3 ms·
So, what does HN consider to be good brands to buy these days, when it comes to SSDs? Sorry for bluntly asking.
by amelius 3y ago
So, what does HN consider to be good brands to buy these days, when it comes to SSDs? Sorry for bluntly asking.
- whitepoplar 3y agoSamsung Pro-series NVMe seems to be liked quite a bit.
- formerly_proven 3y agoRecently had a bad firmware version causing orders of magnitude increased NAND wear. Apparently took Puget Systems bugging Samsung about it to get a fix issued. Nobody’s perfect.
- ls612 3y agoThe best case scenario is that the product ships with perfectly bug free software. The second best case is that when a major bug is discovered it gets fixed quickly which Samsung did. IMO that’s a bigger measure of a company’s quality, how they react when something inevitably goes wrong.
- CamperBob2 3y agoPuget Systems' handling of that issue earned them a $10K order from me after the HN story appeared. Super impressed with the customer focus exhibited by them.
- SomeoneFromCA 3y agoThey have problems in Linux with autotrim, to the extent Ubuntu put trimming script into cron, if it detects Samsung SSD when installed.
- whitepoplar 3y agoWhat's the best NVMe drive for linux these days? (Seamless firmware updates, autotrim, rock-solid reliability)
- ciupicri 3y agoBy the way, but I've discovered that Seagate has published the source code for openSeaChest [1]: > openSeaChest is a collection of comprehensive, easy-to-use command line diagnostic tools and programming libraries for storage devices that help you quickly determine the health and status of your storage product. The collection includes several tests that show device information, properties and settings. It includes several tests which may modify the storage product such as power management features or firmware download. It includes various commands to examine the physical media on your storage device. Close to 200 commands and sub-commands are available in the various openSeaChest utilities. These are described in more detail below. I also recommend reading the Arch Linux wiki [2]. I'd also have a look at the Linux kernel quirks [3][4]. [1]: https://github.com/Seagate/openSeaChest https://github.com/Seagate/openSeaChest [2]: https://wiki.archlinux.org/title/Solid_state_drive/NVMe#Firmware_update https://wiki.archlinux.org/title/Solid_state_drive/NVMe#Firm... [3]: https://www.kernel.org/doc/html/latest/nvme/feature-and-quirk-policy.html#quirks https://www.kernel.org/doc/html/latest/nvme/feature-and-quir... [4]: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/nvme/host/pci.c?h=v6.4#n3314 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/lin...
- feedsmgmt 3y agoThey have been having issues https://arstechnica.com/gadgets/2023/02/pc-maker-pulls-samsung-pro-ssds-after-users-report-abnormal-health-drops/amp/ https://arstechnica.com/gadgets/2023/02/pc-maker-pulls-samsu...
- jeffbee 3y agoDamn, it's incredible how much HN hates AMP, and how much better and faster this AMP presentation is compared to Ars' usual hostile layout.
- s3p 3y agoSamsung NVMe and Sabrent Rocket in my opinion. Price goes down every year too, which is a plus.
- xoa 3y agoAll the major manufacturers have had significant issues at one point or another particularly with consumer drives, and it seems to go through cycles as well so there is no satisfying perfect answer here imo. I also haven't seen as much hard data on SSDs and they've been evolving so fast, so everything feels a lot murkier. At least with spinning rust we have Backblaze's drive stats going back many years now. - If it's a single drive situation and reliability is very important, something aimed at "enterprise" tends to be a good bet. More firmware testing, more over provisioning, and so on. And if nothing else there tends to be less BS around getting warranty coverage honored. Also costs significantly more. Micron, Intel D3s etc. - If it's multi-drive, don't focus on finding the one best brand, use ones from different major brands with redundancy. The goal isn't to try to ensure no drive failures, but rather lower the risk of correlated drive failures so that when they happen the system has time to restore redundancy and there's no issue. - If it's single drive and some down time is acceptable for cost then just roll the dice, though even there I'd stay away from QLC and the ultra mega bargain basement stuff. In all cases buy from somewhere reputable that actually sells drives directly, like B&H or CDW or the like. If you buy from Amazon beware of counterfeits and also "new" stuff that is heavily used. Always check SMART stats right away and make sure drive lifetime is where you expect (not hundreds/thousands of hours on something "new"). And of course, always have good backups and/or some way to rapidly bring stuff back up.
- everfree 3y agoWhy stay away from QLC? Does it not have its place in high-read low-write consumer workloads, such as software or media libraries?
- xoa 3y ago>Why stay away from QLC? Same basic idea as SMR hard drives: the savings are too minor to be worth the asterisks. If you're running at megascale and can afford to customize your workloads, software, OS, and firmware, around it and at scale the ROI crunches out as worth it that's fine of course. But at the individual or even SMB level? Nah. Like, taking Samsung as a fairly large, decent brand example, their 4TB TLC EVO drive is ~$220 right now, whereas their 4TB QLC QVO is... $200. Heck for whatever reason their 1TB at B&H costs (a very little) more, $65 for the QVO, while the EVO is $60. Is $5-20 over terabytes and years worth the slightest hassle or weirdness? QLC is IMO essentially a scam at the consumer level, dramatically impacting reliability/flexibility/performance consistency for really minimal cost savings. >Does it not have its place in high-read low-write consumer workloads, such as software or media libraries? Frankly no. I mean, putting aside whether SSDs themselves make sense in such an application at all vs spinning rust (same money will buy you a nice 14TB hard drive), but to the extent they do I still think the above holds. QLC is just physically fundamentally sketchier tech. And how sure are you (and how much trouble do you want to go through) about what your workload will be like long term? If you put the money/time investing into getting together a decent little SSD NAS with a bunch of terabytes, and then a year or two down the road want to have some VMs on it too, or have a fileshare and then it gets hit harder than you expected, or there's some bug that causes some write amplification, and then your QLC drives end up hosed was that really worth a few cups of coffee? Maybe so! SMR really is worth it to a few places as well. Or ZFS dedup for a software example (with hardware implications). But all of this for me falls under the heading of "if you don't have a specific reason for it, don't".
- creer 3y agoWe still hear comments about specific brands and models but the lessons on these are always learned after months or years of the devices being sold. We can still ask the question, sure. But a more practical lesson seems to go for redundancy because this will still happen to us: Not purchasing all items from one manufacturing lot; Not pairing more items from one lot than you can afford to loose; Staggering device lifetime when possible; Running long tests when putting in service; Keeping an eye on drive statistics once in production and acting on them; Running schemes (RAID, backup, whatever) where it's actually practical to replace one device while in operation; Testing schemes so they will actually work out; Effective backups; Being ready to lose the cost of occasional drives; etc.
- zerocrates 3y agoI've been quite happy with my SK Hynix NVME drives. But I think to a certain extent it's a bit of a crapshoot as to whether you're hitting a manufacturer while they have an undiscovered issue. I bought a whole series of Samsung drives in both NVME and SATA and they've all been perfect but this just all happened to be before their more recent issues. Before that I bought a bunch of Intels that all died in annoying ways (I believe they turned out to have a pretty bad write amplification bug tied to their power saving behavior).
- tagyro 3y agoMight not be the answer you want, but I found it easier to not trust any brand and instead opt for a mirror-RAID approach. Yes, it's more expensive (and still has its own risks) but (IMO) it's better than trusting any manufacturer.
- momirlan 3y agoRAID5 for me. i had 3 disks die over the years, but no data loss.
- linsomniac 3y agoIn production we try to use only Intel, and have had good luck with them over the decades. I say "try" because Dell makes it hard to tell who makes the different drives, but sometimes you can track it down by the specs or part numbers. Our buying cycle a couple years ago we ended up with some "Toshiba THNSN81Q92CSE 1.92TB SATA eSSD" in two systems, and ran into some severe performance issues, system IO wait time going through the roof. Ended up spending significant time with Dell support, them identifying a couple drives in each box to replace, and also doing full "RAID chain" replacements (controller, cabling, backplane). All of the replacement drives we've gotten have been Intel 1.92TB, so maybe Dell is seeing problems with Toshiba? Since then the systems have been solid, so maybe it was part of the RAID chain. Aside: Decades ago, right when Intel was coming out with their SSDs, I had a consulting client that was a Rural ISP. Their NFS server was on 8x 10K WD hard drives in RAID-10. The continually struggled with performance, and we kept tuning, but would regularly just outrun the IOPS, mostly with e-mail access. A friend of mine happened to have a 600GB Intel SSD production sample, so I asked if I could borrow it for a few weeks. Sent this laptop SSD that would fit in the palm of my hand out to this client, had them hook it up to the NFS server, and then did "pvmove" from the RAID array over to the laptop drive. Steadily, over the next few hours, the performance issues were gone. It was pretty fantastic to see an 8 drive RAID array bested by a laptop form factor drive.
- jeffbee 3y agoThis can hardly have been a surprise. Intel's very first mass market SSD from 15 years ago had enough IOPS to embarrass a whole room full of HDDs, and unless you know exactly what you are doing, a Linux NFS server will not benefit from any width of RAID-0 due to concentrating all filesystem metadata load on a single disk or pair of disks in the array.
- dehrmann 3y agoWith only a handful of brands out there, you're bound to get burnt by all of them sooner or later.
- _cenw 3y agoDon't have one to recommend for NVMe. For SATA though... we have several Crucial MX500s 1TB that are approaching 2PB written, and none have failed. It's a shame this didn't transfer to their NVMe SSDs- they are all pretty harsh underperformers (though none broke).