Re: [PATCH, 3.7-rc7, RESEND] fs: revert commit bbdd6808 to fallocateUAPI

[Ric Wheeler]

On 12/08/2012 08:52 AM, Howard Chu wrote:
> Dave Chinner wrote:
> > On Fri, Dec 07, 2012 at 03:25:53PM -0800, Howard Chu wrote:
> > > I have to agree that, if this is going to be an ext4-specific
> > > feature, then it can just be implemented via an ext4-specific ioctl
> > > and be done with it. But I'm not convinced this should be an
> > > ext4-specific feature.
> > >
> > > As for "fix the problem properly" - you're fixing the wrong problem.
> > > This type of feature is important to me, not just because of the
> > > performance issue. As has already been pointed out, the performance
> > > difference may even be negligible.
> > >
> > > But on SSDs, the issue is write endurance. The whole point of
> > > preallocating a file is to avoid doing incremental metadata updates.
> > > Particularly when each of those 1-bit status updates costs entire
> > > blocks, and gratuitously shortens the life of the media. The fact
> > > that avoiding the unnecessary wear and tear may also yield a
> > > performance boost is just icing on the cake. (And if the perf boost
> > > is over a factor of 2:1 that's some pretty damn good icing.)
> >
> > That's a filesystem implementation specific problem, not a generic
> > fallocate() or unwritten extent conversion problem.
>
> > Besides, ext4 doesn't write back every metadata modification that is
> > made - they are aggregated in memory and only written when the
> > journal is full or the metadata ages out. Hence unwritten extent
> > conversion has very little impact on the amount of writes that are
> > done to the flash because it is vastly dominated by the data writes.
> >
> > Similarly, in XFS you might see a few thousand or tens of thousands
> > of metadata blocks get written once every 30s under such a random
> > write workload, but each metadata block might have gone through a
> > million changes in memory since the last time it was written.
> > Indeed, in that 30s, there would have been a few million random data
> > writes so the metadata writes are well and truly lost in the
> > noise...
>
> That's only true if write caching is allowed. If you have a transactional 
> database running, it's syncing every transaction to media.

The math just does not add up - no device sustains millions of random IO's per 
second.

Each class of device has so many IOs it can do per second. S-ATA disks do say 
40-50 IOPS, SAS maybe twice that, enterprise arrays 10k IOPS and PCI-E cards 
100k IOPS.

Transactional databases accumulate multiple updates in memory and commit to disk 
*in transactions* to pack as much as possible into the IOPS that the device has. 
Batching is a core principle of database performance (just like we do in the 
guts of XFS or ext4 in our file system transactions).

If you use a transactional DB, you should pre-allocate the table space (and 
probably the log file), but it would be wise to pre-allocate it and zero the 
blocks since DB's are long lived items. Pay once for the initialization and off 
you go. More expensive for really large tables on slow devices of course.

Again, we are back to the core need here - we want to improve the performance of 
a random, small IO workload. This workload was historically *so* painful 
(remember that 40-50 IOP's for a s-ata disk :)), that pretty much every sane 
application avoided random IO like the plague.

With some of the newer devices like the new SSD's, random IO gets to be more 
reasonable and we need to fix the performance to accommodate workloads that were 
not normal.

Ric


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/