In April I did some research on why synchronous NFSv3 performance on XFS is so
rotten when compared to local filesystem performance. The workload I chose to
work with is tar.
After taking some measurements I came to the conclusion that one of the big
problems is that we're not treating the log as stable storage. By calling
write_inode_now() we've written the changes to the log first and then gone and
also written them out to the inode on disk.
In a short discussion of this issue on the xfs-oss list it was suggested that I
post the patches here for discussion.
The following series is adds a 'wsync' export option to nfsd. It is intended
to be used on XFS with the wsync mount option. When you already have a
synchronous log there is no need to sync metadata separately.
This is barely tested, YMMV, I could have this all wrong, etc, etc. Here are
some very unscientific measurements taken over gigabit ethernet.
# time tar -xvf /mnt2/quilt-0.47.tar > /dev/null
No XFS wsync, no NFS wsync:
0m13.177s 0m13.301s 0m13.528s
XFS wsync set, no NFS wsync:
0m13.019s 0m13.232s 0m13.094s
XFS wsync set, NFS wsync set:
0m8.361s 0m8.400s 0m8.301s
Curious to hear if this is a reasonable thing to do. Suggestions welcome.
Thanks!
Ben
---
Ben Myers (4):
Add 'wsync' export option.
Add datasync argument to nfsd_sync_dir().
If 'wsync' call vfs_fsync() instead of write_inode_now().
If 'wsync' pass datasync=1 to vfs_fsync().
fs/nfsd/export.c | 1 +
fs/nfsd/nfs4recover.c | 2 +
fs/nfsd/vfs.c | 62 +++++++++++++++++++++++++++++++------------
fs/nfsd/vfs.h | 2 +
include/linux/nfsd/export.h | 4 ++-
5 files changed, 50 insertions(+), 21 deletions(-)
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
