Options for performance improvements on very big Filesystems and high IOWAIT

Question

I have a Ubuntu 16.04 Backup Server with 8x10TB HDD via a SATA 3.0 Backplane. The 8 Harddisks are assembled to a RAID6, an EXT4 Filesystem is in use. This Filesystem stores a huge amount of small files with very many SEEK operations but low IO throughput. In fact there are many small files from different servers which get snappshotted via rsnapshot every day (multiple INODES direct to the same files. I have a very poor performance since the file system (60TB net) exceeded 50% usage. At the moment, the usage is at 75% and a

du -sch /backup-root/

takes several days(!). The machine has 8 Cores and 16G of RAM. The RAM is totally utilized by the OS Filesystem Cache, 7 of 8 cores always idle because of IOWAIT.

Filesystem volume name:   <none>
Last mounted on:          /
Filesystem UUID:          5af205b0-d622-41dd-990e-b4d660c12bd9
Filesystem magic number:  0xEF53
Filesystem revision #:    1 (dynamic)
Filesystem features:      has_journal ext_attr dir_index filetype needs_recovery extent 64bit flex_bg sparse_super large_file huge_file uninit_bg dir_nlink extra_isize
Filesystem flags:         signed_directory_hash 
Default mount options:    user_xattr acl
Filesystem state:         clean
Errors behavior:          Continue
Filesystem OS type:       Linux
Inode count:              912203776
Block count:              14595257856
Reserved block count:     0
Free blocks:              4916228709
Free inodes:              793935052
First block:              0
Block size:               4096
Fragment size:            4096
Group descriptor size:    64
Blocks per group:         32768
Fragments per group:      32768
Inodes per group:         2048
Inode blocks per group:   128
RAID stride:              128
RAID stripe width:        768
Flex block group size:    16
Filesystem created:       Wed May 31 21:47:22 2017
Last mount time:          Sat Apr 14 18:48:25 2018
Last write time:          Sat Apr 14 18:48:18 2018
Mount count:              9
Maximum mount count:      -1
Last checked:             Wed May 31 21:47:22 2017
Check interval:           0 (<none>)
Lifetime writes:          152 TB
Reserved blocks uid:      0 (user root)
Reserved blocks gid:      0 (group root)
First inode:              11
Inode size:               256
Required extra isize:     28
Desired extra isize:      28
Journal inode:            8
First orphan inode:       513933330
Default directory hash:   half_md4
Directory Hash Seed:      5e822939-cb86-40b2-85bf-bf5844f82922
Journal backup:           inode blocks
Journal features:         journal_incompat_revoke journal_64bit
Journal size:             128M
Journal length:           32768
Journal sequence:         0x00c0b9d5
Journal start:            30179

I'm lacking experience with this kind of filesystem usage. What options do I have to tune this. What filesystem would perform better with this scenario? Are there any options to involve RAM for other caching options than the OS-build-in one?

How do You handle very large amounts of small files on large RAID assemblies?

Thanks, Sebastian

Answer 1

I have a similar (albeit smaller) setup, with 12x 2TB disks in a RAID6 array, used for the very same purpose (rsnapshot backup server).

First, it is perfectly normal for du -hs to take so much time on such a large, and used, filesystem. Moreover du accounts for hardlinks, which cause considerable and bursty CPU load in addition to the obvious IO load.

Your slowness is due to the filesystem metadata being located in very distant (in LBA terms) blocks, causing many seeks. As a normal 7.2K RPM disk provides about ~100 IOPS, you can see how hours, if not days, are needed to load all metadata.

Something you can try to (non-destructively) ameliorate the situation:

• be sure to not having mlocate/slocate indexing your /backup-root/ (you can use the prunefs facility to avoid that), or metadata cache trashing will severly impair your backup time;
• for the same reason, avoid running du on /backup-root/. If needed, run du only on the specific subfolder interested;
• lower vfs_cache_pressure from the default value (100) to a more conservative one (10 or 20). This will instruct the kernel to prefer metadata caching, rather than data caching; this should, in turn, speed up the rsnapshot/rsync discovery phase;
• you can try adding a writethrough metadata caching device, for example via lvmcache or bcache. This metadata device should obviously be an SSD;
• increase your available RAM.
• as you are using ext4, be aware of inode allocation issues (read here for an example). This is not directly correlated to performance, but it is an important factor when having so many files on an ext-based filesystem.

Other things you can try - but these are destructive operations:

• use XFS with both -ftype and -finobt option set;
• use ZFS on Linux (ZoL) with compressed ARC and primarycache=metadata setting (and, maybe, an L2ARC for read-only cache).

Answer 2

This Filesystem stores a huge amount of small files with very many SEEK operations but low IO throughput.

🎉

This is thing that catches lots of people nowadays. Alas, conventional FSes do not scale any well here. I can give you probably just a few advices when it comes to the set-up you already have: EXT4 over RAID-6 on HDDs:

1. Lower vm.vfs_cache_pressure down, say to 1. It'd change cacheing bias towards preserving more metadata (inode, dentry) instead of data itself and it should have positive effect in reducing number of seeks
2. Add more RAM. Although it might look strange for a server that doesn't run any piggy apps, remember: the only way to reduce seeks is to keep more metadata in faster storage, given that you have 16 GB only it seems that it should be relatively easy to increase the RAM amount
3. As I've said EXT4 isn't good choice for the use case you have, but still you can put in use some of the features it poses to soothe pain:
   • external journal is supported so you can try adding SSD (better mirrored) and place the journal there. Check out "ext4: external journal caveats"
   • Try switching journal mode to "all data's being journaled" mounting with data=journal
4. Try moving files outside of single FS scope. For e. g., if you have LVM-2 here you can create volumes of a lesser size and use them for a time being, then when it gets full, create another one and so on.
   • If you don't have LVM-2 you can try doing that with /dev/loop but it's not that convenient and probably less performant

UPD.: since it's turned out to be Linux Software RAID (LSR) RAID-6, here goes additional item:

5. LSR has own tuning options that many people seem to overlook
   • Stripe cache, that can be set thus to maximum: echo 32768 | sudo tee /sys/devices/virtual/block/md*/md/stripe_cache_size — But do this with care (use lesser value if needed) since the size is chunk-size multiple and depending on the chunk size you've chosen it would take different amount of RAM
   • External journal which can be also on those mirrored SSDs (but currently MD device created w/o journal can't be converted to use one).

— That's probably most of what can be improved w/o from scratch re-design.

I have a very poor performance since the file system (60TB net) exceeded 50% usage. At the moment, the usage is at 75%

That's very serious issue because that high disk space occupancy level only worsen fragmentation. And more fragmentation means more seeks. Wonder no longer why it gave more-or-less acceptable performance before reaching 50 %. Lots of manuals have clear recommendations to do not allow FSes grow up behind 75—80 %.

Answer 3

Thanks to everyone who gave an answer to my question.

This is, how I solved it:

First of all, I added the maximum amount of RAM to the board. Unfortunately, the Board only supports up to 64GB of RAM. I observed the behaviour after the expansion, and it was disappointing. Although all of the available RAM was used for IO Cache, the performance of the RSNAPSHOT-Backup didn't improve measurably.

So I had to pull the large mace. I added two 1TB NVME disks and assembled them to a RAID 1. The RAID 6 consisting of 8x 10TB HDDs got disassembled to one RAID 1 (consisting 2x 10TB HDD, ext4) and one RAID 5 (consisting 6x10TB HDD). The RAID 1 now contains the Operating System and the working copy of the Servers (which get rsynced 4 times a day to this drive).

The RAID5 is now a BCACHE backed device, backed by the NVME-RAID 1 and formatted with ext4. This drive contains the RSNAPSHOT-Copies. Every night, the files get rsynced from the RAID1 to the RAID5, which halves the IO-throughput of the RAID5 compared to the former RAID6, which contained the working copies AND the backup snapshots. Thanks to the BCache, not literally every single file gets written to the Disks, but all changes across one Block are written once, even if it contains several hunderths single file changes. This further decreased IOps on the HDDs.

Finally, I changed my RSnapshot configuration. Formerly, there were 31 daily snapshots and 18 monthly snapshots, which resulted in 49 backup generations. Now, I have the classical 7d / 4w / 12m / 1y-Design, which reduces the amount of backup generations down to 24.

After these changes (and with the 64GB RAM mentioned above), the duration for one snapshot got down from ~20hrs to 1.5 hours. The BCache devices has a Cache-Hit-rate of 82% (after 6 weeks of regular operation).

Mission accomplished. Thanks to all of you for your thoughts and input.

Answer 4

RAID6 does not help you much in this case, something like ZFS might enable much faster metadata and directory access while keeping speeds about the same.

Answer 5

RAID-6 stripes drives, therefore all IO goes to all drives. That's pretty inefficient with many small files. However this probably isn't your main problem which is...

Ext4 isn't well suited for big filesystems with millions of files. Use XFS. I have XFS filesystems running as big as 1,2 PB and with as many as 1 billion files, no problem. Simply use XFS.

Answer 6

I realize I'm late to this party, but if you're still using this setup, I don't see any mention of noatime (or relatime) being set on the ext4 backup file systems. If you don't have that set, the OS will update the access times on every file and its parent directory when its is touched which is a huge IO hit from all the file system metadata writes. Like if you're doing an incremental snapshot, the original file is checked (creating two writes), and the destination file is checked (two more writes) before its even written over if its changed when you don't have noatime set.

I hope this can still help for you or anyone else reading this.