[lustre-discuss] trimming flash-based external journal device

[Andreas Dilger]

On Aug 5, 2021, at 17:44, Nathan Dauchy - NOAA Affiliate via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>> wrote:

On Thu, Aug 5, 2021 at 3:23 PM Andreas Dilger <adilger at whamcloud.com<mailto:adilger at whamcloud.com>> wrote:
On Aug 5, 2021, at 13:29, Nathan Dauchy wrote:
Andreas, thanks as always for your insight.  Comments inline...

On Thu, Aug 5, 2021 at 10:48 AM Andreas Dilger <adilger at whamcloud.com<mailto:adilger at whamcloud.com>> wrote:
On Aug 5, 2021, at 09:28, Nathan Dauchy via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>> wrote:
Question:  Is it possible that a flash journal device on an ext4 filesystem can reach a point where there are not enough clean blocks to write to, and they can suffer from very degraded write performance?
...
Another related question would be how to benchmark the journal device on it's own, particularly write performance, without losing data on an existing file system; similar to the very useful obdfilter-survey tool, but at a lower level.  But I am primarily looking to understand the nuances of flash devices and ldiskfs external journals a bit better.
While the external journal device has an ext4 superblock header for identification (UUID/label), and a feature flag that prevents it from being mounted/used directly, it is not really an ext4 filesystem, just a flat "file".  You'd need to remove it from the main ext4/ldiskfs filesystem, reformat it as ext4 and mount locally, and then run benchmarks (e.g. "dd" would best match the JBD2 workload, or fio if you want random IOPS) against it.  You could do this before/after trim (could use fstrim at this point) to see if it affects the performance or not.

OK, thanks for confirming that there is no magic ext4 journal benchmarking tool.  I'll stop searching.  ;-)
Note that there *are* some journal commit statistics - /proc/fs/jbd2/<dev>/info that you might be able to compare between devices.  Probably the most interesting is "average transaction commit time", which is how long it takes to write the blocks to the journal device after the transaction starts to commit.


Oh, that is interesting!

The "average transaction commit time" seems to fluctuate, possibly with load, and doesn't have an obvious correlation to the slower OSTs.  But perhaps I'll look at it again when running a clean benchmark during a future dedicated time.

I _did_ find a stark difference in other metrics though:

# pdsh -g oss "grep 'handles per' /proc/fs/jbd2/md*/info" | sort
lfs4n04:   17 handles per transaction
lfs4n05:   18 handles per transaction
lfs4n06:   18 handles per transaction
lfs4n07:   18 handles per transaction
lfs4n08:   17 handles per transaction
lfs4n09:   17 handles per transaction
lfs4n10:   18 handles per transaction
lfs4n11:   18 handles per transaction
lfs4n12:   18 handles per transaction
lfs4n13:   17 handles per transaction
lfs4n16:   192 handles per transaction
lfs4n17:   178 handles per transaction
lfs4n18:   198 handles per transaction
lfs4n19:   192 handles per transaction
# pdsh -g oss "grep 'logged blocks per' /proc/fs/jbd2/md*/info" | sort
lfs4n04:   24 logged blocks per transaction
lfs4n05:   24 logged blocks per transaction
lfs4n06:   25 logged blocks per transaction
lfs4n07:   25 logged blocks per transaction
lfs4n08:   24 logged blocks per transaction
lfs4n09:   24 logged blocks per transaction
lfs4n10:   25 logged blocks per transaction
lfs4n11:   24 logged blocks per transaction
lfs4n12:   24 logged blocks per transaction
lfs4n13:   24 logged blocks per transaction
lfs4n16:   103 logged blocks per transaction
lfs4n17:   98 logged blocks per transaction
lfs4n18:   106 logged blocks per transaction
lfs4n19:   103 logged blocks per transaction

The last 4 nodes are the expansion OSTs which are performing better.  What does that difference indicate?

I don't think you can read too much into those numbers.  "handles per transaction" usually depend on the IO load, due to more threads joining a single transaction before it commits due to sync.  The "blocks per transaction" is strongly correlated, in this case it looks like only 1 or 2 blocks per handle (maybe random 4KB writes, with occasional metadata updates?)

On the OSTs the transactions are usually forced to commit quickly to allow clients to stop pinning their dirty pages in cache.  The pages are kept in case an RPC resend is needed, but can be dropped as soon as they commit to storage on the server.  On the MDTs there is less memory pressure, so transaction commits are often timed out (5s) to aggregate as many changes as possible.

If the new OSTs are less full than the old ones (maybe 4x more free space?) then they would naturally get more objects allocated due to OST space balance (about the 4x seen here).  That would lead to 4x more blocks per commit, which would make it appear that the new OSTs are doing better, but really this is just more aggregation of the very small write sizes.

Cheers, Andreas
--
Andreas Dilger
Lustre Principal Architect
Whamcloud







-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20210806/3166f898/attachment.html>