Universal enterprise, reliable SSD brands?

[lapsio]

Hello

I'm quite new to enterprise grade hardware so I'd like to ask about brands that produce really high quality SSD drives as I don't know any of them.

I'm running DIY Linux server with 20TB software RAID6 attached to 128gb SSD cache. Unfortunately I thought Samsung PRO series is kind of semi-enterprise grade but it seems it was terrible misconception. Luckily this was my first SSD ever and I heard they're usually faulty so i didn't really trust it in first place and used cache in write-through mode but still it failed like bilion of times with I/O errors and required array remount at least once a week and cache zero-ing up to point I removed cache at all. I was experiencing exactly the same failures with EVO series but I don't have any of those issues with budget GoodRAM CX100 so it's rather not software related issue. All SSDs tested in the same setup.

So this time I'd like to get something that will work. It doesn't need to be speed daemon but preferably something faster than CX100, But priority is reliability. I thought about HP drives but they seem to be dedicated to HP servers? I'm not sure. What are respected brands of server grade SSDs? Probably I'd be interested in eMLC or even SLC. Compatible with SATA 6. I probably wouldn't like to pay more than 250$ for 128gb though.

 

[ttabbal]

I would use an Intel S3500 or S3700.

 

[T_Minus]

Intel
Hitachi
Sandisk
Samsung's latest Enterprise SSD are much lower latency than their previous gens and are looking to be contenders moving forward, esp. for price.

Check out the "GREAT DEALS" section and pickup Intel S3710 400GB for $200 or less
You'll be happy.

 

[lapsio]

Yes, i know Samsung SSDs have awesome performance, especially for those 40 minutes before they crash on i/o error

How about Intel S3610 100gb? It sounds really affordable and still better than S3510 which from specs looks a bit like another 850 PRO...

Shoot, there's no S3710 100gb. 200gb version is quite expensive at least in Poland (around 250$) and I don't think I need 200gb just for cache.

Is there big difference between 3610 and 3710? Intel says 3710 is for write intensive tasks. I'm having quite high cache hit rate so I wouldn't say in my case caching is write intensive. If 3610 is similar quality to 3710 I think I'd go for it. If not then it's hard decision

My problem is not really performance but that 850 PRO and EVO just don't work, they constantly crash

 

[keybored]

...
Is there big difference between 3610 and 3710? Intel says 3710 is for write intensive tasks. I'm having quite high cache hit rate so I wouldn't say in my case caching is write intensive. If 3610 is similar quality to 3710 I think I'd go for it. If not then it's hard decision ...

Per Anandtech, Intel uses the same type of NAND in the 3610 as in the 3710, but it's binned lower: Intel Launches SSD DC S3610 & S3710 Enterprise SSDs
3610 also posts lower sequential write and 4k write numbers, but if your workload is mostly reads then you won't see much of a difference.

My problem is not really performance but that 850 PRO and EVO just don't work, they constantly crash

Make sure it's not this issue somehow: When Solid State Drives are not that solid
I'd think that it should be patched by now, but who knows.

 

[lapsio]

I know about it. I have disabled TRIM on data partitions. Well in fact just didn't enable it explicitly as LVM with encryption has disabled TRIM by default due to privacy concerns. Also i think with dm-cache you can't really enable TRIM as there's no filesystem here.

To provide more background I'm using btrfs filesystem. Crashes occur during any more intensive i/o especially scrub (filesystem check). Performing filesystem check on EVO 1tb in other machine is almost impossible due to i/o errors interrupting process. Basically i/o errors everywhere. It's definitely not something difficult to reproduce I can purposely crash Samsung SSD in 10 minutes if someone asked me for that. In all of my PCs using Samsung SSDs.

 

[EffrafaxOfWug]

I don't think there's any such thing as a universally reliable SSD brand - everyone's had at least one turkey I think

I assume you're also using mdadm + lvm/dm-cache...? I use dirt-cheap 240GB half-enterprise Intel 3500's for my cache drives (they house the OS as well) and despite being slower on paper than the Crucial M500's that preceded them (OS doesn't need any speed at all TBH so this not a problem for me), they hold up under sustained IO loads much better. The S37xx's will be even faster and AFAICT just as reliable, they're just a fair bit pricier than the plodding S3500.

 

[keybored]

...
It's definitely not something difficult to reproduce I can purposely crash Samsung SSD in 10 minutes if someone asked me for that. In all of my PCs using Samsung SSDs.

What sort of I/O errors do you get? Once you crash an SSD, does it self-recover once you stop accessing it or do you have to secure-erase it to bring it back up? Do these I/O errors translate to anything in SMART data?

 

[lapsio]

What sort of I/O errors do you get? Once you crash an SSD, does it self-recover once you stop accessing it or do you have to secure-erase it to bring it back up? Do these I/O errors translate to anything in SMART data?

I'm getting i/o error. kernel panic, then motherboard doesn't detect ssd, I need to physically unplug machine from power and later as btrfs didn't finish transactions i need to zero cache to avoid filesystem damage. Nothing special in SMART.

 

[keybored]

Anything special about the SATA controller you're using? Is it setup to use AHCI? If you're using an add-on controller card, you could try connecting your SSD to the mobo directly and see if your Samsung drives run better with that configuration. Or do the opposite if you're using the mobo controller, assuming you can borrow a SATA add-on card.

 

[lapsio]

Nope, nothing unusual just normal mobo controller. I was having issues with GRUB and /boot partition encryption support when using pci-e controller.

 

[Patrick]

Working very hard to get the STH recycled data center SSD data online today. Still a few graphs to generate and I have a lunch in 40 minutes so it will likely be later this afternoon.

 

[Patrick]

There you go: Used enterprise SSDs: Dissecting our production SSD population