On January 18, 2017, the Storage Networking Industry Association (SNIA) held its annual Persistent Memory Summit. This year’s Persistent Memory Summit was the best attended yet. The SNIA had more than double the registrations this year than they had last year. This means more professionals are realizing that persistent memory is real and coming to an infrastructure near you.
The day started off with a discussion about how persistent memory will bring a whole new structure to the data infrastructures in the next generation datacenter for both monolithic and distributed environments. Servers, networking, and storage will all take advantage of communicating data to storage like memory. This means not only will the performance of DRAM memory be a magnitude faster than flash memory, but also the durability of this media is higher as well. Persistent memory technology is nothing new; it has been used in mainframes in small amounts as persistent storage for decades. What is new is the capacity, density, price, and uses cases for modern applications.
Today, we use a standard SCSI stack in the operating system to communicate with hard disk drives (HDDs) and solid state disks (SSDs). The SCSI protocol has served us well for decades for reliable communication between applications and media. But SCSI is also one of the largest performance bottlenecks for communicating with persistent memory. Being able to communicate with solid state media more like memory instead of a mechanical device cuts out a lot of communication overhead. For persistent memory, there needs to be new protocols that communicate more like memory. This does not just include communication to block devices, but fabrics as well. The SNIA has not only developed a programming model for block communications, but it is also working on a standardized programming model for over fabric. With this in mind, there were a several sessions to discuss new protocols and filesystems, like NOVA. The rest of the afternoon was dedicated to application use of persistent memory with Microsoft Windows 2016, Red Hat Linux, and SAP HANA.
The day ended with a discussion about enabling next generation compute platforms and data centric compute architectures to handle storage latencies of 250 nanoseconds. There are several approaches under consideration to accomplish this, which includes 3D XPoint, NS-NVM, ST-MRAM, NVDIMM-P, and others.