The Samsung 950 Pro PCIe SSD Review(anandtech.com) |
The Samsung 950 Pro PCIe SSD Review(anandtech.com) |
Since coreboot (with TianoCore or SeaBIOS) and the impressive nouveau and radeon projects, there's fewer places binary blobs can hide on a modern computing system.
One of the bigger exceptions is the drive controllers (which processes AHCI/NVMe messages and things like wear levelling). Besides restrict freedom and performance experimentation, it's also a security issue, given the scope for man-in-the-middle attacks [1]
Now at the start of NVMe's reign, it's probably an especially good time to start a project to make an open-source replacement to the proprietary firmware blobs of SSDs/flash memory! :)
I mean, on the 'remapping bad sectors' bit, right now? to the OS, it looks like you are doing a sequential read, but if there was a remapped sector in there? you're doing a random read. We'd be much better off, at least on server systems with decent raid subsystems, just handling the bad sectors in software.
Right now, I pay almost double for my spinning disk because I want slightly better firmware that is designed to fail outright rather than retry, because all my disks are in raid. some consumer disks allow you to adjust the time-limited error recovery paramiters, but in my experience, it's super unreliable.
Interesting article: http://www.tomshardware.co.uk/intel-micron-3d-xpoint-updates...
Take a look at any of the recent reviews comparing drives with very similar flash hardware, yet with wildly different performance characteristics. Firmware matters.
I'd certainly like to see such devices with Open Source firmware, but don't discount the difficulty of the problem.
So how come, that Samsung SM951 is much much cheaper, yet just as fast, or even faster in every metric? The only thing that is slower, is the read speed(2150MB/s vs 2200MB/s). Better write speed, more IOPS, and cheaper.....so how is the 950 Pro a flagship?
SM951:
http://www.scan.co.uk/products/256gb-samsung-sm951-m2-(22x80...
950 Pro:
http://www.scan.co.uk/products/256gb-samsung-950-pro-m2-(22x...
The 512GB 950 Pro is $349, 512GB SM951 is as much as $490 if you can manage to find it. The 256GB SM951 is more available, but it's the same price as the 950 Pro at the very best.
And that's before you start to see sales on the 950 Pro, which you will definitely see.
SM951 is an OEM product and you're generally hard pressed to find it available as a consumer in any substantial quantity. The 950 Pro is a mass market consumer product.
Just look at Amazon, Newegg, etc.
In other words, if your workload isn't super heavy, this is quite fast - but then, why the heck are you buying this expensive, low capacity power hog if your workload is almost certainly fast enough on a slower drive?
At the very least, this drive should be beating or equaling samsung's own SATA drives reliably - but it's not. The power consumption is also worrying, especially due to the tiny form factor...
http://arstechnica.com/gadgets/2015/08/samsung-unveils-2-5-i...
Their 4TB models are around $1/gb mark; the higher capacity models are much lower $/gb. So far we've only tested the 2.5" models and haven't had a chance to play with the 8/16TB PCIe add-in cards.
Here are some benchmarks: http://techreport.com/review/29221/samsung-950-pro-512gb-ssd...
SuperMicro is selling chassis that have a mix of NVMe and regular sata ports in form factors similar to the traditional 2.5" sata hot swap bays, which is awesome... in my application, at least, it's not acceptable to have to unrack and take apart a server when a drive fails, so without that part (without chassis that have externally accessible hot-swap bays for NVMe) it's pretty much a no-go.
But all the pieces are here. I just need to get together the cash (see aformentioned dayjob) and then figure out what I've gotta buy quality wise. (the samsung pros are on the list, and of course, the nicer Intel stuff) - SSDs really are quite different from spinning disk in many ways; especially backups without zfs. just doing a snapshot and full dd-image of a fast ssd like this would probably be acceptable once a day; something I wouldn't dream of with spinning disk.
For some background, I've fallen way behind my competition, and I'm still entirely spinning disk while most of them are entirely ssd. I'm most of the way to matching them on ram/disk per dollar at this point (not on the front page, but most of my customers are upgraded and the rest will be done soonish) but even then, I'm on spinning disk and they are on ssd.
UBIfs seems to solve that pretty well for me.
"NVM Express: Linux driver support decoded" (2015.06.30)
https://communities.intel.com/community/itpeernetwork/blog/2...
Really exciting to see changes like these come. I can't wait to start adopting these into my PC builds. I might just try one for a new HTPC I am building (OS drive).
Why not 8000 ?
The 850 pro has a 6000 "NAND P/E Cycles"[1] ( P/E = program/erase cycles ) and rated at 300TB ( 512Gb model )
[1] http://www.anandtech.com/print/8239/update-on-samsung-850-pr...
The problems I have with stock firmware isn't that the stock firmware isn't good... it's that the stock firmware hides a lot of information that the rest of my system could use to make the access to the disk more sequential. It also makes the assumption that I want this drive to function by itself without complex zfs-style software on top of it that can handle errors.
The idea being that the drives are setup with a firmware that makes the drives okay in a configuration where it's a single drive running ms-dos. In a multi-drive, redundant configuration with something like zfs doing the raid, a lot of the "features" there (retrying for a long time on errors, remapping bad sectors) are problems that are handled far better at a higher level.
So... if your point is that I could just spend the money and get a storage appliance from emc or netapp, you are right, but those things cost on the order of 10x what local storage costs.
If your point is just that I'm being cheap, that's fair too, and that's probably why drive makers don't want open-source firmware; with closed source firmware, they can charge extra for the feature-set that is optimal for a RAID configuration, and less for the feature-set that is optimal for a single-disk configuration. Even though both of those feature sets might be pretty good for what they do, they are largely the same drive, physically, and even if they weren't, I'd be happy to have a somewhat less-reliable drive that reliably returned errors rather than hanging, if it saved me the rather large difference in price between 'consumer' and 'enterprise' drives. - right now, I'm paying the extra for the 'enterprise' drives even thought the overall reliability of consumer drives would be just fine for my use.
Open firmware would be great for various other reasons like being able to have confidence in a secure erase solution or being able to have better telemetry to predict failures. That said, I wouldn't put some huge secret performance gains at the top of that list.
the remapping I'm talking about happens when a sector goes bad. if sector 5 fails, it's remapped to the bit of disk set aside as spares, no? no secret there. but when my computer reads sector 1 through sector 10, my computer thinks it's reading sequentially... but really there's a big random seek right in the middle of it. By my understanding, that happens every time I go to read sector 5.
>Given how much the mechanical disk vendors are threatened by solid state technology, if there were a simple way to increase performance on drives, I'm sure they would provide it at this point. Even if they didn't do it for retail, their big customers would have access to it. I'm not aware of anything of the sort, so I doubt this is as big an issue as you claim.
The point here is that it's not simple, and it's not a conspiracy or anything... and a lot of this wouldn't be a win for computers running windows with a single disk, and the big customers do have access to this sort of thing; like I said, emc and netapp and the other big vendors have their own custom firmware.
My guess is that the vendors don't want to do anything that requires the user to run special software, and all the changes I'm talking about would end very badly for the user if they weren't running a filesystem/raid system that was built to handle that sort of thing.
On the remapping of sectors, that happens very infrequently. Look up smart reallocation count indicators for more info on this. Once a few percentage points of space has been reallocation, smart will assume the drive is failing.
Intel is also moving ahead with their Infiniband implementation as the "supercomputing" cross-chassis bus (QPI is a clone of Hypertransport, but Intel decided not to use it for supercomputing bus fabrics like AMD does with their externally switched Hypertransport implementation), so future Xeons and future Phis will have on-die True Scale controllers.