XFS aka reflink store

[mtelvers]

Format and mount an XFS volume like this:

apt install xfsprogs
mkfs.xfs -f -d agcount=64 -i maxpct=90 /dev/nvme0n1p1
mount -t xfs -o noatime /dev/nvme0n1p1 /var/cache/obuilder

agcount increases the number of Allocation Groups, which should increase the concurrent performance. I have set this to the number of concurrent jobs the machine is servicing. The rule of thumb seems to be file system size in GB / 4. maxpct is the maximum percentage of the file system which can be used for inodes. Many inodes are needed as each duplicated file needs an inode.

Run ocluster-worker with --obuilder-store=xfs:/var/cache/obuilder

The result directory can be inspected using xfs_bmap -e -l -p -v -v -v <file> to check that the file blocks are, in fact shared:

$ xfs_bmap -e -l -p -v -v -v /var/cache/obuilder/result/*/rootfs/usr/bin/bash

 EXT: FILE-OFFSET      BLOCK-RANGE          AG AG-OFFSET        TOTAL FLAGS
   0: [0..2471]:       342178000..342180471 14 (312784..315255)  2472 100000
 FLAG Values:
    0100000 Shared extent

[tmcgilchrist]

Looks good. Thanks for adding the CI tests.

[tmcgilchrist]

Does anything need to happen here to check it has changed?

[tmcgilchrist]

Suggested change

	(** Store build results using rsync. *)
	(** Store builds results using XFS.
	XFS is intended to behave consistently as it scales to large storage and many files, modern-day XFS was originally from SGI Irix. This store uses the *reflink* feature in XFS to share blocks between files, to support fast snapshots of directory trees and deduplicate file data for more efficient use of storage hardware.
	For more details on the XFS implementation see https://blogs.oracle.com/linux/post/xfs-data-block-sharing-reflink and https://blogs.oracle.com/linux/post/upcoming-xfs-work-in-linux-v48-v49-and-v410-by-darrick-wong *)