This isn't directly related to Cassandra, but we did a bunch of I/O and disk load testing about 2 years ago when we started migrating to a new MTA platform looking specifically at filesystem performance. We compared EXT2, EXT3, GFS2, XFS and EXT4 under RHEL5/CentOS5 on a commodity box using 6 internal SAS drives in a Raid1+0 configuration. EXT2/3 were by far the worst performing, having almost exponential disk usage growth as the number of inodes in use increased. After the directory structure had more than a few thousand files in it the seek and read times for the directory structures became almost unusable (40+ seconds before an ls -l would start returning data). GFS2 as a local file system performed rather well, however that usage is being deprecated, leaving us with XFS and EXT4 - both of which performed very well under our benchmark loads - they could both handle hundreds of thousands of files in a single directory structure without significant increases in disk utilization and lookup time vs EXT2/3. We settled on XFS under CentOS as our final solution mainly because EXT4 was still in preview when we were doing our benchmarking had a few nasty data loss bugs still. Google also took a hard look at filesystems late last year and settled on EXT4 as it allowed them to upgrade their EXT2 filesystems in place. You can read an archive of their post about it here: http://lists.openwall.net/linux-ext4/2010/01/04/8
On Aug 24, 2010, at 11:58 AM, S Ahmed wrote: > Is there a particular linux flavor that plays best with Cassandra? > > I believe the file system plays big role also, any comments in this regard? > > thanks.
