All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
paper with benchmarks of new Intel Optane DIMMs
https://arxiv.org/abs/1903.05714
This is pretty impressive: Optane DIMMs are almost as fast as DRAM, have a mode where they can be used like DRAM, but are persistent. Unfortunately they are almost as expensive as DRAM ($6.57/GB in the 128GB modules and $10+/GB in 256GB modules, per Anandtech).
I think this has some significant consequences for fast databases: suddenly it becomes more important to have them local and in shared memory so they can be accessed with no context switches. There are probably some architectural improvements TBD, like combining them with hardware-assisted transactional memory (Intel TSX-NI) for shared memory concurrency.
OVH Labs has a super expensive server with Optane disks but those are just an unusually fast SSD. This DIMM stuff is really different. Its cost is at the upper end of what we normally encounter on LET but it potentially brings some new capabilities. @Hetzner_OL you might want to forward that paper to your colleagues if they haven't seen it.
Comments
Wow, this is a big game changer for database heavy sites. I can imagine a load balancing setup with a VM dedicated for database handling using this to tank requests for a super high traffic site at a low cost.
I'm not too impressed. Don't get me wrong, Optane's speed is impressive but once the first hype has cooled down, one sees that Optane actually isn't the wonder pill, or at least only in quite few use cases.
You kind of mentioned it yourself: For DBs one will want RAM and one will stay with it particularly with little price difference. The fact that the Dimm Optane also is persistent isn't that big a killer feature imo.
The really major jump has already happened (a) with SSDs and (b) with SSDs achieving a reasonably high life expectancy/durability.
From what I see what we really need now is is long term storage; the hard disk/direct access storage problem is solved, but the "where to reliably long-term store the massive amounts of data we generate" problem is not. LTO can't grow forever and it's not a reliable LT Storage anyway (in fact potentially less reliable than SSDs). What we'll need will be some non-magnetic (probably optical) LT storage for massive amounts of data (as in "50 TB per medium").
I was of course impressed by Optane. But I also see that it isn't that big a jump, it's expensive, and it's not yet well established/proven/trusted.
Hoping to see this as a commercial product in my lifetime: https://ark-invest.com/research/dna-storage-solution
This stuff seems to be much denser than ram. You can get it in 512GB sticks. Anandtech has been testing a server with 3TB of it installed. The persistence is relevant because servers do sometimes unexpectedly have to be rebooted. Though yeah, with an in-ram database you could probably log updates to a hard drive.
Yes our grandkids will be able to pick up 512GB Optane sticks for $150/each eventually. Just us early adopters pay the premium price LOL
It looks nice but I know my DBAs will never be happy.
Me: "Here's four 40Gbps links to our EMC XtremIO SANs."
DBAs: "How many PCI NVMe drives will fit in the server for caching?"
I guess, holographic (optical) storage will be available a lot sooner. Plus we already have a lot of technology and experience dealing with optical storage; with DNS-like storage not so much.
Absolutely. But keep in mind that the write load is usually very significantly lower than the read load (typ. in the 1:x hundred region) and for all but the most extreme profiles the difference between optane and SSD (think NVMe) read performance is neglegible (because processing requests and data and sending data out is much more expensive). We crossed into the sweet spot with SSD (as opposed to spindles).
Also keep in mind that 3TB is seriously much in terms of RAM, in terms of persistent storage, however, it is not and will often be way too small (just think of all the DBs with fat (basically file based) blobs attached to many/most rows. I've seen quite a few DBs in the 10s or 100s of GB range (the DB itself) but needing TBs for attached data).
Again, for some cases Optane will bring paradise, for most however the cost vs. performance ratio will kill wet dreams.
.> @KuJoe said:
I don't have the slightest idea what you are talking about * whistle *
But I suggest to have a good close look at upcoming AMD Epyc Rome systems (with plenty PCIe lanes ...)
We're virtualizing everything so local storage is a thing of the past for us. The DBAs are the last holdouts but now we can easily fit 512GB of RAM on a single blade so we're slowly converting them over.
I think blobs aren't that important for this purpose since they are opaque and can be off in secondary storage. The very high speed of ram or optane dimms is useful when you have to access a lot of different table fragments or indexes to handle a query and find the result set, or alternatively when you do an update you also have to change a lot of indexes. Blobs are for stuff you show after the result set is already known.
@jsg I'm curious to know if you have played with NVDIMM's and if you had any thoughts on it. Thanks!