Over the last few decades, computer systems have implemented the memory-storage hierarchy shown in Figure 1. The memory-storage hierarchy utilizes the principle of locality which keeps frequently accessed data closest to the CPU. Successive generations of technologies have iterated on the number, size, and speed of caches to ensure the CPUs have access to the most frequently used data. CPU speeds have continued to get faster, adding more cores and threads with each new CPU generations as they try to maintain Moore's Law. The capacity, price and speed of volatile DRAM and non-volatile storage such as NAND SSDs or Hard Disk Drives have not kept pace and quickly become a bottleneck for system and application performance.
Persistent Memory (PMEM), also referred to as Non-Volatile Memory (NVM) or Storage Class Memory (SCM), provides a new entry in the memory-storage hierarchy shown in Figure 2 that fills the performance/capacity gap.
With persistent memory, applications have a new tier available for data placement. In addition to the memory and storage tiers, the persistent memory tier offers greater capacity than DRAM and significantly faster performance than storage. Applications can access persistent memory like they do with traditional memory, eliminating the need to page blocks of data back and forth between memory and storage.
The Storage Network Industry Association (SNIA) and several technology industry companies derived several standards, including the NVM Programming Model, to enable application development for persistent memory. NVM Programming Model Version 1.2 can be found on the SNIA Persistent Memory website. Each year SNIA hosts a Non-Volatile conference. Papers and recordings of previous conferences can be found on the SNIA Persistent Memory home page.
The introduction of a persistent memory tier offers application developers a choice of where to put data and data structures. Traditionally data was read and written to volatile memory and flushed to non-volatile persistent storage. When the application is started, data has to be read from storage into volatile memory before it can be accessed. Depending on the size of the working dataset, this can take seconds, minutes, or hours. With clever application design, developers and application architects can now take advantage of this new technology to improve performance and reduce application startup times.
Persistent Memory introduces some new programming concerns, which did not apply to traditional, volartile memory. These include:
Stores are not guaranteed to be persistent until flushed. Although this is also true for the decades-old memory-mapped file APIs (like mmap() and msync() on Linux), many programmers have not dealt with the need to flush to persistence for memory. Following the standard API (like msync() to flush changes to persistence) will work as expected. But more optimal flushing, where the application flushes stores from the CPU caches directly, instead of calling into the kernel, is also possible.
CPUs have out-of-order CPU execution and cache access/flushing. This means if two values are stored by the application, the order in which they become persistent may not be the order that the application wrote them.
8-byte stores are powerfail atomic on the x86 architecture -- if a powerfail happens during an aligned, 8-byte store to PMEM, either the old 8-bytes or the new 8-bytes (not a combination of the two) will be found in that location after reboot.
Anything larger than 8-bytes on x86 is not powerfail atomic, so it is up to software to implement whatever transactions/logging/recovery is required for consistency. Note that this is specific to x86 -- other hardware platforms may have different atomicity sizes (PMDK is designed so applications using it don't have to worry about these details).
Memory leaks to persistent storage are persistent. Rebooting the server doesn't change the on-device contents. In the current volatile model, if an application leaks memory, restarting the application or system frees that memory.
Byte Level Access:
Application developers can read and write at the byte level according to the application requirements. The read/writes no longer need to be aligned or equal to storage block boundaries, eg: 512byte, 4KiB, or 8KiB. The storage doesn't need to read an entire block to modify a few bytes, to then write that entire block back to persistent storage. Applications are free to read/write as much or as little as required. This improves performance and reduces memory footprint overheads.
Applications may need to detect and handle hardware errors directly. Since applications have direct access to the persistent memory media, any errors will be returned back to the application as memory errors.
Application developers can use traditional system calls such as memcpy(), memmap(), etc. to access the persistent memory devices. This is a challenging route to take as it requires intimate knowledge of the persistent memory device and CPU features. Applications developed for one server platform may not be portable to a different platform or CPU generation. To help application developers with these challenges, an open source Persistent Memory Development Kit can be downloaded and freely used. More information can be found in the PMDK Introduction, Installing PMDK, and PMDK Programmers Guide found within this document collection. Tuned and validated on both Linux and Windows, the libraries build on the Direct Access (DAX) feature of those operating systems which allows applications to access persistent memory as memory-mapped files, as described in the SNIA NVM Programming Model.
Managing persistent memory devices on Linux is achieved using the
ndctl utility. Read the NDCTL Introduction, Installing NDCTL, then NDCTL Users Guide within this collection. Windows users can use the
Get-PhysicalDisk power-shell command. More information can be found on the Microsoft Documents site.
The following is a list of documents in this collection in the suggested reading order:
Getting Started Guide (this document): Describes how to install and configure the Persistent Memory Development Kit (PMDK) and Non-Volatile Device Control (NDCTL). It is designed to get users up and running quickly with the software.
Changelog: Describes the historical log or record of all notable changes made to the projects.
NDCTL Users Guide: Describes how to use the ndctl, daxctl, and daxio utilities to configure, manage, and monitor real or emulated Non-Volatile Devices.