10

This has been asked a million times but every time the answer is subject to "it depends on your requirements" condition, so I cannot extract a general guideline to apply to my case. So I ask again.

I have a 24 bays disk server (dual Xeon Silver 4210R with 128 GB RAM and CentOS 7) and 16 TB disks to store scientific data, organized in large files (size ~ GB) which are typically written once and then processed many times (the output of this processing does not matter for what follows). Data is mission critical but to some extent recoverable from other storage sites, so a failure with data loss is a big issue but probably not a killer. Available disk space should be maxed within the previous constraints. To summarize, in order of decreasing importance my constraints are:

  1. data integrity
  2. read performance
  3. available disk space

My tentative solution is to use a hardware RAID 60 level with two RAID 6 arrays of 12 disks each, and the ZFS filesystem. In my poor understanding, RAID 60 should provide a more reliable and read-performant solution than RAID 6 with a reasonable loss in terms of available space, and ZFS is a good choice for a fault-tolerant filesystem. I have no clue about possible downsides of this configuration (e.g. array rebuild time? A different filesystem?) nor about possible better alternatives, so I'd like to hear some informed opinion.

Thanks in advance for any suggestion.

14
  • 2
    This is a great candidate for ZFS. Do you have the details of the specific hardware in use? Vendor/make/models.
    – ewwhite
    Jan 15, 2021 at 12:28
  • 1
    @ewwhite It's a Supermicro 6049P-E1CR24H with 24 Seagate ST16000NM002G, dual Xeon Silver 4210R and 128 GB RAM. Jan 15, 2021 at 13:36
  • 2
    If you pick ZFS, I would suggest benchmarking with the last few LTS kernel versions, as your hardware has many features not supported in the GA kernel of CentOS7, which went EOL 3 years ago. You may also be able to use compression to improve disk performance at the expense of cpu if you have enough bandwidth to the server to make use of it. What sort of bandwidth do you have to the server? Jan 17, 2021 at 9:55
  • 1
    @RichieFrame I thought about it but I have to integrate the disk server into a CentOS 7 computing farm. The IT guy won't be happy to see OS fragmentation. Jan 18, 2021 at 8:17
  • 1
    @warren No, I bought it from an Italian company selling Supermicro products on the national market. Jan 20, 2021 at 7:27

4 Answers 4

11

ZFS doesn't like to be on top of hardware RAID. Probably you might just use ZFS on raw disks, and configure it in raidz2 or raid60 mode. Also, probably it's good to have a replacement drive nearby, or even leave a hot spare(s) in the rack.

See the performance benchmarks here: https://calomel.org/zfs_raid_speed_capacity.html

8

For such a big setup (384 TB raw space) I strongly suggest using ZFS, as its data integrity (and repair) guarantees are simply too valuable to ignore.

If for "read performance" you mean sequential read speed, I would use a ZFS RAIDZ2 array configured with 2x 12-wide vdevs. Moreover, a large recordsize and lz4 compression should be two good choices. If going down that route, please keep in mind that it is generally better to avoid hardware RAID when using ZFS.

If you need high random read performance (unlikely, based on you description) you need to use smaller ZFS RAIDZ2 vdevs or even mirrors (if losing 50% free space is tolerable).

The non-ZFS alternative would be to use an hardware-based RAID60 array (having at least 2+ GB of powerloss-protected writeback cache) and a classical non-CoW filesystem (ie: XFS). In this case, you can use lvmthin as volume manager and snapshot layer. That said, go with ZFS if you can.

9
  • Each client will read mostly sequentially but simultaneous access from several clients effectively result in random reads, right? If true then I'd avoid over-optimizing for sequential reads. Also, I'm a noob and didn't know that ZFS dislikes hardware RAID; has the ZFS+RAIDZ2 some disadvantage against hardware RAID + different FS (e.g. slower performance being a software RAID?). Would it be better to use a JBOD card instead of a RAID controller to setup ZFS+RAIDZ2? Jan 15, 2021 at 13:34
  • @NicolaMori it is unlikely that multiple clients reading will turn large sequential reads in totally random, small ones. Moreover, ZFS has special provision for multi-stream prefetch reads. If your RAID controller supports pass-through operation, it should be fine with ZFS; otherwise, I suggest you to use a JBOD/HBA card. That said, performant storage at this scale is hard; you should really take your time to learn and experiment.
    – shodanshok
    Jan 15, 2021 at 13:42
  • I'm not after the ultimate performance, just trying to avoid trivial errors and ending up with a largely under-performing setup, since I'm not an IT expert. I'll start by studying the ZFS+ZRAID2 setup as you suggested. Jan 15, 2021 at 13:52
  • It seems that my controller allows only for exposing disks as RAID. Would creating 24 single-disk RAID0 arrays do the trick for using ZFS? Jan 15, 2021 at 15:17
  • 1
    I would not use single-disk RAID0 arrays for simulating individual disk drives. However this is getting way off topic. I strongly suggest you to take your time to research, learn and experiment. If needed, talk to (or engage) a professional sysadmin.
    – shodanshok
    Jan 15, 2021 at 16:42
7

Another recommendation is to have your OS separate to your data disks.

That supermicro chassis has two additional slots in the rear for 2.5" SATA disks. These should be RAID1 and contain the OS and any swap. The 24 disks out the front should just be for data in whatever RAID array or ZFS setup you choose.

enter image description here

3
  • 1
    I know its not really an answer, but can't put a picture in a comment. Also this does separate the OS raid from the data, which simplifies things.
    – Criggie
    Jan 16, 2021 at 7:44
  • 1
    Agreed about not answering the question, but definitely agree on keeping data and processing on separate drives - looking at replacing my NAS later this year, but currently running everything via an external SSD (that's backed up as not RAID) specifically to solve this!
    – Rycochet
    Jan 17, 2021 at 9:51
  • 1
    Thanks, I already went down that way with a 240 GB SSD in the rear bay for the OS. Jan 18, 2021 at 8:19
1

I know you'd lose more capacity to parity but I'd personally go with R60 using 3 x 8 disk arrays, simply for the rebuild time, it won't benefit you in any way, but 12 x 16TB disks is a bit much for me personally - yes it'll work.

The other option given you want to use ZFS is to use ZRAID, I'm no expert but there are several here who are.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .