The XFS filesystem is generally known for good performance when working with large files. Cassandra SSTables fit that description, which is why a variety of reputable sources recommend XFS for Cassandra data partitions.
We are currently reimaging nodes for multi-instance operation. I think it is worth using this opportunity to configure those nodes to use XFS. As far as I can tell, all it would take is a small patch in puppet/modules/install_server/files/autoinstall/partman/.
Earlier IRC discussion
[10:11] <gwicke> I have seen others recommend this, and it seems to make sense as cassandra sstables are basically large files written sequentially
[10:11] <gwicke> example: https://tobert.github.io/pages/als-cassandra-21-tuning-guide.html
[10:11] <gwicke> /cc urandom ^^
[10:12] <gwicke> https://books.google.com/books?id=6Q68AQAAQBAJ&pg=PT182&lpg=PT182&dq=cassandra+xfs&source=bl&ots=XplnkYkubr&sig=aaVMaHC52sxFhFg76Z1qgcDGCdc&hl=en&sa=X&ved=0ahUKEwi506T8oLvJAhVQ32MKHQEqB1EQ6AEIZTAM#v=onepage&q=cassandra%20xfs&f=false
[10:13] <godog> I'm not opposed in principle, but I'd like to see evidence that we have an actual problem or might in the future
[10:14] <gwicke> the selling point is that XFS tends to perform better for this kind of workload, and is as mature as ext4
[10:14] <gwicke> we use it for other similar workloads as well
[10:16] <godog> might be, my experience with swift and high load due to xfs deadlocks has been different
[10:17] <gwicke> a quick grep suggests that it's being used for dbs, elasticsearch, graphite, backend varnishes
[10:19] <gwicke> I don't feel super strongly about it, just thought that the re-image could be a good opportunity to tweak this
[10:21] <gwicke> the scylladb folks recommend xfs as well, and they are quite knowledgeable about linux internals and performance
[10:23] <gwicke> "DataStax recommends deploying Cassandra on XFS or ext4." -- http://docs.datastax.com/en/cassandra/2.1/cassandra/planning/architecturePlanningHardware_c.html
[10:25] <godog> I'm of the opinion of not introducing differences if we can avoid it, but if there's a case to be made that it performs better than ext4 for us then for sure