuration, data flows through the file system and storage stacks as normal. You can use this configuration in combination with NTFS and ReFS, and it is recommended for most use cases.
- Direct access (DAX), which operates like memory to get the lowest latency. You can only use DAX in combination with NTFS. If you don't use DAX correctly, there is potential for data loss. We strongly recommend that you use DAX with Block translation table (BTT) turned on to mitigate the risk of torn writes. To learn more, see Understand and configure DAX.
Warning
DAX isn't supported on Azure Stack HCI environments. Azure Stack HCI only supports block access, with BTT turned on.
Regions
A region is a set of one or more persistent memory modules. Regions are often created as interleaved sets in which multiple persistent memory modules appear as a single logical virtual address space to increase throughput. To increase available bandwidth, adjacent virtual addresses are spread across multiple persistent memory modules. Regions can usually be created in a server platform's BIOS.
PmemDisks
To use persistent memory as storage, you must define at least one PmemDisk, which is a virtual hard disk (VHD) on the host that enumerates as a PmemDisk inside a virtual machine (VM). A PmemDisk is a contiguously addressed range of non-volatile memory that you can think of like a hard disk partition or LUN. You can create multiple PmemDisks using Windows PowerShell cmdlets to divide up the available raw capacity. Each persistent memory module contains a Label Storage Area (LSA) that stores the configuration metadata.
Block translation table
Unlike solid-state drives, persistent memory modules do not protect against "torn writes" that can occur in the case of a power failure or system outage, putting data at risk. BTT mitigates this risk by providing atomic sector update semantics for persistent memory devices, essentially enabling block-like sector writes so that apps can avoid mixing old and new data in a failure scenario. We strongly recommend turning on BTT in nearly all cases. Because BTT is a property of the PmemDisk, it must be turned on when the PmemDisk is created.
In block access mode, we recommend using BTT because all data will be using block semantics. BTT is also useful in DAX mode because metadata operations still use block semantics, even if the application's data operations don't. Even if all application operations are using memory-mapped files with DAX semantics, torn writes could still happen for the metadata operations; therefore, turning on BTT is still valuable.
Supported hardware
The following table shows supported persistent memory hardware for Azure Stack HCI and Windows Server. Persistent memory is fully supported in Windows Server 2019, including Storage Spaces Direct.
| Persistent Memory Technology | Windows Server 2016 | Azure Stack HCI v20H2/Windows Server 2019 |
|---|---|---|
| NVDIMM-N in persistent mode | Supported | Supported |
| Intel Optane™ DC Persistent Memory in App Direct Mode | Not Supported | Supported |
| Intel Optane™ DC Persistent Memory in Memory Mode | Supported | Supported |
Intel Optane DC Persistent Memory supports both Memory (volatile) and App Direct (persistent) operating modes. To use persistent memory modules as storage, which is the primary use case for server workloads, you must use App Direct mode. Memory mode essentially uses persistent memory as slower RAM, which doesn't usually meet the performance requirements of server workloads. Memory mode is distinct from DAX, which is a persistent storage volume that can be accessed using memory-like semantics.
Operating mode is often preconfigured by the original device manufacturer.
Note
When you restart a system that has multiple Intel® Optane™ persistent memory modules in App Direct mode that are divided into multiple PmemDisks, you might lose access to some or all of the related logical storage disks. This issue occurs on Windows Server 2019 versions that are older than version 1903.
This loss of access occurs because a persistent memory module is untrained or otherwise fails when the system starts. In such a case, all the PmemDisks on any persistent memory module on the system fail, including those that do not physically map to the failed module.
To restore access to all the PmemDisks, replace the failed module.
If a module fails on Windows Server 2019 version 1903 or newer versions, you lose access only to PmemDisks that physically map to the affected module; others are not affected.
Configure persistent memory
If you're using Intel Optane persistent memory, follow the instructions here. If you're using persistent memory modules from another vendor, consult their documentation.
To create a PmemDisk that supports BTT, use the New-VHD cmdlet:
New-VHD E:\pmemtest.vhdpmem -Fixed -SizeBytes 1GB -AddressAbstractionType BTT
The VHD extension must be "vhdpmem".
You can also convert a VHD that doesn't have BTT enabled into one that does (and vice-versa) using the Convert-VHD cmdlet:
Convert-VHD .\pmemtest_nobtt.vhdpmem -AddressAbstractionType BTT -DestinationPath pmemtest_btt.vhdpmem
After converting, the new VHD will have the same namespace GUID as the original one. That can lead to problems, especially if they're both attached to the same VM. To create a new namespace UUID for the converted VHD, use the Set-VHD cmdlet:
Set-VHD -ResetDiskIdentifier .\pmemtest_btt.vhdpmem
Understand interleaved sets
Interleaved sets can usually be created in a server platform's BIOS to make multiple persistent memory devices appear as a single disk to the host operating system, increasing throughput for that disk.
Note
Windows Server 2016 doesn't support interleaved sets of persistent memory modules.
Recall that a persistent memory module resides in a standard DIMM (memory) slot, which puts data closer to the processor. This configuration reduces latency and improves fetch performance. To further increase throughput, two or more persistent memory modules create an n-way interleaved set to stripe read/write operations. The most common configurations are two-way or four-way interleaving.
You can use the Get-PmemDisk PowerShell cmdlet to review the configuration of such logical disks, as follows:
Get-PmemDisk
DiskNumber Size HealthStatus AtomicityType CanBeRemoved PhysicalDeviceIds UnsafeShutdownCount
---------- ---- ------------ ------------- ------------ ----------------- -------------------
2 252 GB Healthy None True {20, 120} 0
3 252 GB Healthy None True {1020, 1120} 0
We can see that the logical PMem disk 2 uses the physical devices Id20 and Id120, and logical PMem disk 3 uses the physical devices Id1020 and Id1120.
To retrieve further information about the interleaved set that a logical drive uses, run the Get-PmemPhysicalDevice cmdlet:
(Get-PmemDisk)[0] | Get-PmemPhysicalDevice
DeviceId DeviceType HealthStatus OperationalStatus PhysicalLocation FirmwareRevision Persistent memory size Volatile memory size
-------- ---------- ------------ ----------------- ---------------- ---------------- ---------------------- --------------------
20 Intel INVDIMM device Healthy {Ok} CPU1_DIMM_C1 102005310 126 GB 0 GB
120 Intel INVDIMM device Healthy {Ok} CPU1_DIMM_F1 102005310 126 GB 0 GB
Configure interleaved sets
To configure an interleaved set, run the Get-PmemUnusedRegion cmdlet to review all the persistent memory regions that are not assigned to a logical persistent memory disk on the system:
Get-PmemUnusedRegion
RegionId TotalSizeInBytes DeviceId
-------- ---------------- --------
1 270582939648 {20, 120}
3 270582939648 {1020, 1120}
To see all the PMem device information in the system, including device type, location, health and operational status, and so on, run the Get-PmemPhysicalDevice cmdlet:
Get-PmemPhysicalDevice
DeviceId DeviceType HealthStatus OperationalStatus PhysicalLocation FirmwareRevision Persistent memory size Volatile
memory size
-------- ---------- ------------ ----------------- ---------------- ---------------- ---------------------- --------------
1020 Intel INVDIMM device Healthy {Ok} CPU2_DIMM_C1 102005310 126 GB 0 GB
1120 Intel INVDIMM device Healthy {Ok} CPU2_DIMM_F1 102005310 126 GB 0 GB
120 Intel INVDIMM device Healthy {Ok} CPU1_DIMM_F1 102005310 126 GB 0 GB
20 Intel INVDIMM device Healthy {Ok} CPU1_DIMM_C1 102005310 126 GB 0 GB
Because we have an available unused PMem region, we can create new persistent memory disks. We can use the unused region to create multiple persistent memory disks by running the following cmdlets:
Get-PmemUnusedRegion | New-PmemDisk
Creating new persistent memory disk. This may take a few moments.
After this is done, we can see the results by running:
Get-PmemDisk
DiskNumber Size HealthStatus AtomicityType CanBeRemoved PhysicalDeviceIds UnsafeShutdownCount
---------- ---- ------------ ------------- ------------ ----------------- -------------------
2 252 GB Healthy None True {20, 120} 0
3 252 GB Healthy None True {1020, 1120} 0
It is worth noting that we can run Get-PhysicalDisk | Where MediaType -eq SCM instead of Get-PmemDisk to get the same results. The newly created persistent memory disk corresponds one-to-one with drives that appear in PowerShell and in Windows Admin Center.
Replace persistent memory
If you have to replace a failed module, you have to reprovision the PMem disk (refer to the steps that we outlined previously).
When you troubleshoot, you might have to use Remove-PmemDisk. This cmdlet removes a specific persistent memory disk. We can remove all current persistent memory disks by running the following cmdlets:
Get-PmemDisk | Remove-PmemDisk
cmdlet Remove-PmemDisk at command pipeline position 1
Supply values for the following parameters:
DiskNumber: 2
This will remove the persistent memory disk(s) from the system and will result in data loss.
Remove the persistent memory disk(s)?
[Y] Yes [A] Yes to All [N] No [L] No to All [S] Suspend [?] Help (default is "Y"): Y
Removing the persistent memory disk. This may take a few moments.
Important
Removing a persistent memory disk causes data loss on that disk.
Another cmdlet you might need is Initialize-PmemPhysicalDevice. This cmdlet initializes the label storage areas on the physical persistent memory devices, and can clear corrupted label storage information on the devices.
Get-PmemPhysicalDevice | Initialize-PmemPhysicalDevice
This will initialize the label storage area on the physical persistent memory device(s) and will result in data loss.
Initializes the physical persistent memory device(s)?
[Y] Yes [A] Yes to All [N] No [L] No to All [S] Suspend [?] Help (default is "Y"): A
Initializing the physical persistent memory device. This may take a few moments.
Initializing the physical persistent memory device. This may take a few moments.
Initializing the physical persistent memory device. This may take a few moments.
Initializing the physical persistent memory device. This may take a few moments.
Important
Initialize-PmemPhysicalDevice causes data loss in persistent memory. Use it only as a last resort to fix persistent memory-related issues.
Persistent memory in action at Microsoft Ignite 2018
To see some of the benefits of persistent memory, let's look at this video from Microsoft Ignite 2018.
Any storage system that provides fault tolerance necessarily makes distributed copies of writes. Such operations must traverse the network and amplify backend write traffic. For this reason, the absolute largest IOPS benchmark numbers are typically achieved by measuring reads only, especially if the storage system has common-sense optimizations to read from the local copy whenever possible. Storage Spaces Direct is optimized to do so.
When measured by using only read operations, the cluster delivered 13,798,674 IOPS.
If you watch the video closely, you'll notice that what's even more jaw-dropping is the latency. Even at over 13.7 M IOPS, the file system in Windows is reporting latency that's consistently less than 40 µs! (That's the symbol for microseconds, one-millionth of a second.) This speed is an order of magnitude faster than what typical all-flash vendors proudly advertise today.
Together, Storage Spaces Direct in Windows Server 2019 and Intel® Optane™ DC persistent memory delivered breakthrough performance. This HCI benchmark of over 13.7M IOPS, accompanied by predictable and extremely low latency, is more than double our previous industry-leading benchmark of 6.7M IOPS. What's more, this time we needed only 12 server nodes—25 percent fewer than before.
The test hardware was a 12-server cluster that was configured to use three-way mirroring and delimited ReFS volumes, 12 x Intel® S2600WFT, 384 GiB memory, 2 x 28-core "CascadeLake," 1.5 TB Intel® Optane™ DC persistent memory as cache, 32 TB NVMe (4 x 8 TB Intel® DC P4510) as capacity, 2 x Mellanox ConnectX-4 25 Gbps.
The following table shows the full performance numbers.
| Benchmark | Performance |
|---|---|
| 4K 100% random read | 13.8 million IOPS |
| 4K 90/10% random read/write | 9.45 million IOPS |
| 2 MB sequential read | 549 GB/s throughput |
Next steps
For related information, see also:
Feedback
Coming soon: Throughout 2024 we will be phasing out GitHub Issues as the feedback mechanism for content and replacing it with a new feedback system. For more information see: https://aka.ms/ContentUserFeedback.
Submit and view feedback for