On Wed, 6 May 2015, Zach Brown wrote: > On Wed, May 06, 2015 at 03:19:13PM -0700, Sage Weil wrote: > > On Wed, 6 May 2015, Trond Myklebust wrote: > > > Hi Zach, > > > > > > On Wed, May 6, 2015 at 6:00 PM, Zach Brown <zab@xxxxxxxxxx> wrote: > > > > > > > > Add the O_NOMTIME flag which prevents mtime from being updated which can > > > > greatly reduce the IO overhead of writes to allocated and initialized > > > > regions of files. > > > > > > > > ceph servers can have loads where they perform O_DIRECT overwrites of > > > > allocated file data and then sync to make sure that the O_DIRECT writes > > > > are flushed from write caches. If the writes dirty the inode with mtime > > > > updates then the syncs also write out the metadata needed to track the > > > > inodes which can add significant iop and latency overhead. > > > > > > > > The ceph servers don't use mtime at all. They're using the local file > > > > system as a backing store and any backups would be driven by their upper > > > > level ceph metadata. For ceph, slow IO from mtime updates in the file > > > > system is as daft as if we had block devices slowing down IO for > > > > per-block write timestamps that file systems never use. > > > > > > > > In simple tests a O_DIRECT|O_NOMTIME overwriting write followed by a > > > > sync went from 2 serial write round trips to 1 in XFS and from 4 serial > > > > IO round trips to 1 in ext4. > > > > > > > > file_update_time() checks for O_NOMTIME and aborts the update if it's > > > > set, just like the current check for the in-kernel inode flag > > > > S_NOCMTIME. I didn't update any other mtime update sites. They could be > > > > added as we decide that it's appropriate to do so. > > > > > > > > I opted not to name the flag O_NOCMTIME because I didn't want the name > > > > to imply that ctime updates would be prevented for other inode changes > > > > like updating i_size in truncate. Not updating ctime is a side-effect > > > > of removing mtime updates when it's the only thing changing in the > > > > inode. > > > > > > > > The criteria for using O_NOMTIME is the same as for using O_NOATIME: > > > > owning the file or having the CAP_FOWNER capability. If we're not > > > > comfortable allowing owners to prevent mtime/ctime updates then we > > > > should add a tunable to allow O_NOMTIME. Maybe a mount option? > > > > > > > > > > Just out of curiosity, if you need to modify the application anyway, > > > why wouldn't use of fdatasync() when flushing be able to offer a > > > similar performance boost? > > > > Although fdatasync(2) doesn't have to update synchronously, it does > > eventually get written, and that can trigger lots of unwanted IO. > > And the unwanted IO is per file. Are there circumstances where the > write:file ratio is small enough that dirty inode writes could start to > add up to meaningful write amplification? Yeah, exactly: in some not-so-uncommon workloads it's approaching 1:1. sage -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
