A few weeks ago Intel started releasing their Optane product, a commercialization of the 3D Xpoint (Crosspoint) technology they’ve been talking about for a few years. Predictably, there has been a lot of commentary in all directions. Did you know it’s game changing, or that it’s a solution looking for a problem? It’s storage. It isn’t storage. It’s RAM. It isn’t RAM. It’s too slow to be RAM. It’s too small for storage. It’s useful now. Nobody will use it for years.
Yup. Confusion. It’s because Optane is a bunch of different things. It’s consumer and enterprise, and it’s both storage and memory.
There are plenty of articles out there on the technology itself. There’s a small M.2 version for desktops that acts as a cache, which is thoroughly uninteresting to me. I’d rather have a real SSD in one of my precious M.2 slots than a cache that I overrun with three photos from my Nikon SLR. Not to mention I need a 7th generation Intel Core CPU (Kaby Lake) to do this at all.
The real action is with the data center version, the P4800X. The first version is a 375 GB PCIe NVMe card. 375 GB isn’t very much space, but Intel says they’ll have 750 GB and 1.5 TB models out this year. The technology is a lot faster than the NAND flash typically found in SSDs, and the endurance is a lot higher, too (writes to SSDs use voltages that stress, and eventually destroy, the cells in the SSD). Intel says this thing can do 500,000 write IOPS, which makes it a hell of a write cache for something like VMware vSAN, even if it is a bit small. As a storage device, though, Optane is interesting but really just an evolution of NVMe flash technology.
Memory Drive Technology
What’s really interesting to me is the “Memory Drive” component, which seems intent on blurring the lines between memory and storage. You can use the P4800X to create a pool of something that looks to an OS like memory. It is an order of magnitude slower than regular DRAM, but several orders of magnitude faster than an SSD. Given that you could theoretically put 24 TB of Optane in a two socket server — for a lot less money than 24 TB of DRAM – there are some pretty interesting implications. Think about being able to hold a whole enterprise database in memory. The best I/O is one you don’t do, and having all that data close by means a lot less read traffic on your storage, not to mention it being a lot faster.
There aren’t a lot of details about Memory Drive, though. The product brief says it’s Linux only, and that it’s a software layer of some sort. Recently, though, I found a piece over at AnandTech which actually had details around this (link below, kudos to the author, Mr. Tallis, for digging into this). That post indicates it’s a paid add-on, and something like a hypervisor that boots from a USB device, or an IDE controller before the OS loads.
Amateur Hour at Intel
USB or IDE? An extra hypervisor? Paid? What is this, amateur hour? Intel wants me to pay extra for the privilege of booting my servers from a $5 USB drive, which can’t be mirrored or otherwise protected, so that I can load a software layer that basically makes my OS completely unsupported and more complicated? Oooh, sign me up.
Here’s my prediction: no self-respecting enterprise will use this because it is an operational disaster (lack of boot device redundancy, lack of IDE devices, lack of support for popular operating systems, lack of visibility into the Memory Drive layer, even just the nightmare of hardware licensing). As such, nobody will buy the add-on software. A company like Intel charges for features like this to gauge interest, and Intel will eventually incorrectly conclude that the lack of sales is an indicator that nobody is interested. They will then discontinue the product, and because Intel is effectively a monopoly that’ll be the end of this technology. Long live the status quo! Death to the unholy union of DRAM and storage!
On a parallel track, because the poor implementation means little interest from enterprise users, OS vendors won’t be pressured by users, application vendors, or Intel to develop anything for this new layer of addressable storage. That’s a damn shame because there’s real promise here. If Optane support were simply built into the server CPUs and chipsets moving forward, as a native part of what we get for paying the Intel price premiums, people would use it en masse. It should be as easy as plugging an Optane card in and flipping a switch in the BIOS to make it SSD or memory, non-volatile or volatile.
If that happened we’d start seeing real support for it in OSes, applications adapting to use it, and real, interesting, and positive change happening in our data centers. As it stands, though, I fear that Memory Drive is destined to die a slow death for the wrong reasons, at the hands of the ignorant-of-their-customers Ferengis running Intel.