The Intel Optane Memory (SSD) Preview: 32GB of Kaby Lake Caching | Hacker News Reader

The Intel Optane Memory (SSD) Preview: 32GB of Kaby Lake Caching

anandtech.com

87 points2bluesc39 comments

Might be somewhat niche, but this is looks to be a near perfect match for SLOG usage in ZFS based storage schemes where a certain amount of synchronous writes are important. SLOGs don't need much raw space or massive throughput, they need low latency and reliability (including of course persistence in the case of power loss). Traditionally this has been handled either (suboptimally) by flash drives (preferably SLC) or by battery backed RAM drives, which are extremely expensive and have their own potential failure modes. This though looks almost exactly right and could become the new basic answer. In fact it's cheap enough and good enough that it may make ZFS users who previously decided it wasn't cost effective enough to bother with even if some gains were available reconsider.

This is great for anything that needs high iops when doing sync writes. Ceph journal, fsync system calls, database systems and so on.

Something to consider for SLOG (or other journals, e.g., bcache) is that write endurance of the Optane 16-32 GB models is pretty poor. You might want a bigger Optane (or SSD) as a writeahead log / WB cache not because you need the capacity, but because you need the write endurance.

Absolutely an important consideration, and it certainly depends on the use case. The IOPS also isn't remotely as high as the larger models or a ram drive (though SLOGs can be striped for more). But it's $44/$77 whereas the P4800X starts at ~$1500 and something like the DDRdrive X1 (4GB RAM/4GB SLC NAND/power backing) is ~$2k. For a lot of individual applications (and remember, sync behavior can be set per dataset and SLOGs are assigned per pool) even the limited write endurance will be sufficient for a significant period of time, and even if these initial Optanes were burned through within a single year it could still be dramatically more cost effective: at these prices we're talking decade(s) of yearly replacements before equalling the cost of one higher end drive. Also consider where we are on the development curve for this tech. This is effectively 1.0 of XPoint technology, so it's reasonable to expect significant improvements in capacity/$, performance, and so on over the next few years, which in turn means it may be desirable to cycle them relatively fast anyway.

What I had in mind wasn't that these would be replacements for existing solutions used for high sustained intensity applications directly making money. Those are covered by ram drives or striped SLC already (and perhaps the P4800X series soon). I was more thinking the Optane Memory might provide a solid route for experimentation by a new class of users given the capex of a few pizzas. A lot those cases are going to be relatively low sustained intensity, but with burst performance still desirable. Heck, for forty someodd bucks some might go for it just to do some profiling and get a feel for what a separate ZIL would do for them at all.

If anything a significant limiting factor right now may simply be the relative availability/preciousness of M.2 slots in deployed home/soho appliances, servers and workstations.

Yep, those are all good points. Thanks for sharing your thoughts :-).

LinusTechTips just tested this (video is not on YT yet though), and their benchmarks show that it's only really useful as a HDD cache. When used as a SSD cache the only real difference can be found in synthetic benchmarks.

It's unfortunate that it's only available for Kaby Lake, since I imagine that most people in need of a HDD cache run much older systems.

Here's the video I talked about: https://www.youtube.com/watch?v=xgTYSsaNU_A

Intel has a feature where you can use one of their SSDs as a cache. It's fantastic. Except they limit you to 64 gigs no matter what the size of the SSD. Incredibly annoying since what I need cached doesn't remotely fit in 64 gigabytes.

I get to experience lovely things like performance falling off a cliff when stupid programs preallocate 30-40 gigs at a time. If I could use the rest of the SSD it would be no problem.

What desktop user needs fast write caching for synchronous writes? I mean really. It's all buffered writes for programs that work right. Cache size is going to give power desktop users what they need from spinning rust not a small fast cache that won't even fit a single AAA game. 32 gigabytes is a step backwards.

And what is with requiring hardware RAID? It really makes me think all the rapid response stuff is just for marketing.

> "And what is with requiring hardware RAID?"

Intel doesn't really do hardware RAID, just software RAID that the motherboard firmware has a compatible implementation of for boot purposes. There is one hardware feature that Intel relies on for RAID or caching involving NVMe drives: the PCH can remap the PCI registers of NVMe drives so that the NVMe drive is only accessible through non-standard interfaces on the SATA controller. This prevents non-Intel NVMe drivers from binding to the NVMe drive and getting in the way of their caching or RAID. It's a hardware hack Intel created to work around poor software architecture.

Optane menory is a really nice solution for write-back caches to slow HDDs (e.g. ZFS Slog, bcache, ...). Or a cache for often-requested HDD blocks which didn't fit anymore to RAM (ZFS L2ARC).

Whats the different between this and a regular hybrid SSD/HDD besides being more expensive and exclusive to Intel? No benchmarks and pretty bad write speeds(even if latency is low). I bet this would get beat by a SSD/HHD hybrid for a lot less money.

Queue depth is being hyped but means basically nothing for commercial SSD's that have battery backed ram cache. The SSD just queues writes until queue depth is deep enough.

Most SSHDs offer 8GB of flash storage max, and often post numbers that indicate that the flash storage is mediocre. This will absolutely annihilate the vast majority (or all) SSHDs. Is it really so much more expensive? $44 for 32GB of high speed caching is impressive.

Extraordinarily few SSDs, even commercial, have battery backed caches (and those that do often turn into a reliability nightmare), and when they do it tends to be relatively tiny compared to the scope. Queue depth is profoundly important in the real world.

> $44 for 32GB of high speed caching is impressive.

You can get 120GB of SSD for that price, and it would probably do a better job as a hard drive cache.

But yes the better performance at doing persistent writes is very useful for some cases.

One more thing is that every single SSHD ran @ 5400rpm. I ended up buying a conventional HDD @ 7200rpm and bundled a portion of my SSD with PrimoCache.

Yeah, the SSHD offerings from Seagate et al. are atrocious. It's almost as if they don't want to sell them. Perhaps they're afraid of cannibalizing other product lines?

From what I remember a lot of them don't even use the flash to speed up writes, throwing away a lot of the potential performance gains.

Not sure where this will be useful.. Only major advantage seems to be the super low read latency.

I was wondering if this will be any helpful for a product I am working on, but as I use mmap for reading our data, don't really see a benefit after the initial read.

When you have more data than RAM. There's a cost/performance curve between RAM -> SSD -> HDD. This is something between RAM and SSD. Makes sense.

Yes, but this is just 16GB or 32GB, which is not really big.

Think of it as a write-back cache to sliw HDDs. You will be able to have very fast consistent DURABLE cache writes which will be written back async to slow HDDs. 32GB is more than enough for this case.

Or your filesystem caching often requested disk blocks which had to be removed because of memory pressure.

Sure, but why this instead of a 32GB SLC SSD?

Are 32GB SLC SSDs even available any more?

They are, from companies like Transcend. But they're targeted for industrial applications and use relatively low performance SATA controllers.

Because this is supposed to faster and have much longer operational life than a NAND-flash based SSD used for the same purpose.

Back when Intel first announced their crosspoint memory they predicted that it would have a much longer operational life than traditional SSDs. Now that they've finally managed to create a working product with it they're only saying that it has a lifetime a few times longer than SLC NAND.

Since write endurance is spread over the whole SSD you can easily get a cache four as big as the Optane equivalent for I presume the same amount of money and end up with the same endurance, higher throughput, slightly more latency, and a much higher cache hit rate.

That's actually great, because most drives these days are the much lower endurance MLC, rather than SLC. So we get much higher density AND higher endurance, for less money (assuming SLC is still pricey, I haven't looked recently).

well there are 1tb memory machines not sure if a 32 gb ssd cache would help.

The 32GB one from this article is the consumer product with the same Optane technology as the enterprise/datacenter one. The DC-P4800X (the datacenter one) has 375GB, 750GB, and 1.5TB variants[1].

So, no, if you have 1TB of memory a 32GB Optane SSD isn't going to make a difference, but if you have 1TB of memory, you're not the target for the 32GB consumer version, you're the target for the 375GB+ datacenter version.

[1]: http://www.anandtech.com/show/11209/intel-optane-ssd-dc-p480...

> you're the target for the 375GB+ datacenter version

Keeping in mind that that version costs $4/GB, which is about half the price of adding more RAM. Which means that while it can be good storage for ZFS or a database, it's still not a particularly effective cache in such a machine.

It costs $4 per GB today. In 6 months it will probably be $3, and in a few years it will be $2. Meanwhile RAM will still be $8 per GB or more. Also 64GB dimms are around $13 per GB, and have been around that price for awhile.

well at 1tb memory there aren't that many dimm slots free and some processors also don't support that much more.

but still what I wanted to say is that it is prolly more useful as a ssd read cache. and not to have it behind your big memory.

As xoa pointed out, it's very well suited for use as a persistent cache.

I don't quite understand a few things about this review. One is that they're using a 960 EVO 250GB as their benchmark SSD, but it's not the best SSD, the 960 PRO 2TB is. How does it compare to that.

The other thing I don't get is that Intel wants to sell Optane drives to "eSports gamers". Why would they need that? I've heard people on the Overwatch forums complaining about slow map loading from laptop hard drives... but if you're playing organized eSports, they're not going to start the match without you. Matches are paused mid-match all the time for technical glitches. If you're just playing on your machine at home, that's not really eSports, that's just playing computer games at home, perhaps with a higher than average dedication to winning. And map loading time is miniscule with SSDs, so I don't see the advantage you could possibly derive from there. (It all fits in RAM anyway, which is cheap.)

I dunno, I kind of see this whole thing as a solution without a problem, at least in the consumer market.

Good point, eSports likely won't really benefit from Optane that much over a 960 EVO or PRO, but if they get into the eSports scene early they'll get a lot of free advertisement from all the various CS GO / LOL / Overwatch streamers who tend to play many other games that might benefit from it.

I only had a few days to test for this initial review, so I picked flash SSDs that are comparable in price to the Optane+HDD cached setup rather than the best flash SSD money can buy. The 525GB Crucial MX300 and the 250GB Samsung 960 EVO are both close to but slightly cheaper than the 32GB Optane Memory plus the 1TB hard drive.

Intel knows that Optane caching isn't an attractive high-end solution; they're not trying to pitch it against the 2TB Samsung 960 PRO. They'll be coming out with higher capacity Optane SSDs for primary storage use later this year, though still sacrificing a lot of capacity for the improved performance relative to NAND flash.

With this kind of endurance and $/GB, there is no reason at this time to buy Optane over more RAM, NAND SSDs, and battery backups.

Maybe the next generation of Optane will be worth it, but this generation certainly is not.

I wonder, what would be better for desktop: traditional SSD or Intel Optane SSD?

Define 'better'. In terms of performance Optane absolutely wins. In terms of performance per dollar it's hard to justify.

Better in terms of visible performance of typical tasks (development, browsing, gaming). I thought of assembling new PC this year, I wanted to go with Samsung Pro SSD (something like 512G), but now I'm considering to wait for this new tech.

Reply on news.ycombinator.com