[LU-16017] Suboptimal dnode size used for ZFS

• Type: Bug
• Resolution: Unresolved
• Priority: Minor
• Affects Version/s: Upstream, Lustre 2.15.0
• Labels: llnl zfs
• Environment: Any Lustre filesystem using ZFS-based servers.

Epic/Theme

zfs

• Severity: 3

Epic

metadata performance server zfs

• Assignee: Brian Behlendorf
• Reporter: Brian Behlendorf

Description

When allocating new objects for a ZFS-based server the dnode should be sized large enough to store all of the base Lustre extended attributes (trusted.lma, trusted.fid, trusted.version) in order to avoid requiring a spill block.  In practice, this means a minimum size of 1K is required to accommodate the packed xattr nvlist in the bonus area.

> zdb -e -p /tmp/ -dddd lustre-ost1/ost1 1500

Dataset lustre-ost2/ost2 [ZPL], ID 391, cr_txg 8, 18.8M, 614 objects, rootbp DVA[0]=<0:16440a00:200> DVA[1]=<0:ada00:200> [L0 DMU objset] fletcher4 lz4 unencrypted LE contiguous unique double size=1000L/200P birth=89L/89P fill=614 cksum=dd5e644f1:50403505c9e:f29fa6075ec6:1fd1540cabc5f5

    Object  lvl   iblk   dblk  dsize  dnsize  lsize   %full  type
      1500    1   128K    64K    64K      1K    64K  100.00  ZFS plain file
                                               356   bonus  System attributes
        dnode flags: USED_BYTES USERUSED_ACCOUNTED USEROBJUSED_ACCOUNTED
        ...

        SA xattrs: 204 bytes, 3 entries

                trusted.lma = \010\000\000\000\000\000\000\000\000\000\001\000\001\000\000\000\360\000\000\000\000\000\000\000
                trusted.fid = \001\004\000\000\002\000\000\000\335\001\000\000\000\000\000\000\000\000\020\000\001\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000
                trusted.version = \357\000\000\000\001\000\000\000

However, by default 512b dnodes are created forcing a spill block to be allocated for each object which significantly increases the required storage and reduces performance due to the additional I/O.

It appears the issue is caused by OSD_BASE_EA_IN_BONUS being set incorrectly.  This value needs to account not just for the data size (which it does), but also for the xattr key text and XDR encoding overhead for the packed nvlist.  After everything is taken in to account the correct size is 204 bytes according to zdb.

It seems to me the correct fix here is to update OSD_BASE_EA_IN_BONUS accordingly.  Making the follow change fixes the issue in my testing.  It'd be great if someone else could review this change to make sure additional changes aren't needed.

+/*
+ * The base extended attribute SA size including the keys, values,
+ * and XDR encoding overhead as reported by zdb.
+ *
+ * SA xattrs: 204 bytes, 3 entries
+ *     trusted.lma = \000\000\000\000...
+ *     trusted.fid = \000\000\000\000...
+ *     trusted.version = \000\000\000...
+ */
+#define OSD_BASE_EA_IN_BONUS   (ZFS_SA_BASE_ATTR_SIZE + 204)

Until a fix is merged a reasonable workaround for this is to explicitly set the dnodesize property for the ZFS datasets to 1K on pools containing OSTs, and possibly larger on pools containing MDTs.

zfs set dnodesize=1k pool/ost

Comments

[Olaf Faaland]

Cory and Alexander, sorry that took so long.  Reviews appreciated.

[Gerrit Updater]

"Olaf Faaland <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/48646
Subject: LU-16017 osd-zfs: OSD_BASE_EA_IN_BONUS should include names
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 010a36889fc430d498b3b0dfb41838aaf8aae024

[Cory Spitz]

ofaaland, were you still planning to push a patch?

[Olaf Faaland]

Hi Alexander, I took this over from Brian and haven't pushed a patch yet. Probably later today.

[Alexander Boyko]

Hi behlendorf, I see that you was going to prepare a fix. Was it pushed ? I don't see any link here.
Thanks.

[Olaf Faaland]

For my reference, our local ticket is TOSS5732

[Brian Behlendorf]

Right, I'll push a proper patch to for review in the next couple of days.  I'm happy to structure it as you suggested.

[Andreas Dilger]

Brian, it would probably be best for review if you pushed a patch that embodied your change.

That said, it would be preferable IMHO if the change used "sizeof(foo)" instead of a fixed number, possibly with a fudge-factor (2x or 7/4 or whatever) to compensate for encoding and other overhead.

[Brian Behlendorf]

Assuming a 1k dnode it looks like we've got  476 byes of bonus space still available for an OST regular file object and 336 bytes available for an MDT regular file object.  How much space does a PFL layout typically need?

Then it sounds like 1k dnodes are sized reasonably for the OSTs to avoid wasting space, but we probably want to go larger by default of the MDTs.  The code automatically sizes the dnode based on the expected xattr size so it seems we'll need to teach it about those additional MDT xattrs.