Hi,
It is about time to post Tux3 patches for review. Almost. Running the
classic kernel untar test on an older 60 MB/sec hard disk turned up a
performance gap versus Ext4 by a factor of six. Hmm, that is in the
range where it could be a basic design bug, we need to do something.
Tux3 performed fine on the actual untar, but syncing those thousands of
small files to disk was slow. The cause turned out to be multiple
issues. Two were unimplemented design features. For every file, we were
writing a btree root and a leaf block with pointers to data extents.
Instead, most btrees can be "zero depth" with just a leaf block and no
index block. And many files are just a single extent so the btree is not
needed at all. When writing lots of small files, we were transferring up
to three times more blocks than necessary.
The kernel tarball showed a metadata-to-data ratio of about 1.2. This is
way too high, it should be less than a tenth of that. After factoring
out those extra blocks, a huge performance gap still remains. This must
have something to do with seeking, but our disk layout for this load is
pretty good, so what is going on?
It turns out to be important not only to write to the right place, but
at the right time. The block scheduler tries to merge physically
contiguous requests even when submitted out of order, but if the
requests are too far apart in time, an earlier request may have already
left the queue by the time an adjacent request shows up. This causes
extra, costly seeks on spinning disks. Ideally, we want our writes
contiguous in both time and space. Then the disk hardware should be able
to make seek costs effectively disappear and stream the data out at
close to media speed.
Fixes for these issues took the form of patches from Hirofumi to
eliminate redundant btree roots and submit metadata writes in a better
order for block scheduling, and a patch from me to implement a planned
"direct extent" feature to eliminate btrees completely for small files.
Iterative improvements went something like this:
* Start: 60 seconds to sync but Ext4 only needs 10 seconds
* Eliminate most btree roots => now sync in 53 seconds
* Submit file btree together with data => sync in 19 seconds
* Eliminate btree complete for small files => 15 seconds
* Flush inode table blocks after all data => 8 seconds
We ended up with:
untar:
real 0m2.706s
user 0m0.360s
sys 0m2.160s
sync:
real 0m8.651s
user 0m0.000s
sys 0m0.024s
Ext4 takes about 4 seconds to untar and 10 seconds to sync, turning in a
respectable 50 MB/sec write bandwidth on a 62 MB/sec disk. Tux3 now
syncs at 60 MB/sec, or 97% of raw media bandwidth. So we went from 500%
slower to 23% faster, woohoo. The cost for this is that we dropped out
of sight for a few weeks. Maybe it was worth it because the performance
artifact was so big that it could have been a major design deficiency
instead of what it really was: leaving some details for later.
When we checked read performance on the untarred tree, we immediately
saw mixed results. Re-tarring the kernel tree is faster than Ext4, but
directory listing is slower by a multiple. So we need to analyze and fix
ls without breaking the good tar and untar behavior. The question is, is
it worth another delay before putting Tux3 patches up for review?
I think not. In fact, by going quiet when we hit these things, we
detract from the spectator sport aspect of open source. It might not be
any faster to work in public, but it is more fun. Plus, it engraves a
record on the internet as a guide for the next effort to invent a new
and wonderful filesystem. It is hard to overstate the value to our
project of all the historical chatter about design and development
process for Ext4 and other Linux filesystems. We often find ourselves
retracing the same learning processes. By doing this work in public, we
give something back.
Improving ls performance to Ext4 standards may just be a matter of
implementing inode table readahead, or it might be that plus something
else. In any case, this will go on the longish list of important issues
that are not central.
Inline data is a related item already on that list. There is a nice plan
for it, where the same design feature handles inline files and tail
packing, similar to extended attributes. In particular, most directories
in the kernel tarball are small enough to inline, which could speed up
ls significantly. With most files and directories inlined, a Tux3
filesystem becomes a single, fatter btree, with different issues and
tradeoffs. On the whole, I expect substantial improvements in both space
utilization and performance. The final chapter in the tar performance
saga has yet to be written.
We also need to ask why we are putting so much effort into performing
well on spinning disks, which are rapidly disappearing. Two reasons.
First, spinning disks are not gone yet, they are just migrating to a
backend storage role. Second, optimizations for spinning disk are
helpful for solid state storage more often than not. In this case,
keeping related requests close together in time lets the flash
translation layer pack its erase blocks better, reducing write
multiplication during space recovery, and in turn increasing media life
and performance. So it is too soon to forget about the idiosyncrasies
and challenges of traditional rotating media, perhaps ten or twenty
years too soon.
Regards,
Daniel
Yesterday I wrote:
> When we checked read performance on the untarred tree, we immediately saw
> mixed results. Re-tarring the kernel tree is faster than Ext4, but
directory listing is
> slower by a multiple. So we need to analyze and fix ls without
breaking the good tar
> and untar behavior. The question is, is it worth another delay before
putting Tux3
> patches up for review?
Hirofumi would not let me slink cowardly away from that open question.
We noticed that Tux3 does slightly more than one seek per directory,
which is entirely reasonable. But Ext4 goes way beyond that and does
some special magic to read multiple directories per seek. The only
possible way to do that is, pack directories together so there are many
per track. A bit of sleuthing confirmed that this is indeed the case,
and apparently comes from a patch posted by Ted T'so a few years back:
lwn.net/Articles/319829/
[PATCH, RFC] ext4: New inode/block allocation algorithms for
flex_bg filesystems
That patch was aimed at speeding up fsck and the huge ls speedup appears
to have gone unnoticed.
Thus inspired, Hirofumi whipped up a prototype patch to allocate new
directories first, per delta. Result: Tux3 went from 400% slower to 25%
faster than Ext4, for "ls -R" of the kernel source. Even better, tar and
untar performance stayed about the same with Tux3 topping the untar test
at 20% faster and tar at 350%. (The lopsided tar result looks like a
performance bug for Ext4.)
This optimization only applies to spinning disk. It is pretty hard to
think of a reason why packing directories together would benefit flash.
Maybe, directories that are written together are more likely to be
updated together? But it doesn't hurt flash either, and is another data
point to support our theory that optimizing for spinning disk also
optimizes for flash. We are still waiting for the first counterexample
to show up.
There are a few reasons why Tux3 has an edge for the case exercised by
the kernel source loads:
* Defer everything
Tux3 takes the idea of delayed allocation much further and delays nearly
everything. Directory updates and inode number selection are the only
exceptions. (In future we will attempt to defer the namespace updates as
well.)
* Front back separation
Besides enabling defer-everything, this simplifies locking and reduces
contention a lot, for both read and update. For now, a naive locking
strategy serves us well. Eventually we will multithread the backend,
which will help with high processor core counts, once we get there.
* Big deltas
Under heavy update load, Tux3 deltas grow as big as cache will allow, so
per-delta layout algorithms have a big data window available to optimize
over. With our current strategy, we observe an effect similar to Ext4
flex_bg, where directories and other metadata tend to self-organize
along delta boundaries, with beneficial performance effects. We might
control this behavior more explicitly in future.
* More inodes per inode table block.
Tux3 stores about 57 inodes per block while Ext4 typically has 16 or
less. Multiple inodes per block already resembles a kind of inode table
readahead. Without this, there would be two seeks per directory even
with directories packed together.
Anyway, I don't think we need to hang our heads in shame for performance
reasons at this point, even though plenty of major optimization issues
still remain on the list. For example, you can embarras Tux3 just by
running a benchmark with 10,000 files per directory. The answer to that
one is Shardmap, which needs a couple of months to bring up and solves a
problem that does not come up on your home server or phone. Not a reason
to get sidetracked again..
Regards,
Daniel