Recordsize/ahift: in both cases default values were used (but on different
versions of Lustre). How can I check different values for recordsize/ashift for
actual values to compare?
zpool mirroring is quite different though – bad drive is a simple raidz2:
raidz2-0 ONLINE 0 0 0
sdd ONLINE 0 0 0
….
errors: No known data errors
the good drive uses 10 mirrors:
NAME STATE READ WRITE CKSUM
mdt0000 ONLINE 0 0 0
mirror-0 ONLINE 0 0 0
sdd ONLINE 0 0 0
sde ONLINE 0 0 0
mirror-1 ONLINE 0 0 0
sdf ONLINE 0 0 0
sdg ONLINE 0 0 0
mirror-2 ONLINE 0 0 0
sdh ONLINE 0 0 0
sdi ONLINE 0 0 0
mirror-3 ONLINE 0 0 0
sdj ONLINE 0 0 0
sdk ONLINE 0 0 0
mirror-4 ONLINE 0 0 0
sdl ONLINE 0 0 0
sdm ONLINE 0 0 0
mirror-5 ONLINE 0 0 0
sdn ONLINE 0 0 0
sdo ONLINE 0 0 0
mirror-6 ONLINE 0 0 0
sdp ONLINE 0 0 0
sdq ONLINE 0 0 0
mirror-7 ONLINE 0 0 0
sdr ONLINE 0 0 0
sds ONLINE 0 0 0
mirror-8 ONLINE 0 0 0
sdt ONLINE 0 0 0
sdu ONLINE 0 0 0
mirror-9 ONLINE 0 0 0
sdv ONLINE 0 0 0
sdw ONLINE 0 0 0
mirror-10 ONLINE 0 0 0
sdx ONLINE 0 0 0
sdy ONLINE 0 0 0
thanks
Michael
From: Andreas Dilger <[email protected]>
Sent: Monday, November 11, 2019 14:42
To: Hebenstreit, Michael <[email protected]>
Cc: Mohr Jr, Richard Frank <[email protected]>; [email protected]
Subject: Re: [lustre-discuss] changing inode size on MDT
There isn't really enough information to make any kind of real analysis.
My guess would be that you are using a larger ZFS recordsize or ashift on the
new filesystem, or the RAID config is different?
Cheers, Andreas
On Nov 7, 2019, at 08:45, Hebenstreit, Michael
<[email protected]<mailto:[email protected]>> wrote:
So we went ahead and used the FS – using rsync to duplicate the existing FS.
The inodes available on the NEW mdt (which is almost twice the size of the
second mdt) are dropping rapidly and are now LESS than on the smaller mdt (even
though the sync is only 90% complete). Both FS are running almost identical
Lustre 2.10. I cannot say anymore which ZFS version was used to format the good
FS.
Any ideas why those 2 MDTs behave so differently?
old GOOD FS:
# df -i
mgt/mgt 81718714 205 81718509 1% /lfs/lfsarc02/mgt
mdt0000/mdt0000 458995000 130510339 328484661 29% /lfs/lfsarc02/mdt
# df -h
mgt/mgt 427G 7.0M 427G 1% /lfs/lfsarc02/mgt
mdt0000/mdt0000 4.6T 1.4T 3.3T 29% /lfs/lfsarc02/mdt
# rpm -q -a | grep zfs
libzfs2-0.7.9-1.el7.x86_64
lustre-osd-zfs-mount-2.10.4-1.el7.x86_64
lustre-zfs-dkms-2.10.4-1.el7.noarch
zfs-0.7.9-1.el7.x86_64
zfs-dkms-0.7.9-1.el7.noarch
new BAD FS
# df -ih
mgt/mgt 83M 169 83M 1% /lfs/lfsarc01/mgt
mdt0000/mdt0000 297M 122M 175M 42% /lfs/lfsarc01/mdt
# df -h
mgt/mgt 427G 5.8M 427G 1% /lfs/lfsarc01/mgt
mdt0000/mdt0000 8.2T 3.4T 4.9T 41% /lfs/lfsarc01/mdt
# rpm -q -a | grep zfs
libzfs2-0.7.9-1.el7.x86_64
lustre-osd-zfs-mount-2.10.8-1.el7.x86_64
lustre-zfs-dkms-2.10.8-1.el7.noarch
zfs-0.7.9-1.el7.x86_64
zfs-dkms-0.7.9-1.el7.noarch
From: Andreas Dilger <[email protected]<mailto:[email protected]>>
Sent: Thursday, October 03, 2019 20:38
To: Hebenstreit, Michael
<[email protected]<mailto:[email protected]>>
Cc: Mohr Jr, Richard Frank <[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: Re: [lustre-discuss] changing inode size on MDT
On Oct 3, 2019, at 20:09, Hebenstreit, Michael
<[email protected]<mailto:[email protected]>> wrote:
So bottom line – don’t change the default values, it won’t get better?
Like I wrote previously, there *are* no default/tunable values to change for
ZFS. The tunables are only for ldiskfs, which statically allocates everything,
but is will cause problems if you guessed incorrectly at the instant you format
the filesystem.
The number reported by raw ZFS and by Lustre-on-ZFS is just an estimate, and
you will (essentially) run out of inodes once you run out of space on the MDT
or all OSTs. And I didn't say "it won't get better", actually I said the
estimate _will_ get better once you actually start using the filesystem.
If the (my estimate) 2-3B inodes on the MDT is insufficient, you can always add
another (presumably mirrored) VDEV to the MDT, or add a new MDT to the
filesystem to increase the number of inodes available.
Cheers, Andreas
From: Andreas Dilger <[email protected]<mailto:[email protected]>>
Sent: Thursday, October 03, 2019 19:38
To: Hebenstreit, Michael
<[email protected]<mailto:[email protected]>>
Cc: Mohr Jr, Richard Frank <[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: Re: [lustre-discuss] changing inode size on MDT
On Oct 3, 2019, at 05:03, Hebenstreit, Michael
<[email protected]<mailto:[email protected]>> wrote:
So you are saying on a zfs based Lustre there is no way to increase the number
of available inodes? I have 8TB MDT with roughly 17G inodes
[root@elfsa1m1 ~]# df -h
Filesystem Size Used Avail Use% Mounted on
mdt0000 8.3T 256K 8.3T 1% /mdt0000
[root@elfsa1m1 ~]# df -i
Filesystem Inodes IUsed IFree IUse% Mounted on
mdt0000 17678817874 6 17678817868 1% /mdt0000
For ZFS the only way to increase inodes on the *MDT* is to increase the size of
the MDT, though more on that below. Note that the "number of inodes" reported
by ZFS is an estimate based on the currently-allocated blocks and inodes (i.e.
bytes_per_inode_ratio = bytes_used / inodes_used, total inode estimate =
bytes_free / inode_ratio + inodes_used), which becomes more accurate as the MDT
becomes more full. With 17B inodes on a 8TB MDT that is an bytes-per-inode
ratio of 497, which is unrealistically low for Lustre since the MDT will always
stores multiple xattrs on each inode. Note that the filesystem only has 6
inodes allocated, so the ZFS total inodes estimate is unrealistically high and
will get better as more inodes are allocated in the filesystem.
Formating under Lustre 2.10.8
mkfs.lustre --mdt --backfstype=zfs --fsname=lfsarc01 --index=0
--mgsnid="36.101.92.22@tcp<mailto:36.101.92.22@tcp>" --reformat mdt0000/mdt0000
this translates to only 948M inodes on the Lustre FS.
[root@elfsa1m1 ~]# df -i
Filesystem Inodes IUsed IFree IUse% Mounted on
mdt0000 17678817874 6 17678817868 1% /mdt0000
mdt0000/mdt0000 948016092 263 948015829 1% /lfs/lfsarc01/mdt
[root@elfsa1m1 ~]# df -h
Filesystem Size Used Avail Use% Mounted on
mdt0000 8.3T 256K 8.3T 1% /mdt0000
mdt0000/mdt0000 8.2T 24M 8.2T 1% /lfs/lfsarc01/mdt
and there is no reasonable option to provide more file entries except for
adding another MDT?
The Lustre statfs code will weight in some initial estimates for the
bytes-per-inode ratio when computing the total inode estimate for the
filesystem. When the filesystem is nearly empty, as is the case here, then
those initial estimates will dominate, but once you've allocated a few thousand
inodes in the filesystem the actual values will dominate and you will have a
much more accurate number for the total inode count. This will probably be
more in the range of 2B-4B inodes in the end, unless you also use Data-on-MDT
(Lustre 2.11 and later) to store small files directly on the MDT.
You've also excluded the OST lines from the above output? For the Lustre
filesystem you (typically) also need at least one OST inode (object) for each
file in the filesystem, possibly more than one, so "df" of the Lustre
filesystem may also be limited by the number of inodes reported by the OSTs
(which may themselves depend on the average bytes-per-inode for files stored on
the OST). If you use Data-on-MDT and only have a small files, then no OST
object is needed for small files, but you consume correspondingly more space on
the MDT.
Cheers, Andreas
From: Andreas Dilger <[email protected]<mailto:[email protected]>>
Sent: Wednesday, October 02, 2019 18:49
To: Hebenstreit, Michael
<[email protected]<mailto:[email protected]>>
Cc: Mohr Jr, Richard Frank <[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: Re: [lustre-discuss] changing inode size on MDT
There are several confusing/misleading comments on this thread that need to be
clarified...
On Oct 2, 2019, at 13:45, Hebenstreit, Michael
<[email protected]<mailto:[email protected]>> wrote:
http://wiki.lustre.org/Lustre_Tuning#Number_of_Inodes_for_MDS
Note that I've updated this page to reflect current defaults. The Lustre
Operations Manual has a much better description of these parameters.
and I'd like to use --mkfsoptions='-i 1024' to have more inodes in the MDT. We
already run out of inodes on that FS (probably due to an ZFS bug in early IEEL
version) - so I'd like to increase #inodes if possible.
The "-i 1024" option (bytes-per-inode ratio) is only needed for ldiskfs since
it statically allocates the inodes at mkfs time, it is not relevant for ZFS
since ZFS dynamically allocates inodes and blocks as needed.
On Oct 2, 2019, at 14:00, Colin Faber
<[email protected]<mailto:[email protected]>> wrote:
With 1K inodes you won't have space to accommodate new features, IIRC the
current minimal limit on modern lustre is 2K now. If you're running out of MDT
space you might consider DNE and multiple MDT's to accommodate that larger name
space.
To clarify, since Lustre 2.10 any new ldiskfs MDT will allocate 1024 bytes for
the inode itself (-I 1024). That allows enough space *within* the inode to
efficiently store xattrs for more complex layouts (PFL, FLR, DoM). If xattrs
do not fit inside the inode itself then they will be stored in an external 4KB
inode block.
The MDT is formatted with a bytes-per-inode *ratio* of 2.5KB, which means
(approximately) one inode will be created for every 2.5kB of the total MDT
size. That 2.5KB of space includes the 1KB for the inode itself, plus space
for a directory entry (or multiple if hard-linked), extra xattrs, the journal
(up to 4GB for large MDTs), Lustre recovery logs, ChangeLogs, etc. Each
directory inode will have at least one 4KB block allocated.
So, it is _possible_ to reduce the inode *ratio* below 2.5KB if you know what
you are doing (e.g. 2KB/inode or 1.5KB/inode, this can be an arbitrary number
of bytes, it doesn't have to be an even multiple of anything) but it definitely
isn't possible to have 1KB inode size and 1KB per inode ratio, as there
wouldn't be *any* space left for directories, log files, journal, etc.
Cheers, Andreas
--
Andreas Dilger
Principal Lustre Architect
Whamcloud
Cheers, Andreas
--
Andreas Dilger
Principal Lustre Architect
Whamcloud
Cheers, Andreas
--
Andreas Dilger
Principal Lustre Architect
Whamcloud
_______________________________________________
lustre-discuss mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
