The Best RAID for Small SSD Servers?

The Graid SupremeRAID SR-1001 is a high-performance, cost-effective RAID solution tailored for smaller NVMe SSD configurations. The SR-1001 demonstrates its ability to enhance storage potential better than software RAID and more cost-effectively than hardware RAID (faster too).
Full Review -
www.storagereview.com/review/...
SMCI Server Review -
www.storagereview.com/review/...
SR Discord -
/ discord
SR Merch -
store.storagereview.com
#storage #datacenter #raid

[@rugbyiloveyou]

Oh my, this RAID card is like magic! Finally found a solid way to keep my machine learning data safe!!!😍😍😍

[@AileenXu-uo2rj]

This is really interesting. It's a well-known fact that GPUs are faster than CPUs. Running a software RAID on a GPU, like Graid SupremeRAID does, totally makes sense. With its computing power, it can definitely perform better and provide more features compared to software RAID running on a CPU.

[@StorageReview]

Different tools for different workloads, but in this case, yes, offloading to a GPU is what makes the difference here, freeing up CPU cycles. We are working on another piece that looks at the CPU impact under extreme load.

[@samegoi]

I'm looking forward @@StorageReview

[@nadtz]

Not only is this an interesting raid solution but siena gets so little coverage even though it's pretty great for what it is meant for. Cool video, thanks!

[@StorageReview]

Siena is very cool - AMD didn't do much to push it since it's really a SKU for service providers and app owners. But we really like it for edge use cases where you need a good cores/power balance. With Graid in this case, it's a wicked little combo server.

[@udirt]

What i don't get is how we can't get somewhere without copying stuff around pcie, like, make that thing with nvme ports and you can drive 100 lane's worth over x16.

[@dancalmusic]

If I were to install Proxmox or TrueNAS on the server, choosing ZFS as the file system, would this RAID controller still be advisable? I know it's never a good idea to intermediate O/S and disks when using ZFS. Thank you!

[@joshhardin666]

if you were to install proxmox or truenas on the server, you wouldn't need graid at all, you'd just use zfs directly on your (enterprise) nvme and performance is generally great and zfs *does* do advanced data checking, closes the raid write hole, and does scrubbing. - it does use the cpu instead of an outboard accelerator, but modern systems, even something like the low-core epycs depicted, it's plenty performant, and instead of the graid controller, I would drop a sas HBA in a server like this, hook it up to a few jbods, and use this as a controller for a large hybrid storage array that would be extremely performant.

[@dancalmusic]

@@joshhardin666 Thanks

[@kelleyosburn4951]

Graid supports Proxmox and they recently posted a new Configuration guide. It is avaialble on the Graid web site site in their resource library. You can use ZFS as the file system and still take advantage of Graid for the high performance RAID protection.

[@michaelpettit1263]

Fast and loose with JBOD? With a software raid system like Storage Spaces, if a JBOD chassis or chassis component goes, the volumes simply go offline until you replace the JBOD chassis and filling it with the existing drives. Nothing happens to the drives. I'm assuming that using something like a Broadcom or Graid card, if more than the resiliency count goes offline at once it simply takes it offline instead of grinding the volumes into the pavement. The bigger problem we've had with Storage Spaces and SMC JBODs is that the individual SAS backplanes report as separate enclosures, and the rest of the JBOD components like the fans, temperature sensors and voltage regulators are unevenly split between them. Thus, Windows can't see them properly for enclosure awareness. We have to do that sort of by hand in the pool and vdisk creation scripts. That said, we have found other ways to get Storage Spaces to eat 200 drive pools because other drives died during the finicky drive replacement process. Thankfully, though, a whole JBOD going offline convinces it to simply take the pool and volumes offline.

[@StorageReview]

JBODs are fine, but this is really designed for flash, in which case, the JBOD doesn't really work any more...at least not very well with NVMe. Not impossible, we've done it...but not easy.

[@DataHoarders]

0:40 how many exabytes is a "bazillion Petabytes?" lol

[@StorageReview]

12 or so?

[@paulwais9219]

omg hardware raid is not dead! where is wendell ???

[@StorageReview]

Broadcom sells a lot of RAID cards for that part of the business to be dead...lol

[@fteoOpty64]

Wendell meant for consumer market. Enterprise always got all sorts of accelerator cards mainly due to huge number of PCieX16 slots and esoteric switches and bridges to aggregate them.

[@kriscyber]

RAID Hardware more powerful than RAID by software, and RAID very important for make sure your DATA and Compute system keep running.

[@keyboard_g]

20GB vs 2.1GB on writes mean for sure its skipping checks and susceptible to write holes in the raid?

[@leanderyu0819]

No RAID will skip parity on write, that's not RAID at all. There are plenty RAID solutions do more than 2.1GB/s on write. You don't assume them skip parity simply because they run faster than Linux MDADM. I believe StorageReview will not post this video if SupremeRAID didn't do RAID at all.

[@keyboard_g]

Didn’t say anything about skipping parity. Checks that the write made it through cache and landed on the storage media.

[@GuoFuTseng]

No parity can be skipped in RAID writes; the reason it can have this level of performance is due to its smart secret sauce. Regarding the write hole issue, the upcoming 1.6 release will have a patent-pending distributed journal feature implemented.

[@JordansTechJunk]

I take the other comment back. THIS is the most smooth brain comment on the video thus far.

[@GuoFuTseng]

@@keyboard_g The NVMe drive takes care of that part. The SupremeRAID solution handles the RAID operations (calculating parity, translating data location) without modifying or relocating user data for writing. The original user data address is passed to the NVMe drive using a standard NVMe write command. The NVMe drive must guarantee that the write is landed on persistent media before acknowledging the job as complete.

[@janliberda9493]

A wise man (Wendell from L1 tech) once said, it doesnt verify crap :) If I'm not mistaken, this is still true... (doesn't do read parity check, relies on error reporting by drives, unresolved raid write hole)

[@mspencerl87]

I recall the same. Sends it all like a firehose with no integrity, able to introduce corruption with no healing taking place. RIP GRAID

[@remomattei]

@@mspencerl87 well as of today no RAID solution does read the parity verifications. As of Graid, like many other RAID solutions out there in the market, they are doing consistency check in the background to do a bit of rot detection. You could check the related instruction user manual as well to verify that. Furthermore, in Wendell's video it compared a file system (ZFS ) and RAID for block device which his not an apple to apple comparison. File system feature the read parity, therefore, if you do really need that feature / option you can always run it on top of Graid SupremeRAID. Lastly, RAID write hole can be resolved if the RAID have the write-through enabled with the a journal file-system in an optimal state. In fact, some major brand like Broadcom 95xx is starting to introduce this into the market now. As of Graid Technology, with the new version released of 1.6 a new distributed journal feature has been released and it has been designed specifically for "DOUBLE FAILURE:, which means if you have unclean shutdown, while the system is recovering from a potential / possible write hole and there is a disk that failed you can still recover and this will guarantee all write hole use cases to be resolved. Keep in mind that this feature is not available or can be found in any traditional HW RAID solution at this time.

[@tnpaquette]

Wendell, a wise man? Two years ago he called out Graid and since he has ignored repeated offers from Graid to take a hands on looks at SupremeRAID or have a conversation with Graid directly. Would a "wise man" do that and not look to get the actual story? It appears that likes and clicks are more important than the truth...

[@keyboard_g]

Yes. Don’t trust the hardware to self report. Firmware at many levels pave over errors. 20GB vs 2.1GB on writes is for sure skipping checks. There is no magic in that old repurposed gpu.

[@leanderyu0819]

I think your information is out-of-date here. SupremeRAID supported background parity check like all RAID solution did 2 years ago. And it recently release v1.6 with distribute journal feature to resolve the data consistency issue during double failure. This resolve the issue without a single point of failure(journal device) like current solution in the market has.